Monte Carlo Ray Tracing

Siggraph 2003 Course 44

Tuesday, July 29, 2003

Organizer
Henrik Wann Jensen
University of California, San Diego

Lecturers
James Arvo
University of California, Irvine
Phil Dutre
Katholieke Universiteit Leuven
Alexander Keller
Universitat Kaiserslautern
Henrik Wann Jensen
University of California, San Diego
Art Owen
Stanford University
Matt Pharr
NVIDIA
Peter Shirley
University of Utah

Abstract

This full day course will provide a detailed overview of state of the art in Monte
Carlo ray tracing. Recent advances in algorithms and available compute power
have made Monte Carlo ray tracing based methods widely used for simulating
global illumination. This course will review the fundamentals of Monte Carlo
methods, and provide a detailed description of the theory behind the latest techniques and algorithms used in realistic image synthesis. This includes path tracing,
bidirectional path tracing, Metropolis light transport, irradiance caching and photon mapping.

Course Syllabus

8:30 Introduction and Welcome

Henrik Wann Jensen

8:40 Why Monte Carlo Ray Tracing?

Peter Shirley
The global illumination problem
Monte Carlo as a general tool

9:00 Fundamentals of Monte Carlo Integration

Peter Shirley
An overview of Monte Carlo integration techniques.
Rejection methods
Importance sampling
Stratified sampling Hammersley points
Arbitrary-edge discrepancy
Applications of quasirandom techniques to distribution ray tracing
Spectral sampling techniques

9:30 A Recipe for Sampling Algorithms

James Arvo
Sampling of special geometries and reflection models
Russian roulette

10:00 Break
10:30 Direct Illumination
Peter Shirley
Sampling of light sources
Special types of light sources
Efficient sampling of many light sources

11:00 Variance Reduction Techniques

James Arvo
Combined estimators
Hybrid sampling methods.

11:00 Quasi-Monte Carlo Techniques

Art Owen
Good sample distributions
Efficient sampling of high-dimensional spaces

12:00 Lunch
2:15 The Rendering Equation
Philip Dutre
The path integral formulation
Path tracing
Russian Roulette
Adjoint techniques
Bidirectional transport
Bidirectional path tracing

3:00 Quasi-Monte Carlo Techniques II

Alexander Keller
Algorithms for (randomized) quasi-Monte Carlo sample points
Variance reduction by quasi-Monte Carlo points
Benefits of correlated sampling
How to boost a ray tracer by quasi-Monte Carlo

4:00 Break
4:15 Metropolis Sampling
Matt Pharr
One dimensional setting
Motion blur
Metropolis light transport

5:00 Biased Monte Carlo Ray Tracing

Henrik Wann Jensen
Biased vs. consistent methods
Filtering Techniques
Irradiance Caching
Photon Mapping

6:00 Conclusion and Questions

Contents
1

Introduction by Henrik Wann Jensen

1.1 Purpose of this Course . . . . . . . . . . . . . . . . . . . . . . .
1.2 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . .

11
12
12
12

Fundamentals of Monte Carlo Integration by Peter Shirley

2.1 Background and Terminology . . . . . . . . . . . . . . . . . . .
2.1.1 One-dimensional Continuous Probability Density Functions
2.1.2 One-dimensional Expected Value . . . . . . . . . . . . .
2.1.3 Multi-dimensional Random Variables . . . . . . . . . . .
2.1.4 Variance . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.5 Estimated Means . . . . . . . . . . . . . . . . . . . . . .
2.2 Monte Carlo Integration . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Quasi-Monte Carlo Integration . . . . . . . . . . . . . . .
2.2.2 Multidimensional Monte Carlo Integration . . . . . . . .
2.3 Choosing Random Points . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Function inversion . . . . . . . . . . . . . . . . . . . . .
2.3.2 Rejection . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3 Metropolis . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . .
2.5 Density Estimation . . . . . . . . . . . . . . . . . . . . . . . . .

13
13
13
14
15
16
17
17
19
20
21
22
24
25
27
28

Direct Lighting via Monte Carlo Integration by Peter Shirley

3.1 Mathematical framework . . . . . . . . . . . . . . . . . .
3.2 Sampling a spherical luminaire . . . . . . . . . . . . . . .
3.3 Non-diffuse Luminaries . . . . . . . . . . . . . . . . . . .
3.4 Direct Lighting from Many Luminaires . . . . . . . . . .

29
29
31
34
35

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3.4.1
3.4.2
4

Constant i . . . . . . . . . . . . . . . . . . . . . . . . .
Linear i . . . . . . . . . . . . . . . . . . . . . . . . . .

36
37

Stratified Sampling of 2-Manifolds by Jim Arvo

4.1 Introduction . . . . . . . . . . . . . . . . . .
4.2 A Recipe for Sampling Algorithms . . . . . .
4.3 Analytic Area-Preserving Parametrizations .
4.3.1 Sampling Planar Triangles . . . . . .
4.3.2 Sampling the Unit Disk . . . . . . . .
4.3.3 Sampling the Unit Hemisphere . . . .
4.3.4 Sampling a Phong Lobe . . . . . . .
4.3.5 Sampling Spherical Triangles . . . .
4.4 Sampling Projected Spherical Polygons . . .
4.4.1 The Cumulative Marginal Distribution
4.4.2 The Sampling Algorithm . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

39
39
41
46
46
47
48
49
49
54
56
57

Combining Sampling Strategies by Jim Arvo

5.1 Introduction . . . . . . . . . . . . . . .
5.2 Using Multiple PDFs . . . . . . . . . .
5.3 Possible Weighting Functions . . . . . .
5.4 Obvious is also Nearly Optimal . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

63
63
64
66
68

.
.
.
.
.
.
.
.

69
70
72
74
77
80
84
86
87

.
.
.
.

89
89
96
98
103

.
.
.
.

Quasi-Monte Carlo Sampling by Art B. Owen

6.1 Crude Monte Carlo . . . . . . . . . . . .
6.2 Stratification . . . . . . . . . . . . . . . .
6.3 Multiple Stratification . . . . . . . . . . .
6.4 Uniformity and Discrepancy . . . . . . .
6.5 Digital Nets and Related Methods . . . .
6.6 Integration Lattices . . . . . . . . . . . .
6.7 Randomized Quasi-Monte Carlo . . . . .
6.8 Padding and Latin Supercube Sampling .

.
.
.
.

.
.
.
.
.
.
.
.

Monte Carlo Path Tracing by Pat Hanrahan

7.1 Solving the Rendering Equation . . . . . .
7.2 Monte Carlo Path Tracing . . . . . . . . . .
7.3 Random Walks and Markov Chains . . . .
7.4 Adjoint Equations and Importance Sampling

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.

109
109
111
114
115
115
116
117
118
118

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

121
123
123
125
126
127
128
129
131
131
133
135
139
139
141
143
144
146
147
147

10 Biased Techniques by Henrik Wann Jensen

10.1 Biased vs. Unbiased . . . . . . . . . . . . . . . . . . . . . . . . .
10.2 Filtering Techniques . . . . . . . . . . . . . . . . . . . . . . . .
10.3 Adaptive Sampling . . . . . . . . . . . . . . . . . . . . . . . . .

149
149
151
153

The Rendering Equation and Path Tracing by Philip Dutre

8.1 Formulations of the rendering equation . . . . . . . . . .
8.2 Importance function . . . . . . . . . . . . . . . . . . . .
8.3 Path formulation . . . . . . . . . . . . . . . . . . . . .
8.4 Simple stochastic ray tracing . . . . . . . . . . . . . . .
8.5 Russian Roulette . . . . . . . . . . . . . . . . . . . . .
8.6 Indirect Illumination . . . . . . . . . . . . . . . . . . .
8.6.1 Hemisphere sampling . . . . . . . . . . . . . . .
8.6.2 Importance sampling . . . . . . . . . . . . . . .
8.6.3 Overview . . . . . . . . . . . . . . . . . . . . .
Metropolis Sampling by Matt Pharr
9.1 Overview . . . . . . . . . . . . . . . . . .
9.1.1 Detailed Balance . . . . . . . . . .
9.1.2 Expected Values . . . . . . . . . .
9.2 One Dimensional Setting . . . . . . . . . .
9.2.1 Mutations and Transition Functions
9.2.2 Start-up bias . . . . . . . . . . . .
9.2.3 Initial Results . . . . . . . . . . . .
9.2.4 Ergodicity . . . . . . . . . . . . . .
9.2.5 Mutations via PDFs . . . . . . . .
9.3 Motion Blur . . . . . . . . . . . . . . . . .
9.3.1 Basic Mutation Strategies . . . . .
9.3.2 Re-normalization . . . . . . . . . .
9.3.3 Adapting for large variation in f . .
9.3.4 Color . . . . . . . . . . . . . . . .
9.4 Metropolis Light Transport . . . . . . . . .
9.4.1 Path Mutations . . . . . . . . . . .
9.4.2 Pixel Stratification . . . . . . . . .
9.4.3 Direct Lighting . . . . . . . . . . .
9.4.4 Participating Media . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

10.4 Irradiance Caching . . . . . .

10.5 Photon Mapping . . . . . . . .
10.5.1 Pass 1: Photon Tracing
10.5.2 The Radiance Estimate
10.5.3 Pass 2: Rendering . .

.
.
.
.
.

References

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

153
156
156
157
159
162

10

Chapter 1

Introduction
By Henrik Wann Jensen

Realistic image synthesis is increasingly important in areas such as entertainment

(movies, special effects and games), design, architecture and more. A common
trend in all these areas is to request more realistic images of increasingly complex
models. Monte Carlo ray tracing based techniques are the only methods that can
handle this complexity. Recent advances in algorithms and compute power has
made Monte Carlo ray tracing the natural choice for most problems. This is a
significant change from just a few years back when the (finite element) radiosity
method was the prefered algorithm for most graphics researchers.
Monte Carlo ray tracing has several advantages over finite element methods. A
recent list from [33] includes:
Geometry can be procedural
No tessellation is necessary
It is not necessary to precompute a representation for the solution
Geometry can be duplicated using instancing
Any type of BRDF can be handled
Specular reflections (on any shape) are easy
Memory consumption is low
The accuracy is controlled at the pixel/image level
11

 Complexity has empirically been found to be O(log N ) where N is number of scene elements. Compare this with O(N log N ) for the fastest finite
element methods [12].
In addition one might add that Monte Carlo ray tracing methods can be very
easy to implement. A basic path tracing algorithm which has all of the above
advantages is a relatively straightforward extension to ray tracing.
The main problem with Monte Carlo ray tracing is variance seen as noise in the
rendered images. This noise can be eliminated by using more samples. Unfortunately the convergence of Monte Carlo methods is quite slow, and a large number
of samples can be necessary to reduce the variance to an acceptable level. Another
way of reducing variance is to try to be more clever; a large part of this course material is devoted to techniques and algorithms for making Monte Carlo ray tracing
more efficient.

1.1

Purpose of this Course

The purpose of this course is to impart upon the attendies a thorough understanding
of the principles of Monte Carlo ray tracing methods, as well as a detailed overview
of the most recently developed methods.

1.2

Prerequisites

The reader is expected to have a good working knowledge of ray tracing and to
know the basics of global illumination. This includes knowledge of radiometric
terms (such as radiance and flux) and knowledge of basic reflection models (such
as diffuse, specular and glossy).

1.3

Acknowledgements

Funding for the authors of these notes include DARPA DABTB63-95-C0085 and
an NSF Career Award (CCR9876332).

12

Chapter 2

Fundamentals of Monte Carlo

Integration
By Peter Shirley

This Chapter discusses Monte Carlo integration, where random numbers are used
to approximate integrals. First some basic concepts from probability are reviewed,
and then they are applied to numerically estimate integrals. The problem of estimating the direct lighting at a point with arbitrary lighting and reflection properties
is then discussed. The next Chapter applies Monte Carlo integration to the direct
light problem in ray tracing.

2.1 Background and Terminology

Before getting to the specifics of Monte Carlo techniques, we need several definitions, the most important of which are continuous random variable, probability
density function (pdf), expected value, and variance. This section is meant as a
review, and those unfamiliar with these terms should consult an elementary probability theory book (particularly the sections on continuous, rather than discrete,
random variables).

2.1.1

One-dimensional Continuous Probability Density Functions

Loosely speaking, a continuous random variable x is a scalar or vector quantity

that randomly takes on some value from the real line R = (, +). The
13

behavior of x is entirely described by the distribution of values it takes. This distribution of values can be quantitatively described by the probability density function,
p, associated with x (the relationship is denoted x p). The probability that x will
take on a value in some interval [a, b] is given by the integral:
Z b
Probability(x [a, b]) =
p(x)dx.
(2.1)
a

Loosely speaking, the probability density function p describes the relative likelihood of a random variable taking a certain value; if p(x1 ) = 6.0 and p(x2 ) = 3.0,
then a random variable with density p is twice as likely to have a value near x1
than it it to have a value near x2 . The density p has two characteristics:
p(x) 0 (Probability is nonnegative),
Z

(2.2)

p(x)dx = 1 (Probability(x R) = 1).

(2.3)

As an example, the canonical random variable takes on values between zero

(inclusive) and one (non-inclusive) with uniform probability (here uniform simply
means each value for is equally likely). This implies that the probability density
function q for is:

1 if 0 1
q() =
0 otherwise
The space over which is defined is simply the interval [0, 1). The probability that
takes on a value in a certain interval [a, b] [0, 1) is:
Z b
Probability(a b) =
1dx = b a.
a

2.1.2

One-dimensional Expected Value

The average value that a real function f of a one dimensional random variable with
underlying pdf p will take on is called its expected value, E(f (x)) (sometimes
written Ef (x)):
Z
E(f (x)) = f (x)p(x)dx.
The expected value of a one dimensional random variable can be calculated by
letting f (x) = x. The expected value has a surprising and useful property: the
14

expected value of the sum of two random variables is the sum of the expected
values of those variables:
E(x + y) = E(x) + E(y),
for random variables x and y. Because functions of random variables are themselves random variables, this linearity of expectation applies to them as well:
E(f (x) + g(y)) = E(f (x)) + E(g(y)).
An obvious question is whether this property holds if the random variables being
summed are correlated (variables that are not correlated are called independent).
This linearity property in fact does hold whether or not the variables are independent! This summation property is vital for most Monte Carlo applications.

2.1.3

Multi-dimensional Random Variables

The discussion of random variables and their expected values extends naturally to
multidimensional spaces. Most graphics problems will be in such higher-dimensional
spaces. For example, many lighting problems are phrased on the surface of the
hemisphere. Fortunately, if we define a measure on the space the random variables occupy, everything is very similar to the one-dimensional case. Suppose the
space S has associated measure , for example S is the surface of a sphere and
measures area. We can define a pdf p : S 7 R, and if x is a random variable with
x p, then the probability that x will take on a value in some region Si S is
given by the integral:
Z
Probability(x Si ) =
p(x)d
(2.4)
Si

Here Probability(event) is the probability that event is true, so the integral is the
probability that x takes on a value in the region Si .
In graphics S is often an area (d = dA = dxdy), or a set of directions (points
on a unit sphere: d = d = sin dd). As an example, a two dimensional
random variable is a uniformly distributed random variable on a disk of radius
R. Here uniformly means uniform with respect to area, e.g., the way a bad dart
players hits would be distributed on a dart board. Since it is uniform, we know that
p() is some constant. From Equation 2.3, and the fact that area is the appropriate
15

measure, we can deduce that p() = 1/(R2 ). This means that the probability
that is in a certain subset S1 of the disk is just:
Z
1
Probability( S1 ) =
dA.
2
R
S1
This is all very abstract. To actually use this information we need the integral in
a form we can evaluate. Suppose Si is the portion of the disk closer to the center
than the perimeter. If we convert to polar coordinates, then is represented as a
(r, ) pair, and S1 is where r < R/2. Note that just because is uniform does
not imply that theta or r are necessarily uniform (in fact, theta is, and r is not
uniform). The differential area dA becomes r dr d. This leads to:
Z 2 Z R
2
R
1
Probability(r < ) =
r dr d = 0.25.
2
2
R
0
0
The formula for expected value of a real function applies to the multidimensional case:
Z
E(f (x)) =
f (x)p(x)d,
S

Where x S and f : S 7 R, and p : S 7 R For example, on the unit square

S = [0, 1] [0, 1] and p(x, y) = 4xy, the expected value of the x coordinate for
(x, y) p is:
Z
E(x) =
f (x, y)p(x, y)dA
S
Z 1Z 1
=
4x2 y dx dy
0

2
=
3
Note that here f (x, y) = x.

2.1.4

Variance

The variance, V (x), of a one dimensional random variable is by definition the

expected value of the square of the difference between x and E(x):
V (x) E([x E(x)]2 ).
Some algebraic manipulation can give the non-obvious expression:
V (x) = E(x2 ) [E(x)]2 .
16

The expression E([x E(x)]2 ) is more useful for thinking intuitively about variance, while the algebraically equivalent expression E(x2 ) [E(x)]2 is usually
convenient for calculations. The variance of a sum of random variables is the sum
of the variances if the variables are independent. This summation property of variance is one of the reasons it is frequently used in analysis of probabilistic models.
The square root of the variance is called the standard deviation, , which gives
some indication of expected absolute deviation from the expected value.

2.1.5

Estimated Means

Many problems involve sums of independent random variables xi , where the variables share a common density p. Such variables are said to be independent identically distributed (iid) random variables. When the sum is divided by the number
of variables, we get an estimate of E(x):
N
1 X
E(x)
xi .
N
i=1

As N increases, the variance of this estimate decreases. We want N to be large

enough that we have confidence that the estimate is close enough. However,
there are no sure things in Monte Carlo; we just gain statistical confidence that our
estimate is good. To be sure, we would have to have n = . This confidence is
expressed by Law of Large Numbers:
#
N
1 X
Probability E(x) = lim
xi = 1.
N N
"

i=1

2.2

Monte Carlo Integration

In this section the basic Monte Carlo solution methods for definite integrals are
outlined. These techniques are then straightforwardly applied to certain integral
problems. All of the basic material of this section is also covered in several of
the classic Monte Carlo texts. This section differs by being geared toward classes
of problems that crop up in Computer Graphics. Readers interested in a broader
treatment of Monte Carlo techniques should consult one of the classic Monte Carlo
texts [27, 72, 26, 98].
17

As discussed earlier, given a function f : S 7 R and a random variable x p,

we can approximate the expected value of f (x) by a sum:
Z
f (x)p(x)d

E(f (x)) =
xS

N
1 X
f (xi ).
N

(2.5)

i=1

Because the expected value can be expressed as an integral, the integral is also
approximated by the sum. The form of Equation 2.5 is a bit awkward; we would
usually like to approximate an integral of a single function g rather than a product
f p. We can get around this by substituting g = f p as the integrand:
Z
g(x)d
xS

N
1 X g(xi )
.
N
p(xi )

(2.6)

i=1

For this formula to be valid, p must be positive where g is nonzero.

So to get a good estimate, we want as many samples as possible, and we want
the g/p to have a low variance (g and p should have a similar shape). Choosing p
intelligently is called importance sampling, because if p is large where g is large,
there will be more samples in important regions. Equation 2.5 also shows the fundamental problem with Monte Carlo integration: diminishing return. Because the
variance of the estimate is proportional to 1/N , the standard deviation is propor
tional to 1/ N . Since the error in the estimate behaves similarly to the standard
deviation, we will need to quadruple N to halve the error.
Another way to reduce variance is to partition S, the domain of the integral,
into several smaller domains Si , and evaluate the integral as a sum of integrals
over the Si . This is called stratified sampling. Normally only one sample is taken
in each Si (with density pi ), and in this case the variance of the estimate is:
!


N
N
X
X
g(xi )
g(xi )
.
(2.7)
var
=
var
pi (xi )
pi (xi )
i=1

i=1

It can be shown that the variance of stratified sampling is never higher than unstratified if all strata have equal measure:
Z
Z
1
p(x)d =
p(x)d.
N S
Si
The most common example of stratified sampling in graphics is jittering for pixel
sampling [14].
18

method
importance
importance
importance
importance
stratified

sampling function

variance

(6 x)/(16)
1/4
(x + 2)/16
x/8
1/4

56.8N 1
21.3N 1
6.3N 1
0
21.3N 3

samples needed for

standard error of 0.008
887,500
332,812
98,437
1
70

Table 2.1: Variance for Monte Carlo Estimate of

R4
0

x dx

As an example of the Monte Carlo solution of an integral I set g(x) to be x

over the interval (0, 4):
Z 4
I=
x dx = 8.
(2.8)
0

The great impact of the shape of the function p on the variance of the N sample
estimates is shown in Table 2.1. Note that the variance is lessened when the shape
of p is similar to the shape of g. The variance drops to zero if p = g/I, but I is
not usually known or we would not have to resort to Monte Carlo. One important
principle illustrated in Table 2.1 is that stratified sampling is often far superior to
importance sampling. Although the variance for this stratification on I is inversely
proportional to the cube of the number of samples, there is no general result for the
behavior of variance under stratification. There are some functions where stratification does no good. An example is a white noise function, where the variance is
constant for all regions. On the other hand, most functions will benefit from stratified sampling because the variance in each subcell will usually be smaller than the
variance of the entire domain.

2.2.1

Quasi-Monte Carlo Integration

Although distribution ray tracing is usually phrased as an application of Equation 2.6, many researchers replace the i with more evenly distributed (quasirandom) samples (e.g. [13, 53]). This approach can be shown to be sound by
analyzing decreasing error in terms of some discrepancy measure [99, 97, 53, 67]
rather than in terms of variance. However, it is often convenient to develop a sampling strategy using variance analysis on random samples, and then to turn around
and use non-random, but equidistributed samples in an implementation. This approach is almost certainly correct, but its justification and implications have yet to
19

be explained.
For example, when evaluating a one dimensional integral on [0, 1] we could
use a set of N uniformly random sample points (x1 , x2 , , xN ) on [0, 1] to get
an approximation:
Z 1
N
1 X
f (xi ).
f (x)dx
N
0
i=1

Interestingly, we can replace the points (x1 , x2 , , xN ) with a set of non-random

points (y1 , y2 , , yN ), and the approximation will still work. If the points are
too regular, then we will have aliasing, but having correlation between the points
(e.g. using one dimension Poisson disk sampling), does not invalidate the estimate (merely the Monte Carlo argument used to justify the approximation!). In
some sense, this quasi-Monte Carlo method can be thought of as using the equidistributed points to estimate the height of f . This does not fit in with the traditional
quadrature approaches to numerical integration found in most numerical analysis
texts (because these texts focus on one-dimensional problems), but is no less intuitive once you are used to the idea.
The relative advantages of Monte Carlo versus QMC for graphics is still an
open question. QMC does have better convergence subject to certain conditions,
but these conditions are often not true in graphics. Also, there are rarely enough
samples taken in practice for the asymptotic analysis to apply. To further complicate matters, QMC sometimes produces aliasing. However, this aliasing is sometimes not visually objectionable in often looks better than the noise produced by
traditional Monte Carlo. For more information on this topic, see the recent work
of Alexander Keller.

2.2.2

Multidimensional Monte Carlo Integration

Applying Equation 2.6 to multidimensional integrals is straightforward, except that

choosing the multidimensional sampling points can be more involved than in the
one dimensional case.
As an example in two dimensions, suppose we want to integrate some function
f on the origin centered square [1, 1]2 . This can be written down as a integral
over a single two dimensional variable x:
Z
I=
f (x)dA.
[1,1]2

20

Applying Equation 2.6 to this gives us:

N
1 X f (xi )
I
,
N
p(xi )
i=1

where each xi is a two dimensional point distributed according to a two dimensional density p. We can convert to more explicit Cartesian coordinates and have a
form we are probably more comfortable with:
Z

f (x, y)dxdy

I=
y=1

x=1

N
1 X f (xi , yi )
.
N
p(xi , yi )
i=1

This is really no different than the form above, except that we see the explicit
components of xi to be (xi , yi ).
If our integral is over the of radius R, nothing really changes, except that the
sample points must be distributed according to some density on the disk. This is
why Monte Carlo integration is relatively easy: once the sample points are chosen,
the application of the formula is always the same.

2.3

Choosing Random Points

We often want to generate sets of random or pseudorandom points on the unit

square for applications such as distribution ray tracing. There are several methods
for doing this such as jittering and Poisson disk sampling. These methods give us
a set of N reasonably equidistributed points on the unit square: (u1 , v1 ) through
(uN , vN ).
Sometimes, our sampling space may not be square (e.g. a circular lens), or
may not be uniform (e.g. a filter function centered on a pixel). It would be nice if
we could write a mathematical transformation that would take our equidistributed
points (ui , vi ) as input, and output a set of points in our desired sampling space
with our desired density. For example, to sample a camera lens, the transformation
would take (ui , vi ) and output (ri , i ) such that the new points were approximately
equidistributed on the disk of the lens.
There are several ways to generate such non-uniform points, and this section
reviews the three most often used: function inversion, rejection, and Metropolis.
21

2.3.1

Function inversion

If the density is a one dimensional f (x) defined over the interval x [xmin , xmax ],
then we can generate random numbers i that have density f from a set of uniform
random numbers i , where i [0, 1]. To do this we need the cumulative probability distribution function P (x):
Z x
Probability( < x) = P (x) =
f (x0 )d
(2.9)
xmin

To get i we simply transform i :

i = P 1 (i )

(2.10)

where P 1 is the inverse of P . If P is not analytically invertible then numerical

methods will suffice because an inverse exists for all valid probability distribution
functions.
For example, to choose random points xi that have the density p(x) = 3x2 /2

on [1, 1], we see that P (x) = (x3 + 1)/2, and P 1 (x) = 3 2x 1, so we

can warp a set of canonical random numbers (1 , , N ) to the properly dis

tributed numbers (x1 , , xN ) = ( 3 21 1, , 3 2N 1). Of course, this

same warping function can be used to transform uniform Poisson disk samples
into nicely distributed samples with the desired density.
If we have a random variable = (x , y ) with two dimensional density
(x, y) defined on [xmin , xmax ] [ymin , ymax ] then we need the two dimensional
distribution function:
Z y Z x
P rob(x < x and y < y) = F (x, y) =
f (x0 , y 0 )d(x0 , y 0 )
ymin

xmin

We first choose an xi using the marginal distribution F (x, ymax ), and then choose
yi according to F (xi , y)/F (xi , ymax ). If f (x, y) is separable (expressible as g(x)h(y)),
then the one dimensional techniques can be used on each dimension.
For example, suppose we are sampling uniformly from the disk of radius R, so
p(r, ) = 1/(R2 ). The two dimensional distribution function is:
Z 0 Z r0
rdrd
r2
P rob(r < r0 and < 0 ) = F (r0 , 0 ) =
=
2
R
2R2
0
0
This means that a canonical pair (1 , 2 ) can be transformed to a uniform random

point on the disk: (r, ) = (R 1 , 22 ).

22

To choose random points on a triangle defined by vertices p0 , p1 , and p2 , a more

complicated analysis leads to the transformation u = 1 1 1 , v = (1 u)2 ,

and the random point p will is:
p = p0 + u(p1 p0 ) + v(p2 p0 ).
To choose reflected ray directions for zonal calculations or distributed ray tracing, we can think of the problem as choosing points on the unit sphere or hemisphere (since each ray direction can be expressed as a point on the sphere). For
example, suppose that we want to choose rays according to the density:
p(, ) =

n+1
cosn
2

(2.11)

Where n is a Phong-like exponent, is the angle from the surface normal and
[0, /2] (is on the upper hemisphere) and is the azimuthal angle ( [0, 2]).
The distribution function is:
Z Z
P (, ) =
p(0 , 0 ) sin 0 d0 d0
(2.12)
0

The sin 0 term arises because on the sphere d = sin dd. When the marginal
densities are found, p (as expected) is separable and we find that a (1 , 2 ) pair of
canonical random numbers can be transformed to a direction by:
1

(, ) = (arccos((1 r1 ) n+1 ), 2r2 )

One nice thing about this method is that a set of jittered points on the unit square
can be easily transformed to a set of jittered points on the hemisphere with a distribution of Equation 2.11. If n is set to 1 then we have a diffuse distribution needed
for a Monte Carlo zonal method.
For a zonal or ray tracing application, we choose a scattered ray with respect
~ (as opposed to the z axis). To do this we can first
to some unit normal vector N
convert the angles to a unit vector a:
~a = (cos sin , sin sin , cos )
We can then transform ~a to be an ~a0 with respect to by multiplying ~a by a rotation
matrix R (~a0 = R~a). This rotation matrix is simple to write down:

ux vx wx
R = uy vy wy
uz vz wz
23

where ~u = (ux , uy , uz ), ~v = (vx , vy , vz ), w

~ = (wx , wy , wz ), form a basis (an
orthonormal set of unit vectors where ~u = ~v w,
~ ~v = w
~ ~u, and w
~ = ~u ~v ) with
~
the constraint that w
~ is aligned with N :
w
~=

~
N
~|
|N

To get ~u and ~v , we need to find a vector ~t that is not collinear with w.

~ To do this
simply set ~t equal to w
~ and change the smallest magnitude component of ~t to one.
The ~u and ~v follow easily:
~t w
~
~u =
|~t w|
~
~v = w
~ ~u
As an efficiency improvement, you can avoid taking trigonometric functions of
inverse trigonometric functions (e.g. cos arccos ). For example, when n = 1 (a
diffuse distribution), the vector ~a simplifies to
p p
p
~a = (cos (21 ) 2 , sin (21 ) 2 , 1 2 )

2.3.2

Rejection

A rejection method chooses points according to some simple distribution and rejects some of them so that they are in a more complex distribution. There are
several scenarios where rejection is used, and we show several of these by example.
Suppose we want uniform random points within the unit circle. We can first
choose uniform random points (x, y) [1, 1]2 and reject those outside the circle.
If the function r() returns a canonical random number, then the procedure for this
is:
done = false
while (not done)
x = -1 + 2*r()
y = -1 + 2*r()
if (x*x + y*y < 1)
done = true
end while
24

If we want a random number x p and we know that p : [a, b] 7 R, and

that for all x, p(x) < m, then we can generate random points in the rectangle
[a, b] [0, m] and take those where y < p(x):
done = false
while (not done)
x = a + r()*(b-a)
y = r()*m
if (y < p(x))
done = true
end while
A variant of the last methods is common because we can often deal more easily
with boxes than spheres. To pick a random unit vector with uniform directional
distribution, we first pick a random point in the unit sphere and then treat that point
as a direction vector by taking the unit vector in the same direction:
done = false
while (not done)
x = -1 + 2*r()
y = -1 + 2*r()
z = -1 + 2*r()
if ((length2 =x*x + y*y +z*z) <
done = true
end while
length = sqrt(length2)
x /= length
y /= length
z /= length

2.3.3

1)

Metropolis

The Metropolis method uses random mutations to produce a set of samples with a
desired density. This concept is used extensively in the Metropolis Light Transport
algorithm described later in Chapter 9. Suppose we have a random point x0 in a
domain S. Further, suppose for any point x we have a way to generate random
y px . We use the marginal notation px (y) p(x y) to denote this density
function. Now suppose we let x1 be a random point in S selected with underlying
25

density p(x0 x1 ). We generate x2 with density p(x1 x0 ) and so in. In

the limit where we generate an infinite number of samples, it can be proven that
the samples with have some underlying density determined by p regardless of the
initial point x0 .
Now suppose we want to chose p such that the underlying density of samples
we converge to is proportional to a function f (x) where f is a non-negative function with domain S. Further, suppose we can evaluate f but we have little or no
additional knowledge about its properties (such functions are common in graphics). Also suppose we have the ability to make transitions from xi to xi+1 with
underlying density function t(xi xi+1 ). To add flexibility, further suppose we
add the potentially non-zero probability that xi transitions to itself, i.e., xi+1 = xi .
We phrase this as generating a potential candidate y t(xi y) and accepting
this candidate (i.e., xi+1 = y) with probability a(xi y) and rejecting it (i.e.,
xi+1 = xi ) with probability 1 a(xi y). Note that the sequence x0 , x1 , x2 , . . .
will be a random set, but there will be some correlation among samples. They will
still be suitable for Monte Carlo integration or density estimation, but analyzing
the variance of those estimates is much more challenging.
Now suppose that given a transition function t(x y) and a function f (x) we
want to mimic the distribution of, can we use a(y x) such that the points are
distributed in the shape of f , or more precisely:
{x0 , x1 , x2 , . . .} R

f
sf

It turns out this can be forced by making sure the xi are stationary in some strong
sense. If you visualize a huge collection of sample points x, you want the flow
between two points to be the same in each direction. If we assume the density of
points near x and y are proportional to f (x) and f (y) respectively, the the flow in
the two directions as follows should be the same:
(
flow(x y) = kf (x)t(x y)a(x y)
flow(y x) = kf (y)t(y x)a(y x)
where k is some positive constant. Setting these two flows constant gives a constraint on a:
a(y x)
f (x)t(x y)
=
.
a(x y)
f (y)t(y x)
Thus if either a(y x) or a(x y) is known, so is the other. Making them bigger
26

improves the chance of acceptance, so the usual technique is to set the larger of the
two to 1.
An awkward part of using the Metropolis sample generation technique is that it
is hard to estimate how many points are needed before the set of points is good.
Things are accelerated if the first n points are discarded, although choosing n
wisely is non-trivial. Weights can be added if a truly unbiased distribution is desired, as shown later in the context of Metropolis Light Transport.

2.4

Monte Carlo Simulation

For some physical processes, we have statistical models of behavior at a microscopic level from which we attempt to derive an analytic model of macroscopic
behavior. For example, we often think of a luminaire (a light emitting object) as
emitting a very large number of random photons (really pseudo-photons that obey
geometric, rather than physical, optics) with certain probability density functions
controlling the wavelength and direction of the photons. From this a physicist
might use statistics to derive an analytic model to predict how the luminaire distributes its energy in terms of the directional properties of the probability density
functions. However, if we are not interested in forming a general model, but instead
want to know about the behavior of a particular luminaire in a particular environment, we can just numerically simulate the behavior of the luminaire. To do this
we computationally emit photons from the luminaire and keep track of where
the photons go. This simple method is from a family of techniques called Monte
Carlo Simulation and can be a very easy, though often slow, way to numerically
solve physics problems.
The first thing that you might try in generating a highly realistic image is to
actually track simulated photons until they hit some computational camera plane
or were absorbed. This would be very inefficient, but would certainly produce a
correct image, although not necessarily while you were alive. In practice, very
few Monte Carlo simulations model the full physical process. Instead, an analog
process is found that is easier to simulate, but retains all the important behavior of
the original physical process. One of the difficult parts of finding an analog process
is deciding what effects are important.
An analog process that is almost always employed in graphics is to replace
photons with set wavelengths with power carrying beams that have values across
the entire spectrum. If photons are retained as an aspect of the model, then an
27

obvious analog process is one where photons whose wavelengths are outside of the
region of spectral sensitivity of the film do not exist.

2.5 Density Estimation

Often in graphics we have a set of random points {x0 , x1 , . . . , xn1 } on some
domain S and we wish to infer a plausible underlying density function p(x). These
problems usually arise in photon tracing applications, although one can also view
Metropolis Light Transports final screen reconstruction as density estimation in
screen space.
If the underlying form of the density is known, e.g., p(x) = ax + b on the
interval [0, 1] then this becomes a classic problem based on error metric such as
a least squares fit. This is known as parametric density estimation because the
density is a known parametric form.
In graphics we rarely have a parametric form for p. However, we do assume
p is smooth and a variety of techniques then exist estimating a smooth p that is
representative of the points.

28

Chapter 3

Direct Lighting via Monte Carlo

Integration
By Peter Shirley

In this chapter we apply Monte Carlo Integration to compute the lighting at a point
from an area light source in the presence of potential occluders. As the number of
samples in the pixel becomes large, the number if samples on the light will become
large as well. Thus the shadows will become smooth. Much of the material in this
chapter is from the book Realistic Ray Tracing and is used with permission from
the publisher AK Peters.

3.1

Mathematical framework

To calculate the direct light from one luminaire (light emitting object) onto a diffuse surface, we solve the following equations:
Z
R(x)
L(x) = Le (x) +
Le (x,
~ 0 ) cos d,
(3.1)

all
~0
where L(x) is radiance (color) of x, Le (x) is the light emitted at x, R(x) is the
reflectance of the point,
~ 0 is the direction the light is incident from, and is the
angle between the incident light and the surface normal. Suppose we wanted to
restrict this integral to a domain of one luminaire. Instead of all
~ 0 we would
need to integrate over just the directions toward the luminaire. In practice this can
be hard (the projection of a polygon onto the hemisphere is a spherical polygon,
29

x
dA
luminaire

Figure 3.1: Integrating over the luminaire. Note that there is a direct correspondence between dx, the differential area on the luminaire, and d, the area of the
projection of dx onto the unit sphere centered at x.
and the projection of a cylinder is stranger still). So we can change variable to
integrate over just area (Figure 3.1). Note that the differential relationship exists:
d =

dA cos 0
.
kx0 xk2

(3.2)

Just plugging those relationships into Equation 3.1 gives:

Z
R(x)
dA cos 0
L(x) = Le (x) +
Le (x0 ) cos 0
.

kx xk2
all x0
There is an important flaw in the equation above. It is possible that the points x and
x cannot see each other (there is a shadowing object between them). This can
be encoded in a shadow function s(x, x0 ) which is either one or zero depending
on whether or not there is a clear line of sight between x and x. This gives us the
equation:
Z
s(x, x0 )dA cos 0
R(x)
Le (x0 ) cos
L(x) = Le (x) +
.
(3.3)

kx0 xk2
all x0
If we are to sample Equation 3.3, we need to pick a random point x on the surface
of the luminaire with density function p (so x0 p). Just plugging into the Monte
Carlo equation with one sample gives:
L(x) Le (x) +

R(x)
s(x, x0 ) cos 0
Le (x0 ) cos
.

p(x0 )kx0 xk2

30

(3.4)

If we pick a uniform random point on the luminaire, then p = 1/A, where A is the
area of the luminaire. This gives:
L(x) Le (x) +

A s(x, x0 ) cos 0
R(x)
Le (x0 ) cos
.

kx0 xk2

(3.5)

We can use Equation 3.5 to sample planar (e.g. rectangular) luminaires in a straightforward fashion. We simply pick a random point on each luminaire. The code for
one luminaire would be:
spectrum directLight( x, ~n)
pick random point x with normal vector ~n0 on light
d~ = (x0 x)
if ray x + td~ hits at x0 then
~ n0 d)/k
~ dk
~ 4
return ALe (x0 )(~n d)(~
else
return 0
The above code needs some extra tests such as clamping the cosines to zero if they
~ 4 comes from the distance squared term and
are negative. Note that the term kdk
~ cos because d~ is not necessarily a unit vector.
the two cosines, e.g., ~n d~ = kdk
Several examples of soft shadows are shown in Figure 3.2.

3.2

Sampling a spherical luminaire

Although a sphere with center c and radius r can be sampled using Equation 3.5,
this will yield a very noisy image because many samples will be on the back of the
sphere, and the cos 0 term varies so much. Instead we can use a more complex
p(x0 ) to reduce noise. The first nonuniform density we might try is p(x0 ) cos 0 .
This turns out to be just as complicated as sampling with p(x0 ) cos 0 /kx0 xk2 ,
so we instead discuss that here. We observe that sampling on the luminaire this
way is the same as using a density constant function q(~
0 ) = const defined in
the space of directions subtended by the luminaire as seen from x. We now use
a coordinate system defined with x at the origin, and a right-handed orthonormal
basis with w
~ = (c x)/kc xk, and ~v = (w
~ ~n)/k(w
~ ~n)k (see Figure 3.3).
We also define (, ) to be the azimuthal and polar angles with respect to the uvw
coordinate system.
31

Figure 3.2: Various soft shadows on a backlit sphere with a square and a spherical
light source. Top: one sample. Bottom: 100 samples. Note that the shape fof the
light source is less important than its size in determining shadow appearance.

32

luminaire

n
x

r
c

u
x

max

Figure 3.3: Geometry for spherical luminaire.

The maximum that includes the spherical luminaire is given by:
s



2
r
r
max = arcsin
= arccos 1
.
kx ck
kx ck
Thus a uniform density (with respect to solid angle) within the cone of directions
subtended by the sphere is just the reciprocal of the solid angle 2(1 cos max )
subtended by the sphere:
1

q() =

r
2 1

r
kxck

2

!.

And we get


cos

= 1 1 + 1

r
kxck

2
.

22

This gives us the direction to x. To find the actual point, we need to find the first
point on the sphere in that direction. The ray in that direction is just (x+t~a), where
~a is given by:

cos sin
ux vx wx
~a = uy vy wy sin sin .
uz vz wz
cos
33

Figure 3.4: A sphere with Le = 1 touching a sphere of reflectance 1. Where they

touch the reflective sphere should have L(x) = 1. Left: one sample. Middle: 100
samples. Right: 100 samples, close-up.
We must also calculate p(x0 ), the probability density function with respect to the
area measure (recall that the density function q is defined in solid angle space).
Since we know that q is a valid probability density function using the measure,
and we know that ddir = dA(x0 ) cos 0 /kx0 xk2 , we can relate any probability
density function q(~
0 ) with its associated probability density function p(x0 ):
q(~
0) =

p(x0 ) cos 0
.
kx0 xk2

(3.6)

So we can solve for p(x0 ):

cos 0
r

p(x0 ) =
2kx0 xk2

r
kxck

2

!.

A good debugging case for this is shown in Figure 3.4. For further details on
sampling the sphere with p see the article by Wang [92].

3.3

Non-diffuse Luminaries

There is no reason the brightness of the luminaire cannot vary with both direction
and position. It can vary with position if the luminaire is a television. It can
vary with direction for car headlights and other directional sources. Nothing need
change from the previous sections, except that Le (x0 ) must change to Le (x0 ,
~ 0 ).
The simplest way to vary the intensity with direction is to use a phong-like pattern
34

with respect to the normal vector ~n0 . To keep the total light output independent of
exponent, you can use the form:
Le (x0 ,
~ 0) =

(n + 1)E(x0 ) (n1) 0
cos
,
2

where E(x0 ) is the radiant exitance (power per unit area) at point x0 , and n is the
phong-exponent. You get a diffuse light for n = 1.

3.4

Direct Lighting from Many Luminaires

Traditionally, when NL luminaires are in a scene, the direct lighting integral is

broken into NL separate integrals [14]. This implies at least NL samples must
be taken to approximate the direct lighting, or some bias must be introduced (as
done by Ward where small value samples are not calculated [93]). This is what
you should probably do when you first implement your program. However, you
can later leave the direct lighting integral intact and design a probability density
function over all NL luminaires.
As an example, suppose we have two luminaires, l1 and l2 , and we devise two
probability functions p1 (x0 ) and p2 (x0 ), where pi (x0 ) = 0 for x0 not on li and pi (x0 )
is found by a method such as one of those described previously for generating x0
on li . These functions can be combined into a single density over both lights by
applying a weighted average:
p(x0 ) = p1 (x0 ) + (1 )p2 (x0 ),
where (0, 1). We can see that p is a probability density function because its
integral over the two luminaires is one, and it is strictly positive at all points on the
luminaires. Densities that are mixed from other densities are often called mixture
densities and the coefficients and (1 ) are called the mixing weights [82].
To estimate L = (L1 + L2 ), where L is the direct lighting and Li is the lighting
from luminaire li , we first choose a random canonical pair (1 , 2 ), and use it to
decide which luminaire will be sampled. If 0 1 < , we estimate L1 with e1
using the methods described previously to choose x0 and to evaluate p1 (x0 ), and we
estimate L with e1 /. If 1 then we estimate L with e2 /(1). In either case,
once we decide which source to sample, we cannot use (1 , 2 ) directly because we
have used some knowledge of 1 . So if we choose l1 (so 1 < ), then we choose
a point on l1 using the random pair (1 /, 2 ). If we sample l2 (so 1 ),
35

then we use the pair ((1 )/(1 ), 2 ). This way a collection of stratified
samples will remain stratified in some sense. Note that it is to our advantage to
have 1 stratified in one dimension, as well as having the pair (1 , 2 ) stratified in
two dimensions, so that the li we choose will be stratified over many (1 , 2 ) pairs,
so some multijittered sampling method may be helpful (e.g [7]).
This basic idea used to estimate L = (L1 + L2 ) can be extended to NL luminaires by mixing NL densities
p(x0 ) = 1 p1 (x0 ) + 2 p2 (x0 ) + + NL pNL (x0 ),

(3.7)

where the i s sum to one, and where each i is positive if li contributes to the
direct lighting. The value of i is the probability of selecting a point on the li ,
and pi is then used to determine which point on li is chosen. If li is chosen, the
we estimate L with ei /i. Given a pair (1 , 2 ), we choose li by enforcing the
conditions
i1
i
X
X
j < 1 <
j .
j=1

j=1

And to sample the light we can use the pair (10 , 2 ) where
10 =

Pi1

j=1 j

This basic process is shown in Figure 3.5. It cannot be over stressed that it is
important to reuse the random samples in this way to keep the variance low, in
the same way we use stratified sampling (jittering) instead of random sampling in
the space of the pixel To choose the point on the luminaire li given (10 , 2 ), we can
use the same types of pi for luminaires as used in the last section. The question
remaining is what to use for i .

3.4.1

Constant i

The simplest way to choose values for i was proposed by Lange [45] (and this
method is also implied in the figure on page 148 of [36]), where all weights are
made equal: i = 1/NL for all i. This would definitely make a valid estimator
because the i sum to one and none of them is zero. Unfortunately, in many scenes
this estimate would produce a high variance (when the Li are very different as
occurs in most night walkthroughs).
36

1
1=0

3
1=0.45

Actual 1 chosen means we

pick luminaire 3.

1=0

1=0.75

5
1=1.0

1=0.55

1=0.33

1=1.0

Figure 3.5: Diagram of mapping 1 to choose li and the resulting remapping to

new canonical sample 10 .

3.4.2 Linear i
Suppose we had perfect pi defined for all the luminaires. A zero variance solution
would then result if we could set i Li , where Li is the contribution from the
ith luminaire. If we can make i approximately proportional to Li , then we should
have a fairly good estimator. We call this the linear method of setting i because
the time used to choose one sample is linearly proportional to NL , the number of
luminaires.
To obtain such i we get an estimated contribution ei at x by approximating
the rendering equation for li with the geometry term set to one. These ei s (from all
luminaires) can be directly converted to i by scaling them so their sum is one:
i =

ei
.
e1 + e2 + + eNL

(3.8)

This method of choosing i will be valid because all potentially visible luminaires
will end up with positive i . We should expect the highest variance in areas where
shadowing occurs, because this is where setting the geometry term to one causes
i to be a poor estimate of i .
Implementing the linear i method has several subtleties. If the entire luminaire is below the tangent plane at x, then the estimate for ei should be zero. An
37

easy mistake to make is to set ei to zero if the center of the luminaire is below the
horizon. This will make i take the one value that is not allowed: an incorrect zero.
Such a bug will become obvious in pictures of spheres illuminated by luminaires
that subtend large solid angles, but for many scenes such errors are not noticeable
(the figures in [68] had this bug, but it was not noticeable). To overcome this problem, we make sure that for a polygonal luminaires all of its vertices are below the
horizon before it is given a zero probability of being sampled. For spherical luminaires, we check that the center of the luminaire is a distance greater than the
sphere radius under the horizon plane before it is given a zero probability of being
sampled.

38

Chapter 4

Stratified Sampling of
2-Manifolds
By Jim Arvo

4.1

Introduction

Monte Carlo techniques arise in image synthesis primarily as a means to solve

integration problems. Integration over domains of two or higher dimensions is
ubiquitous in image synthesis; indirect global illumination is expressed as an integral over all paths of light, which entails numerous direct illumination problems
such as the computation of form factors, visibility, reflected radiance, subsurface
scattering, and irradiance due to complex or partially occluded luminaires, all of
which involve integration.
The Monte Carlo method stems from a very natural and immediate connection between integration and expectation. Every integral, in both deterministic and
probabilistic contexts, can be viewed as the expected value (mean) of a random
variable; by averaging over many samples of the random variable, we may thereby
approximate the integral. However, there are infinitely many random variables that
can be associated with any given integral; of course, some are better than others.
One attribute that makes some random variables better than others for the purpose of integration is the ease with which they can be sampled. In general, we tend
to construct Monte Carlo methods using only those random variables with convenient and efficient sampling procedures. But there is also a competing attribute.
One of the maxims of Monte Carlo integration is that the probability density func39

tion of the random variable should mimic the integrand as closely as possible. The
closer the match, the smaller the variance of the random variable, and the more
reliable (and efficient) the estimator. In the limit, when the samples are generated
with a density that is exactly proportional to the (positive) integrand, the variance
of the estimator is identically zero [64]. That is, a single sample delivers the exact
answer with probability one.
Perhaps the most common form of integral arising in image synthesis is that
expressing either irradiance or reflected radiance at a surface. In both cases, we
must evaluate (or approximate) an integral over solid angle, which is of the form
Z
f (~
)(~
~n) d,
(4.1)
S

where S is subset of the unit sphere,

~ is a unit direction vector, and ~n is the surface
normal vector, or over surface area, which is of the form
Z
(x ~n)(x ~n0 )
dx,
(4.2)
f (x)
||x||2
A
where A is a surface and x is a point in IR3 . Technically, these integrals differ only
by a change of variable that results from the pullback of surface differentials to
solid angle differentials [1]. The function f may represent the radiance or emissive
power of a luminaire (in the case of irradiance) or it may include a BRDF (in the
case of reflected radiance). In all cases, visibility may be included in the function
f , which can make the integrand arbitrarily discontinuous, thereby drastically reducing the effectiveness of standard numerical quadrature methods for computing
the integral.
To apply Monte Carlo integration most effectively in image synthesis, we seek
sampling algorithms that match the geometries and known light distributions found
in a simulated scene to the extent possible. Consequently, a wide assortment of
sampling algorithms have been developed for sampling both the surfaces of and the
solid angles subtended by various scene geometries [71] so that both forms of the
integral above can be accommodated. In this chapter we will see how to construct
random variables for specific geometries: that is, random variables whose range
coincides with some bounded region of the plane or some bounded surface in IR3 ,
and whose probability density function is constant. For instance, we will see how
to generate uniformly distributed samples over both planar and spherical triangles,
and projected spherical polygons.
40

Figure 4.1: A spherical triangle with uniform samples (left) and stratified samples (right).
Both sets of samples were generated using an area-preserving parametrization for spherical triangles, which we derive below.

All of the sampling algorithms that we construct are based on mappings from
the unit square, [0, 1] [0, 1], to the regions or surfaces in question that preserve
uniform sampling. That is, uniformly distributed samples in the unit square are
mapped to uniformly distributed samples in the range. Such mappings also preserve stratification, also known as jitter sampling [13], which means that uniform
partitionings of the unit square map to uniform partitionings of the range. The ability to apply stratified sampling over various domains is a great advantage, as it is
often a very effective variance reduction technique. Figure 4.1 shows the result of
applying such a mapping to a spherical triangle, both with and without stratified
(jittered) sampling. Figure 4.2 shows the result of applying such a mapping to a
projected spherical polygon, so that the samples on the original spherical polygon
are cosine-distributed rather than uniformly distributed.
All of the resulting algorithms depend upon a source of uniformly distributed
random numbers in the interval [0, 1], which we shall assume is available from
some unspecified source: perhaps drand48, or some other pseudo-random number generator.

4.2

A Recipe for Sampling Algorithms

Although there is a vast and mature literature an Monte Carlo methods, with many
texts describing how to derive sampling algorithms for various geometries and
density functions (see, for example, Kalos and Whitlock [37], Spanier and Gel41

Figure 4.2: A projected spherical polygon with uniform samples (left) and stratified samples (right). Projection onto the plane results in more samples near the north pole of
the sphere than near the equator. Both sets of samples were generated using an areapreserving parametrization for spherical polygons, which we derive below.


area-preserving
parametrization


unit square

2-manifold

(1,1)

M


w
warp

2


t


Ms

parametrization

Figure 4.3: An arbitrary parametrization for a 2-manifold M can be converted into an

area-preserving parametrization, which is useful for uniform and stratified sampling, by
composing it with a warping function. The warping function can be derived directly from
by following a precise procedure.

bard [77], or Rubinstein [64]), these treatments do not provide step-by-step instructions for deriving the types of algorithms that we frequently require in computer
graphics. In this section we present a detailed recipe for how to convert an arbitrary parametrization : [0, 1]2 M, from the unit square to a 2-manifold, into
an area-preserving parametrization, : [0, 1]2 M . That is, a mapping with
the property that
area(A) = area(B) = area([A]) = area([B]),

(4.3)

for all A, B [0, 1] [0, 1]. Such mappings are used routinely in image synthesis
to sample surfaces of luminaires and reflectors. Note that may in fact shrink or
magnify areas, but that all areas undergo exactly the same scaling; hence, it is areapreserving in the strictest sense only when area(M) = 1. A parametrization with
42

this property will allow us to generate uniformly distributed and/or stratified samples over M by generating samples with the desired properties on the unit square
(which is trivial) and then mapping them onto M. We shall henceforth consider
area-preserving parametrizations to be synonymous with sampling algorithms.
Let M represent a shape that we wish to generate uniformly distributed samples on; in particular, M may be any 2-manifold with boundary in IRn , where n is
typically 2 or 3. The steps for deriving a sampling algorithm for M are summarized in Figure 4.4. These steps apply for all dimensions n 2; that is, M may be
a 2-manifold in any space.
Step 1 requires that we select a smooth bijection from [0, 1] [0, 1] to the
2-manifold M. Such a function is referred to as a parametrization and its inverse
is called a coordinate chart, as it associates unique 2D coordinates with (almost)
all points of M. In reality, we only require to be a bijection almost everywhere;
that is, on all but a set of measure zero, such as the boundaries of M or [0, 1]
[0, 1]. In theory, any smooth bijection will suffice, although it may be impractical
or impossible to perform some of the subsequent steps in Figure 4.4 symbolically
(particularly step 4) for all but the simplest functions.
Step 2 defines a function : [0, 1]2 IR that links the parametrization to
the notion of surface area on M. More precisely, for any region A [0, 1]2 , the
function satisfies
Z
= area([A]).
(4.10)
A

That is, the integral of over any region A in the 2D parameter space is the surface
area of the corresponding subset of M under the mapping . Equation (4.4) holds
for all n 2. For the typical cases of n = 2 and n = 3, however, the function
can be expressed more simply. For example, if M is a subset of IR2 then


(s, t) det D(s,t) ,
(4.11)
where D(s,t) is 2 2 Jacobian matrix of at the point (s, t). On the other hand,
if M is a subset of IR3 , then
(s, t) || s (s, t) t (s, t) || ,

(4.12)

which is a convenient abbreviation for equation (4.4) that holds only when n = 3,
as the two partial derivatives of are vectors in IR3 in this case. Non-uniform sampling can also be accommodated by including a weighting function in the definition
of in step 2.
43

1. Select a parametrization for the 2-manifold M IRn . That is,

select a smooth bijection (diffeomorphism) : [0, 1]2 M.
2. Define the function : [0, 1]2 IR by
q
(s s )(t t ) (s t )2
where s =

n
1
s , . . . , s

(4.4)

is the vector of partial derivatives.

3. Define two cumulative distribution functions on [0, 1] by

R1Rs
F (s)

R01 R01
0

Rt
Gs (t)

R 01
0

(u, v) du dv
(u, v) du dv

(s, v) dv
(s, v) dv

(4.5)

(4.6)

4. Invert the two cumulative distribution functions

f (z) F 1 (z)
g(z1 , z2 )

G1
f (z1 ) (z2 )

(4.7)
(4.8)

5. Define the new parametrization : [0, 1]2 M by

(z1 , z2 ) (f (z1 ), g(z1 , z2 )).

(4.9)

Thus, the mapping (z1 , z2 ) 7 (s, t) = (f (z1 ), g(z1 , z2 )) defines the

warping function that converts the original parametrization into
the area-preserving parametrization .
Figure 4.4: Five steps for deriving an area-preserving parametrization from [0, 1]
[0, 1] to any bounded 2-manifold M IRn , beginning from an arbitrary parametrization
for M. The function is suitable for stratified sampling of M. Step 2 simplifies in the
common two- and three- dimensional cases. Step 4 is often the only impediment to finding
a closed-form expression for the area-preserving parametrization.

44

Step 3 can often be carried out without the aid of an explicit expression for .
For example, the cumulative distributions can often be found by reasoning directly
about the geometry imposed by the parametrization rather than applying formulas (4.4), (4.5) and (4.6), which can be tedious. Let Ms denote the family of
sub-manifolds of M defined by the first coordinate of . That is,
h
i
Ms = [0, s] [0, 1] .
(4.13)
See Figure 4.3. It follows from the definition of F and equation (4.10) that
F (s) =

area(Ms )
,
area(M)

(4.14)

which merely requires that we find an expression for the surface area of Ms as a
function of s. Similarly, by equation (4.7) we have


area(Ms )
s = f
.
(4.15)
area(M)
Thus, f is the map that recovers the parameter s from the fractional area of the
sub-manifold Ms . Equation (4.15) can be more convenient to work with than
equation (4.14), as it avoids an explicit function inversion step. While Gs () and
g(, ) do not admit equally intuitive interpretations, they can often be determined
from the general form of , since many of the details vanish due to normalization. A good example of how this can be done is provided by the area-preserving
parametrization derived for spherical triangles, which we discuss below.
Step 4 above is the only step that is not purely mechanical, as it involves function inversion. When this step can be carried out symbolically, the end result is
a closed-form area-preserving transformation from [0, 1]2 to the manifold M.
Closed-form expressions are usually advantageous, both in terms of simplicity and
efficiency. Of the two inversions entailed in step 4, it is typically equation (4.5)
that is the more troublesome, and frequently resists symbolic solution. In such a
case, it is always possible to perform the inversion numerically using a root-finding
method such as Newtons method; of course, one must always weigh the cost of
drawing samples against the benefits conferred by the resulting importance sampling and stratification. When numerical inversion is involved, the area-preserving
transformation is less likely to result in a net gain in efficiency.
The steps outlined in Figure 4.4 generalize very naturally to the construction of
volume-preserving parametrizations for arbitrary k-manifolds. For any 2 k n,
45

step 3 entails a sequence of k cumulative distribution functions, each dependent

upon all of its predecessors, and step 4 requires the cascaded inversion of all k
distributions, in the order of their definition. Step 5 entails a k-way function composition. In the remainder of these notes, we will consider only the case where
k = 2 and n {2, 3}; that is, we will only consider the problem of generating
samples over 2-manifolds (surfaces) in IR2 or IR3 .

4.3

Analytic Area-Preserving Parametrizations

In this section we will apply the recipe given in Figure 4.4 to derive a number
of useful area-preserving parametrizations. Each will be expressed in closed form
since the functions F and Gs will be invertible symbolically; however, in the case
of spherical triangles it will not be trivial to invert F .

4.3.1

Sampling Planar Triangles

As a first example of applying the steps in Figure 4.4, we shall derive an areapreserving parametrization for an arbitrary triangle ABC in the plane. We begin
with an obvious parametrization from [0, 1] [0, 1] to a given triangle in terms of
barycentric coordinates. That is, let
(s, t) = (1 s)A + s(1 t)B + stC.

(4.16)

It is easy to see that is a smooth mapping that is bijective except when t = 0,

which is a set of measure zero. Since the codomain of is IR2 , is simply the
Jacobian of . After a somewhat tedious computation, we obtain
det(D) = 2cs,

(4.17)

where c is the area of the triangle. From equations (4.5) and (4.6) we obtain
F (s) = s2

Gs (t) = t.

and

(4.18)

In both cases the constant c disappears due to normalization. These functions are
trivial to invert, resulting in
f (z) =

and

g(z1 , z2 ) = z2 .
46

(4.19)

SamplePlanarTriangle( real 1 , real 2 )

Compute
the warping function (1 , 2 ) 7 (s, t).
s 1 ;
t 2 ;
Plug the warped coords into the original parametrization.
P (1 s)A + s(1 t)B + stC;
return P;
end
Figure 4.5: Algorithm for computing an area-preserving parametrization of the triangle
with vertices A, B, and C. This mapping can be used for uniform or stratified sampling.

Finally, after function composition, we have

(z1 , z2 ) = (f (z1 ), g(z1 , z2 ))

= (1 z1 )A + z1 (1 z2 )B + z1 z2 C.

(4.20)

Figure 4.5 shows the final algorithm. If 1 and 2 are independent random variables, uniformly distributed over the interval [0, 1], then the resulting points will be
uniformly distributed over the triangle.

4.3.2

Sampling the Unit Disk

Next, we derive an area-preserving parametrization for a unit-radius disk D in the

plane, centered at the origin. We start with the parametrization from [0, 1] [0, 1]
to D given by
(s, t) = s [ cos(2t)x + sin(2t)y ] ,

(4.21)

where x and y are the orthogonal unit vectors in the plane. Again, is a smooth
mapping that is bijective except when s = 0. Computing the Jacobian of , we
obtain
det(D) = 2s.

(4.22)

The remaining steps proceed precisely as in the case of the planar triangle; in fact,
the distributions F and G turn out to be identical. Thus, we obtain

(z1 , z2 ) = ( z1 , z2 )
p
=
1 [cos(22 )x + sin(22 )y] .
(4.23)
47

The resulting algorithm for sampling the unit disk is exactly analogous to the algorithm shown in Figure 4.5 for sampling planar triangles.

4.3.3

Sampling the Unit Hemisphere

As a first example of applying the steps in Figure 4.4 to a surface in IR3 , we shall
derive the well-known area-preserving parametrization for the unit-radius hemisphere centered at the origin. First, we define a parametrization using spherical
coordinates:

sin s
2 cos(2t)

(s, t) = sin s
(4.24)
2 sin(2t) .


s
cos 2
Here the parameter s defines the polar angle and t defines the azimuthal angle.
Since the codomain of is IR3 , we can apply equation (4.12). Since

cos s
sin s
2 cos(2t)
2 sin(2t)

s
s (s, t) t (s, t) = 2 cos s
2 sin(2t) sin 2 cos(2t) ,


sin s
0
2
we obtain
(s, t) = || s (s, t) t (s, t) ||
 s 
= 2 sin
.
2

(4.25)

It then follows easily that

F (s) = 1 cos

 s 

and

Gs (t) = t,

which are trivial to invert, resulting in

f (z) =

2 cos1 (1 z)

and

g(z1 , z2 ) = z2 .

Composing f and g with results in

p
z1 (2 z1 ) cos(2z2 )
p
(z1 , z2 ) =
z1 (2 z1 ) sin(2z2 )

1 z1
48

(4.26)

Here the s coordinate of the parametrization simply selects the z-plane from z = 1
and z = 0, while the t coordinate parameterizes the resulting circle in the z-plane.
The form of can be simplified somewhat by substituting 1 z1 for z1 , which
does not alter the distribution.

4.3.4

Sampling a Phong Lobe

Now suppose that we wish to sample the hemisphere according to a Phong distribution rather than uniformly; that is, with a density proportional to the cosine of
the polar angle to a power. To do this we simply include a weighting function in
the definition of given in equation (4.25). That is, we let
 s 
 s 
cosk
,
(4.27)
(s, t) = 2 sin
2
2
where k is the Phong exponent. It follows that
 s 
F (s) = cosk+1
and
2
which implies that
f (z) =

1
2
cos1 z n+1

Gs (t) = t,

g(z1 , z2 ) = z2 .

and

It follows that
q

2
k+1

q1 z1 cos(2z2 )

2
(z1 , z2 ) =

1 z1k+1 sin(2z2 )

(4.28)

z1k+1

4.3.5

Sampling Spherical Triangles

We shall now derive an area-preserving parametrization for an arbitrary spherical

triangle, which is significantly more challenging than the cases weve considered
thus far. Let T denote the spherical triangle with vertices A, B, and C, as shown
in Figure 4.6. Such a triangle can be parameterized by using the first coordinate
s to select the edge length bs , which in turn defines sub-triangle Ts T, and the
second coordinate t to select a point along the edge BCs , as shown in Figure 4.6.
This parameterization can be expressed as
(s, t) = slerp( B, slerp( A, C, s ), t ) ,
49

(4.29)

B
c

s
as

bs

Cs

Figure 4.6: Parameter s controls the edge length bs , which determines the vertex Cs , and
consequently sub-triangle Ts . Parameter t then selects a point P along the arc between Cs
and B. Not shown is the length of the arc AC, which is b.

where slerp(A, C, s) is the spherical linear interpolation function that generates

points along the great arc connecting A and C (according to arc length) as s varies
from 0 to 1. The slerp function can be defined as
slerp(x, y, s) = x cos(s) + [ y | x ] sin(s),

(4.30)

where = cos1 x y, and [ y | x ] denotes the normalized component of the vector

y that is orthogonal to the vector x; that is
[y|x]

(I xxT ) y
,
||(I xxT ) y ||

(4.31)

where x is assumed to be a unit vector. From the above definition of it is now

possible to derive the function using equation (4.12). We find that is of the
form
(s, t) = h(s) sin( as t )

(4.32)

for some function h, where as is the length of the moving edge BCs as a function of
s. The exact nature of h is irrelevant, however, as it will not be needed to compute
F , and it is eliminated from Gs by normalization. Thus, we have
F(s) =
Gs ( t ) =

area( Ts )
area(T)

(4.33)

1 cos( as t )
.
1 cos as

(4.34)

50

It follows immediately from inversion of equation (4.34) that

h
i
1
g(z1 , z2 ) =
cos1 1 z2 (1 cos az1 ) .
az1

(4.35)

However, solving for f , which is the inverse of F , is not nearly as straightforward.

Our approach will be to derive an expression for the function f directly, using
equation (4.15), rather than starting from F and inverting it. To do this, we require
several elementary identities from spherical trigonometry. Let A denote the surface
area of the spherical triangle T with vertices A, B and C. Let a, b, and c denote the
edge lengths of T,
a = cos1 B C,
b = cos1 A C,
c = cos1 A B,
and let , , and denote the three internal angles, which are the dihedral angles
between the planes containing the edges. See Figure 4.6. Listed below are a few
well-known identities for spherical triangles:
A=++

(4.36)

sin
sin
sin
=
=
sin a
sin b
sin c

(4.37)

cos = cos cos + sin sin cos a

(4.38)

cos = cos cos + sin sin cos b

(4.39)

cos = cos cos + sin sin cos c

(4.40)

Each of these identities will be employed in deriving area-preserving parametrizations, either for spherical triangles or projected spherical polygons, which will be
described in the following section. Equation (4.36) is known as Girards formula,
equation (4.37) is the spherical law of sines, and equations (4.38), (4.39), and (4.40)
are spherical cosine laws for angles [6].
Our task will be to construct f : [0, 1] IR such that f (As /A) = s, where the
parameter s [0, 1] selects the sub-triangle Ts and consequently determines the
area As . Specifically, the sub-triangle Ts is formed by choosing a new vertex Cs
on the great arc between A and C, at an arc length of bs = sb along the arc from
A, as shown in Figure 4.6. The point P is finally chosen on the arc between B and
Cs , according to the parameter t.
51

SampleSphericalTriangle( real 1 , real 2 )

Use one random variable to select the new area.
As 1 A;
Save the sine and cosine of the angle .
p sin(As );
q cos(As );
Compute the pair (u, v) that determines sin s and cos s .
u q cos ;
v p + sin cos c;
Compute the s coordinate
as normalized arc length from A to Cs .

(v
q

u
p) cos v
1
;
s cos1
b
(v p + u q) sin
Compute the third vertex of the sub-triangle.
Cs slerp(A, C, s) ;
Compute the thcoordinate using Cis and 2 .
t

cos1 1 2 (1 Cs B)

;
cos1 Cs B
Construct the corresponding point on the sphere.
P slerp(B, Cs , t) ;
return P;
end
Figure 4.7: An area-preserving parametrization for an arbitrary spherical triangle ABC.
This procedure can be easily optimized to remove the inverse cosines used to compute the
warped coordinates s and t, since the slerp function uses the cosine of its scalar argument.

To find the parameter s that corresponds to the fractional area As /A, we first
solve for cos bs in terms of As and various constants associated with the triangle.
From equations (4.36) and (4.39) we have
cos bs =
=
=

cos s cos + cos s

sin s sin
cos(As s ) cos + cos s
sin(As s ) sin
cos( s ) cos cos s
,
sin( s ) sin
52

(4.41)

where we have introduced As . We now eliminate s to obtain a function

that depends only on area and the fixed parameters: in particular, we shall construct
a function of only , , and c. We accomplish this by using spherical trigonometry
to find expressions for both sin s and cos s . From equation (4.36) and plane
trigonometry it follows that
cos s = cos( s ) = sin sin s cos cos s .

(4.42)

Combining equation (4.42) with equation (4.40) we have

(cos cos ) cos s + (sin + sin cos c) sin s = 0.

(4.43)

Consequently, sin s = ru and cos s = rv where

u cos cos ,
v sin + sin cos c,
and r is a common factor that cancels out in our final expression, so it is irrelevant.
Simplifying equation (4.41) using these new expressions for sin s and cos s , we
obtain an expression for cos bs in terms of , u, v, and . It then follows that


1
1 ( v cos u sin ) cos v
s =
cos
,
(4.44)
b
( v sin + u cos ) sin
since s = bs /b. Note that cos bs determines bs , since 0 < bs < , and that bs
in turn determines the vertex Cs . The algorithm shown in Figure 4.7 computes an
area-preserving map from the unit square onto the triangle T; it takes two variables
1 and 2 , each in the unit interval, and returns a point P T IR3 . If 1 and
2 are uniformly distributed random variables in [0, 1], the algorithm will produce
a random variable P that is uniformly distributed over the surface of the spherical
triangle T.
The procedure in Figure 4.7 explicitly warps the coordinates (1 , 2 ) into the
coordinates (s, t) in such a way that the resulting parametrization is area-preserving.
If implemented exactly as shown, the procedure performs a significant amount of
unnecessary computation. Most significantly, all of the inverse cosines can be
eliminated by substituting the equation (4.30) for the slerp function and then simplifying [3]. Also, cos , sin , cos c, and [ C | A ], which appears in the expression
for slerp(A, C, s), need only be computed once per triangle rather than once per
sample.
53

Results of the algorithm are shown in Figure 4.1. On the left, the samples are
identically distributed, which produces a pattern equivalent to that obtained by rejection sampling; however, each sample is guaranteed to fall within the triangle.
The pattern on the right was generated by partitioning the unit square into a regular
grid and choosing one pair (1 , 2 ) uniformly from each grid cell, which corresponds to stratified sampling. The advantage of stratified sampling is evident in the
resulting pattern; the samples are more evenly distributed, which generally reduces
the variance of Monte Carlo estimates based on these samples. The sampling algorithm can be applied to spherical polygons by decomposing them into triangles and
performing stratified sampling on each component independently, which is analogous to the method for planar polygons described by Turk [84]. This is one means
of sampling the solid angle subtended by a polygon. We discuss another approach
in the following section.

4.4

Sampling Projected Spherical Polygons

In this section we will see an example in which the inversion of the F function
can not be done symbolically; consequently, we will resort to either approximate
inversion, or inversion via a root finder.
The dot product
~ ~n appearing in equation (4.1) is the ubiquitous cosine factor that appears in nearly every illumination integral. Since it is often infeasible to
construct a random variable that mimics the full integrand, we settle for absorbing
the cosine term into the sampling distribution; this compromise is a useful special
case of importance sampling. In this section we address the problem of generating
stratified samples over the solid angle subtended by arbitrary polygons, while taking the cosine weighting into account, as shown in Figure 4.2. The combination
of stratification and importance sampling, even in this relatively weak form, can
significantly reduce the variance of the associated Monte Carlo estimator [3, 64].
We now describe a new technique for Monte Carlo sampling of spherical polygons with a density proportional to the cosine from a given axis which, by Nusselts
analogy, is equivalent to uniformly sampling the projection of the spherical polygon onto the z = 0 plane. The technique handles polygons directly, without first
partitioning them into triangles, and is ideally suited for stratified sampling. The
Jacobian of the bijection from the unit square to the polygon can be made arbitrarily close to the cosine density, making the statistical bias as close to zero as desired.
After preprocessing a polygon with n vertices, which can be done in O(n2 log n)
54

A


v4

C


v1

(a)

v3

3


v2

(b)

Figure 4.8: (a) Spherical triangle T . We consider the projected area of triangle T as a
function of , keeping vertices A and B, and angle fixed. (b) Partitioning a spherical
polygon by great circles passing through the poles and the vertices.

time, each sample can be generated in O(n) time.

Let P denote a spherical polygon. To help in defining the mapping : [0, 1]2
P, we first derive several basic expression that pertain to spherical triangles. Let
T be a spherical triangle, and consider the family of sub-triangles shown in Figure 4.8a, where the unit vectors A and B and the internal angle are all fixed, but
the internal angle is allowed to sweep from 0 to the last vertex. Our task in this
section is to express a (which is the length of edge BC) and cos b as functions of
. From equation (4.38) we have
cos a =

cos + cos cos

.
sin sin

(4.45)

Let a , b , and c denote the outward unit normals for each edges of the triangle,
as shown in Figure 4.8a. Then cos = a b , where b can be expressed as
b = c cos +(c A) sin .
Also, from equation (4.37) we have sin = sin c sin / sin a. Therefore,


sin
1
a() = tan
,
c1 cos c2 sin

(4.46)

(4.47)

where the constants c1 and c2 are given by

c1 =

(a c ) cos + 1
,
sin sin c

c2 =
55

(a c A) cos
.
sin sin c

(4.48)

These constants depend only on the fixed features of the triangle, as the vectors a
and c do not depend on . It is now straightforward to find cos b as a function of
, which we shall denote by z(). Specifically,
z() = (B N) cos a() + (D N) sin a(),

(4.49)

where D is a point on the sphere that is orthogonal to B, and on the great circle
through B and C. That is,
D = (I BB T )C.

(4.50)

We now show how to sample an arbitrary spherical polygon according to a cosine

distribution. The function a() will be used to invert a cumulative marginal distribution over the polygon, as a great arc sweeps across the polygon, while z() will
be used to sample one-dimensional vertical slices of the polygon.

4.4.1

The Cumulative Marginal Distribution

We break the problem of computing the bijection : [0, 1]2 P into two parts.
First, we define a sequence of sub-polygons of P in much the same way that we
parameterized the triangle T above; that is, we define P() to be the intersection
of P with a lune 1 whose internal angle is , and with one edge passing through an
extremal vertex of P. Next we define a cumulative marginal distribution F () that
gives the area of polygon P() projected onto the plane orthogonal to N, which is
simply the cosine-weighted area of P. Then F is a strictly monotonically increasing function of . By inverting this function we arrive at the first component of
our sampling algorithm. That is, if 1 is a uniformly distributed random variable in
[0, 1], and if is given by
= F 1 ( 1 ),

(4.51)

then defines the great circle from which to draw a sample.

To find F , we first consider the spherical triangle T and its family of subtriangles. The projected area of the triangle T , which we denote by , follows
immediately from Lamberts formula for computing the irradiance from a polygonal luminaire [2]. That is
= (aa + bb + cc ) N,
1

A lune is a spherical triangle with exactly two vertices, which are antipodal.

56

(4.52)

where a , b , and c are outward normals of the triangle T , as shown in Figure 4.8a. If we now constrain T to be a polar triangle, with vertex A at the pole of
the hemisphere (A = N), then becomes a very simple function of . Specifically,
() = a() (a N),

(4.53)

where a N is fixed; this follows from the fact that both b and c are orthogonal
to N. Equation (4.53) allows us to easily compute the function F () for any collection of spherical polygons whose vertices all lie on the lune with vertices at A
and A as shown in Figure 4.10, where we restrict our attention to the positive or
upper half of the lune. Thus,
F () =

k
X

i ai ( k ),

(4.54)

i=1

for [k , k+1 ], where the constants 1 , 2 , . . . , k account for the slope and
orientation of the edges; that is, edges that result in clockwise polar triangles are
positive, while those forming counter-clockwise triangles are negative.
We now extend F () to a general spherical polygon P by slicing P into lunes
with the above property; that is, we partition P into smaller polygons by passing
a great arc through each vertex, as shown in Figure 4.8b. Then for any spherical polygon, we can evaluate F () exactly for any value of by virtue of equation (4.47). The resulting function F is a piecewise-continuous strictly monotonically increasing function with at most n 2 discontinuities, where n is the number
of vertices in the polygon. See Figure 4.9. This function is precisely the cumulative marginal distribution function that we must invert to perform the first stage
of cosine-weighted sampling. Because it is monotonically increasing, its inverse is
well-defined.

4.4.2

The Sampling Algorithm

Given two variables 1 and 2 in the interval [0, 1] we will compute the correspond as
ing point P = (1 , 2 ) in the polygon P. We use 1 to determine an angle ,
described above, and 2 to select the height z according to the resulting conditional

density defined along the intersection of the polygon P and the great circle at .
To compute using equation (4.51), we proceed in two steps. First, we find the
lune from which will be drawn. This corresponds to finding the integer k such
57

3
2
1

Figure 4.9: The cumulative marginal distribution function F as a function of the angle
. At each value of i , the abscissa is the form factor from the origin to the polygon that
is within the range [1 , i ]. This function is strictly monotonically increasing, with at most
n 2 derivative discontinuities, where n is the number of vertices in the polygon. The
fluctuations in F have been greatly exaggerated for purposes of illustration.

z3
z
z3

3
2
1

z2
2

z2
z1
z1

1
k

k+1

Figure 4.10: On the left is an illustration of a single lune with a collection of arcs passing
through it, and the points at which a great circle at intersects them. On the right is a
cross-section of the circle, showing the heights z 1 , z 1 , corresponding to these intersection
points.

that
k
k+1
1
.
tot
tot

(4.55)

Next, we must invert F as it is defined on this interval. Given the nature of F , as

defined in equations (4.47) and (4.54), it is unlikely that this can be done symboli58

cally in general, so we seek a numerical approximation. This is the only step in the
algorithm which is not computed exactly; thus, any bias that is introduced in the
sampling is a result of this step alone.
Approximate numerical inversion is greatly simplified by the nature of F within
each lune. Since F is extremely smooth and strictly monotonic, we can approximate F 1 directly to high accuracy with a low-order polynomial. For example, we
may use
F 1 (x) a + bx + cx2 + dx3 ,

(4.56)

where we set

a
b

= V 1
c

k
k + 1
.
k + 2
k+1

(4.57)

Here V is the Vandermonde matrix formed from F (k ), F (k + 1 ), F (k + 2 ),

and F (k+1 ). Coupling this approximation with a Newton iteration can, of course,
compute the function inverse to any desired numerical accuracy, making the bias
effectively zero. However, such a high degree of accuracy is not warranted for a
typical Monte Carlo simulation.
Once the angle has been computed using 1 , we then compute z using 2 .
This corresponds to sampling z according to the conditional density function cor This conditional density function is defined on the
responding to the choice of .
intervals
[z 1 , z 1 ] [z 2 , z 2 ] [z n , z n ],
which correspond to the intersection points shown in Figure 4.10. These intervals
are computed using equation (4.49). The conditional density is proportional to
z 2 within these intervals, which distributes the samples vertically according to the
cosine of the angle from the pole. The most costly part of sampling according to
this density is normalization. We define
Zj

j
X

z 2i z 2i .

(4.58)

i=1

Then Zn is the normalization constant. The random variable 2 then selects the
interval by finding 1 ` n such that
Z`
Z`1
2
Zn
Zn
59

(4.59)

Figure 4.11: A non-convex spherical polygon with cosine-weighted samples generated

with the proposed mapping.

Figure 4.12: A non-convex spherical polygon with cosine-weighted and stratified samples
generated with the proposed mapping.

where Z0 0. Finally, the height of P = (1 , 2 ) is

z =

2 Z`1 + z ` ,

(4.60)

and the point itself is

sin ,
z),
P = ( cos ,

(4.61)

where = 1 z2 .
The algorithm described above also works for spherical polygons P that surround the pole of the sphere. In this case, each lune has an odd number of segments
crossing it, and z n = 1 must be added to the list of heights defined by each in
sampling from the conditional distribution.
The algorithm described above is somewhat more costly than the algorithm for
uniform sampling of spherical triangles [3] for two reasons: 1) evaluating piecewise continuous functions requires some searching, and 2) the cumulative marginal
60

distribution cannot be inverted exactly. Furthermore, the sampling algorithm requires some preprocessing to make both of these operations efficient and accurate.
Pre-processing includes partitioning the polygon into lunes, computing the
constants c1 and c2 defined in equation (4.48) for each resulting edge, and sorting the line segments within each lune into increasing order. In the worst case,
there may be n 2 lunes, with (n) of them containing (n) segments. Thus,
creating them and sorting them requires O(n2 log n) in the worst case. For convex
polygons, this drops to O(n), since there can be only two segments per lune.
Once the pre-processing is done, samples can be generated by searching for
the appropriate [k , k+1 ] interval, which can be done in O(log n) time, and then
sampling according to the conditional distribution, which can be done on O(n)
time. The latter cost is the dominant one because all of the intervals must be formed
for normalization. Therefore, in the worst case, the cost of drawing a sample is
O(n); however, for convex polygons this drops to O(log n).
Figure 4.2 shows 900 samples in a spherical quadrilateral, distributed according
to the cosine distribution. Note that more of the samples are clustered near the pole
than the horizon. Stratification was performed by mapping jittered points from
the unit square onto the quadrilateral. Figures 4.11 and 4.12 show 900 samples
distributed according to the cosine density within a highly non-convex spherical
polygon. These samples were generated without first partitioning the polygon into
triangles. In both of the test cases, the cumulative marginal distribution function F
is very nearly piecewise linear, and its inverse can be computed to extremely high
accuracy with a piecewise cubic curve.

61

62

Chapter 5

Combining Sampling Strategies

By Jim Arvo

5.1

Introduction

In this chapter we explore the idea of constructing effective random variables for
Monte Carlo integration by combining two or more simpler random variables. For
instance, suppose that we have at our disposal a convenient means of sampling the
solid angle subtended by a luminaire, and also a means of sampling a brdf; how
are these to be used in concert to estimate the reflected radiance from a surface?
While each sampling method can itself serve as the basis of an importance sampling scheme, in isolation neither can reliably predict the shape of the resulting
integrand. The problem is that the shape of the brdf may make some directions
important (i.e. likely to make a large contribution to the integral) while the luminaire, which is potentially orders of magnitude brighter than the indirect illumination, may make other directions important. The question that we shall address
is how to construct an importance sampling method that accounts for all such hot
spots by combining available sampling methods, but without introducing statistical bias. The following discussion closely parallels the work of Veach [90], who
was the first to systematically explore this idea in the context of global illumination.
To simplify the discussion, let us assume that we are attempting to approximate
some quantity I, which is given by the integral of an unknown and potentially illbehaved function f over the domain D:
Z
I =
f (x) dx.
(5.1)
D

63

For instance, f may be the product of incident radiance (direct and indirect), a reflectance function, and a visibility factor, and D may be the either a collection of
surfaces or the hemisphere of incident directions; in cases such as these, I may
represent reflected radiance. In traditional importance sampling, we select a probability density function (pdf) p over D and rewrite the integral as



Z 
f (x)
f (X)
I =
p(x) dx =
,
(5.2)
p(X)
D p(x)
where X denotes a random variable on the domain D distributed according to the
pdf p, and hi denotes the expected value of a random variable. The second equality
in equation (5.2) is simply the definition of expected value. It follows immediately
that the sample mean of the new random variable f (X)/p(X) is an estimator for I;
that is, if
N
1 X f (Xi )
E =
,
N
p(Xi )

(5.3)

i=1

for N 1, where X1 , . . . , XN are iid (independent identically distributed) random

variables, each distributed according to the pdf p, then hEi = I. Consequently,
E I, and the quality of the approximation can be improved by increasing N , the
number of samples, and/or by increasing the similarity between the original integrand f and the pdf p. Since evaluating f (Xi ) is potentially very costly, we wish
to pursue the second option to the extent possible. This is precisely the rationale
for importance sampling.

5.2

Using Multiple PDFs

Now let us suppose that we have k distinct pdfs, p1 , p2 , . . . , pk , that each mimic
some potential hot spot in the integrand; that is, each concentrates samples in a
region of the domain where the integrand may be relatively large. For instance, p1
may sample according to the brdf, concentrating samples around specular directions of glossy surfaces, while p2 , . . . , pk sample various luminaires or potential
specular reflections. Let us further suppose that for each pi we draw Ni iid samples, Xi,1 , Xi,2 , . . . , Xi,Ni , distributed according to pi . Our goal is to combine them
into an estimator E that has several desirable properties. In particular, we wish to
ensure that
1. hEi = I,
64

2. E is relatively easy to compute,

3. var(E) is small.
That is, we wish to have the expected value of E match the actual value of the integral, I, to pay a low computational price for drawing each sample, and to reduce
the variance of E as much as possible, thereby reducing the number of samples
required to attain a reliable approximation. The first requirement ensures that the
estimator is unbiased. Unbiased estimators have the highly desirable property that
they allow us to converge to the exact answer by taking a sufficiently large number
of samples. As a general rule, this is the first property that any random variable
designed for Monte Carlo integration should possess [40, 41].
As we shall see, there is a large family of functions that meet property 1,
leaving much flexibility in choosing one that meets both properties 2 and 3. We
begin by identifying such a class of estimators, then imposing the other constraints.
First, consider an estimator of the form
E

Ni
k X
X

i (Xi,j )f (Xi,j ),

(5.4)

i=1 j=1

for some suitable choice of the functions i . That is, let us allow a different function i to be associated with the samples drawn from each pdf pi , and also allow the
weight of each sample to depend on the sample itself. Equation (5.4) is extremely
general, and also reasonable, as we can immediately ensure that E is unbiased by
constraining the functions i to be of the form
i (x)

wi (x)
,
Ni pi (x)

(5.5)

where for all x D and 1 i k, the weighting functions wi satisfy

wi (x) 0,

(5.6)

w1 (x) + + wk (x) = 1.

(5.7)

To see that the resulting estimator is unbiased, regardless of the choice of the
weighting functions wi , provided that they satisfy constraints (5.6) and (5.7), let
us define Ew to be the estimator E where the i are of the form shown in equa65

tion (5.5), and observe that

hEw i =

* k N
i
XX

+
i (Xi,j )f (Xi,j )

i=1 j=1

Ni
k X
X

hi (Xi,j )f (Xi,j )i

i=1 j=1

Ni Z
k X
X
i=1 j=1

i (x)f (x)pi (x) dx

k Z
X
wi (x)
=
f (x)pi (x) dx
p (x)
i=1 D i
" k
#
Z
X
=
f (x)
wi (x) dx
D

i=1

Z
=

f (x) dx
D

= I.
Thus, by considering only estimators of the form Ew , we may henceforth ignore
property 1 and concentrate strictly on selecting the weighting functions wi so as to
satisfy the other two properties.

5.3

Possible Weighting Functions

In some sense the most obvious weighting functions to employ are given by
wi (x)

ci pi (x)
,
q(x)

(5.8)

where
q(x) c1 p1 (x) + + ck pk (x),

(5.9)

is a pdf obtained by taking a convex combination of the original pdfs; that is, the
constants ci satisfy ci 0 and c1 + + ck = 1. Clearly, these wi are positive
and sum to one at each x; therefore the resulting estimator is unbiased, as shown
above. This particular choice is obvious in the sense that it corresponds exactly
66

to classical importance sampling based on the pdf defined in equation (5.9), when
a very natural constraint is imposed on N1 , . . . , Nk . To see this, observe that
Ew

Ni
k X
X
wi (Xi,j )
f (Xi,j )
=
Ni pi (Xi,j )
i=1 j=1

Ni
k
X
X
1
c
i

=
f (Xi,j )
Ni
q(Xi,j )
i=1
j=1

Ni
k
X
X
f (Xi,j )
ci
=
.
Ni
q(Xi,j )
i=1

(5.10)

j=1

Now, let N = N1 + + Nk be the total number of samples, and let us further

assume that the samples have been partitioned among the pdfs p1 , . . . , pk in proportion to the weights c1 , . . . , ck , that is, with Ni = ci N . Then the ratio ci /Ni is
constant, and equation (5.10) simplifies to
Ew

k Ni
f (Xi,j )
1 XX
=
.
N
q(Xi,j )

(5.11)

i=1 j=1

Note that in equation (5.11) all samples are handled in exactly the same manner;
that is, the weighting of the samples does not depend on i, which indicates the
pdfs they are distributed according to. This is precisely the formula we would
obtain if we began with q as our pdf for importance sampling. Adopting Veachs
terminology, we shall refer to this particular choice of weighting functions as the
balance heuristic [90]. Other possibilities for the weighting functions, which are
also based on convex combinations of the original pdfs, include
(
1 if ci pi (x) = maxj cj pj (x)
wi (x) =
,
(5.12)
0 otherwise
and

1
k
X

,
wi (x) = ci pm
cj pm
i (x)
j (x)

(5.13)

j=1

for some exponent m 1. Again, we need only verify that these weighting functions are non-negative and sum to one for all x to verify that they give rise to
unbiased estimators. Note, also, that each of these strategies is extremely simple to
compute, thus satisfying property 2 noted earlier.
67

5.4 Obvious is also Nearly Optimal

Let Ebw denote an estimator that incorporates the balance heuristic, and let Ew be an
estimator with any other valid choice of weighting function. Veach has shown [90]
that


 
1
1
2
b
var Ew var(Ew ) + I

,
(5.14)
Nmin N
where Nmin = mini Ni . Inequality (5.14) indicates that the variance of the estimator
Ebw compares favorably with the optimal strategy, which would be infeasible to
determine in any case. In fact, as the number of samples of the least-sampled pdf
approaches to infinity, the balance heuristic approaches optimality.
Fortunately, the balance heuristic is also extremely easy to apply; it demands
very little beyond the standard requirements of importance sampling, which include the ability to generate samples distributed according to each of the original
pdfs pi , and the ability to compute the density of a given point x with respect to
each of the original pdfs [41]. This last requirement simply means that for each
x D and 1 i k, we must be able to evaluate pi (x). Thus, Ebw satisfies all
three properties noted earlier, and is therefore a reasonable heuristic in itself for
combining multiple sampling strategies.

68

Chapter 6

Quasi-Monte Carlo Sampling

By Art B. Owen

In Monte Carlo (MC) sampling the sample averages of random quantities are used
to estimate the corresponding expectations. The justification is through the law of
large numbers. In quasi-Monte Carlo (QMC) sampling we are able to get a law
of large numbers with deterministic inputs instead of random ones. Naturally we
seek deterministic inputs that make the answer converge as quickly as possible. In
particular it is common for QMC to produce much more accurate answers than MC
does. Keller [39] was an early proponent of QMC methods for computer graphics.
We begin by reviewing Monte Carlo sampling and showing how many problems can be reduced to integrals over the unit cube [0, 1)d . Next we consider
how stratification methods, such as jittered sampling, can improve the accuracy
of Monte Carlo for favorable functions while doing no harm for unfavorable ones.
Method of multiple-stratification such as Latin hypercube sampling (n-rooks) represent a significant improvement on stratified sampling. These stratification methods balance the sampling points with respect to a large number of hyperrectangular
boxes. QMC may be thought of as an attempt to take this to the logical limit: how
close can we get to balancing the sample points with respect to every box in [0, 1)d
at once? The answer, provided by the theory of discrepancy is surprisingly far,
that the result produces a significant improvement compared to MC. This chapter
concludes with a presentation of digital nets, integration lattices and randomized
QMC.
69

6.1 Crude Monte Carlo

As a frame of reference for QMC, we recap the basics of MC. Suppose that the
average we want to compute is written as an integral
Z
I=
f (x)q(x)dx.
(6.1)
D

The set D Rd is the domain of interest, perhaps a region on the unit sphere
or in the unit cube. The function q is a probability density function on D. That is
R
q(x) 0 and D q(x)dx = 1. The function f gives the quantity whose expectation
we seek: I is the expected value of f (x) for random x with density q on D.
In crude Monte Carlo sampling we generate n independent samples x1 , . . . , xn
from the density q and estimate I by
n

1X
I = In =
f (xi ).
n

(6.2)

i=1

The strong law of large numbers tells us that

Pr lim In = I = 1.
n

(6.3)

That is, crude Monte Carlo always converges to the right answer as n increases
without bound.
R
Now suppose that f has finite variance 2 = Var(f (x)) D (f (x)I)2 q(x)dx.
Then E((In I)2 ) = 2 /n so the root mean square error (RMSE) of MC sam
pling is O(1/ n). This rate is slow compared to that of classical quadrature rules
(Davis and Rabinowitz [16]) for smooth functions in low dimensions. Monte Carlo
methods can improve on classical ones for problems in high dimensions or on discontinuous functions.
A given integration problem can be written in the form (6.1) in many different
ways. First, let p be a probability density on D such that p(x) > 0 whenever
q(x)|f (x)| > 0. Then
Z
Z
f (x)q(x)
I=
f (x)q(x)dx =
p(x)dx
p(x)
D
D
and we could as well sample xi p(x) and estimate I by
n

1 X f (xi )q(xi )
Ip = In,p =
.
n
p(xi )
i=1

70

(6.4)

The RMSE can be strongly affected, for better or worse, by this re-expression,
known as importance sampling. If we are able to find a good p that is nearly
proportional to f q then we can get much better estimates.
Making a good choice of density p is problem specific. Suppose for instance,
that one of the components of x describes the angle = (x) between a ray and a
surface normal. The original version of f may include a factor of cos() for some
> 0. Using a density p(x) q(x) cos() corresponds to moving the cosine
power out of the integrand and into the sampling density.
We will suppose that a choice of p has already been made. There is also the
possibility of using a mixture of sampling densities pj as with the balance heuristic
of Veach and Guibas [88, 89]. This case can be incorporated by increasing the dimension of x by one, and using that variable to select j from a discrete distribution.
Monte Carlo sampling of x p over D almost always uses points from a
pseudo-random number generator simulating the uniform distribution on the interval from 0 to 1. We will take this to mean the uniform distribution on the half-open
interval [0, 1). Suppose that it takes d uniform random variables to simulate a
point in the d dimensional domain D. Often d = d but sometimes d = 2 variables from [0, 1) can be used to generate a point within a surface element in d = 3
dimensional space. In other problems we might use d > d random variables to
generate a p distributed point in D Rd . Chapter 4 describes general techniques
for transforming [0, 1)d into D and provides some specific examples of use in ray
tracing. Devroye [17] is a comprehensive reference on techniques for transforming
uniform random variables into ones desired random objects.

Suppose that a point having the U [0, 1)d distribution is transformed into a
point (x) having the density p on D. Then
Z
I=
D

where
by

f (x)q(x)
p(x)dx =
p(x)

Z
[0,1)d

f ( (x))q( (x))
dx
p( (x))

Z
[0,1)d

f (x)dx

(6.5)
incorporates the transformation and the density q. Then I is estimated
n

1 X f ( (xi ))q( (xi ))

1X
I =
=
f (xi )
n
p( (xi ))
n
i=1

(6.6)

i=1

where xi are independent U [0, 1) random variables.

Equation (6.5) expresses the original MC problem (6.1) as one of integrating
a function f over the unit cube in d dimensions. We may therefore reformulate
71

the problem as finding I =

is the old f .

[0,1)d

f (x)dx. The new d is the old d and the new f

6.2 Stratification
Stratified sampling is a technique for reducing the variance of a Monte Carlo integral. It was originally applied in survey sampling (see Cochran [10]) and has been
adapted in Monte Carlo methods, Fishman [21]. In stratified sampling, the domain
S
T
of x is written as a union of strata D = H
Dk = if j 6= k.
h=1 Dh where Dj
An integral is estimated from within each stratum and then combined. Following
the presentation in chapter 6.1, we suppose here that D = [0, 1)d .
Figure 6.1 shows a random sample from the unit square along with 3 alternative
stratified samplings. The unit cube [0, 1)d is very easily partitioned into box shaped
strata like those shown. It is also easy to sample uniformly in such strata. Suppose
that a, c [0, 1)d with a < c componentwise. Let U U [0, 1)d . Then a + (c
a)U interpreted componentwise is uniformly distributed on the box with lower left
corner a and upper right corner c.
In the simplest form of stratified sampling, a Monte Carlo sample xh1 , . . . xhnh
is taken from within stratum Dh . Each stratum is sampled independently, and the
results are combined as
ISTRAT = ISTRAT (f ) =

nh
H
X
|Dh | X
h=1

nh

f (xhi ),

(6.7)

i=1

where |Dh | is the volume of stratum D.

For any x [0, 1)d let h(x) denote the stratum containing x. That is x Dh(x) .
The mean and variance of f within stratum h are
Z
1
h = |Dh |
f (x)dx, and,
(6.8)
Dh
Z
h2 = |Dh |1
(f (x) h )2 dx
(6.9)
Dh

respectively. We can write E(ISTRAT ) as:

nh
H
X
|Dh | X
h=1

nh

i=1

E(f (xhi )) =

H
X

|Dh |h =

h=1

72

H Z
X
h=1 Dh

f (x)dx = I,

Figure 6.1: The upper left figure is a simple random sample of 16 points in [0, 1)2 .
The other figures show stratified samples with 4 points from each of 4 strata.

so that stratified sampling is unbiased.

The variance of stratified sampling depends on the allocation of sample size nh
to strata. We will suppose that nh is allocated proportionally, so that nh = n|Dh |
for the total sample size n. First we note that when x U [0, 1)d , then h(x) is
a random variable taking the value ` with probability |D` |. Then from a standard
73

variance formula
2 = Var(f (x)) = E(Var(f (x) | h(x))) + Var(E(f (x) | h(x)))
=

H
X

|Dh |h2

h=1

H
X

|Dh |(h I)2 ,

(6.10)
(6.11)

h=1

so that 2 is a sum of contributions from within and between strata. Now

Var(ISTRAT ) =

H
X
|Dh |2
h=1

nh

h2 =

1X
2
|Dh |h2
,
n
n

(6.12)

h=1

from (6.10).
Equation (6.12) shows that stratified sampling with proportional allocation
does not increase the variance. Proportional allocation is not usually optimal. Optimal allocations take nh |Dh |h . If estimates of h are available they can be
used to set nh , but poor estimates of h could result in stratified sampling with
larger variance than crude MC. We will assume proportional allocation.
A particular form of stratified sampling is well suited to the unit cube. Haber [24]
proposes to partition the unit cube [0, 1)d into H = md congruent cubical regions
and to take nh = 1 point from each of them. This stratification is known as jittered
sampling in graphics, following Cook, Porter and Carpenter [14].
Any function that is constant within strata is integrated without error by ISTRAT .
If f is close to such a function, then f is integrated with a small error. Let f be the
function defined by f(x) = h(x) , and define the residual fRES (x) = f (x) f(x).
This decomposition is illustrated in Figure 6.2 for a function on [0, 1). The error
ISTRAT I reduces to the stratified sampling estimate of the mean of fRES . Stratified
sampling reduces the Monte Carlo variance from 2 (f )/n to 2 (fRES )/n.

6.3

Multiple Stratification

Suppose we can afford to sample 16 points in [0, 1)2 . Sampling one point from
each of 16 vertical strata would be a good strategy if the function f depended
primarily on the horizontal coordinate. Conversely if the vertical coordinate is the
more important one, then it would be better to take one point from each of 16
horizontal strata.
It is possible to stratify both ways with the same sample, in what is known as
Latin hypercube sampling (McKay, Beckman and W. J. Conover [51]) or n-rooks
74

1.2
1.0
0.8
0.6
0.4
0.2
0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.6 0.4 0.2

0.0

0.2

0.4

0.6

0.0

Figure 6.2: The upper plot shows a piece-wise smooth function f on [0, 1). The
step function is the best approximation f to f , in mean square error, among functions constant over intervals [j/10, (j + 1)/10). The lower plot shows the difference f f using a vertical scale similar to the upper plot.

sampling (Shirley [67]). Figure 6.3 shows a set of 16 points in the square, that are
simultaneously stratified in each of 16 horizontal and vertical strata.
If the function f on [0, 1)2 is dominated by either the horizontal coordinate
or the vertical one, then well get an accurate answer, and we dont even need to
know which is the dominant variable. Better yet, suppose that neither variable is
75

Figure 6.3: The left plot shows 16 points, one in each of 16 vertical strata. The
right plot shows the same 16 points. There is one in each of 16 horizontal strata.
These points form what is called a Latin hypercube sample, or an n-rooks pattern.

dominant but that

f (x) = fH (x) + fV (x) + fRES (x)

(6.13)

where fH depends only on the horizontal variable, fV depends only on the vertical
one, and the residual fRES is defined by subtraction. Latin hypercube sampling will
give an error that is largely unaffected by the additive part fH + fV . Stein [78]
2 /n
showed that the variance in Latin hypercube sampling is approximately RES
2 is the smallest variance of f
where RES
RES for any decomposition of the form (6.13).
His result is for general d, not just d = 2.
Stratification with proportional allocation is never worse than crude MC. The
same is almost true for Latin hypercube sampling. Owen [58] shows that for all
n 2, d 1 and square integrable f , that
Var(ILHS )

2
.
n1

For the worst f , Latin hypercube sampling is like using crude MC with one observation less.
The construction of a Latin hypercube sample requires uniform random permutations. A uniform random permutation of 0 through n 1 is one for which
76

all n! possible orderings have the same probability. Devroye [17] gives algorithms for such random permutations. One choice is to have an array Ai = i
for i = 0, . . . , n 1 and then for j = n 1 down to 1 swap Aj with Ak where k
is uniformly and randomly chosen from 0 through j.
For j = 1, . . . , d, let j be independent uniform random permutations of
0, . . . , n 1. Let Uij U [0, 1)d independently for i = 1, . . . , n and j = 1, . . . , d
and let X be a matrix with
Xij =

j (i 1) + Uij
.
n

Then the n rows of X form a Latin hypercube sample. That is we may take xi =
(Xi1 , . . . , Xid ). An integral estimate I is the same whatever order the f (xi ) are
summed. As a consequence we only need to permute d 1 of the d input variables.
We can take 1 (i 1) = i 1 to save the cost of one random permutation.
Jittered sampling uses n = k 2 strata arranged in a k by k grid of squares
while n-rooks provides simultaneous stratification in both an n by 1 grid and a 1
by n grid. It is natural to wonder which method is better. The answer depends on
whether f is better approximated by a step function, constant within squares of size
1/k 1/k grid, or by an additive function with each term constant within narrower
bins of width 1/n. Amazingly, we dont have to choose. It is possible to arrange
n = k 2 points in an n-rooks arrangement that simultaneously has one point in
each square of a k by k grid. A construction for this was proposed independently
by Chiu, Shirley and Wang [8] and by Tang [81]. The former handle more general
grids of n = k1 k2 points. The latter reference arranges points in [0, 1)d with
d 2 in a Latin hypercube such that every two dimensional projection of xi puts
one point into each of a grid of strata.

6.4

Uniformity and Discrepancy

The previous sections look at stratifications in which every cell in a rectangular

grid or indeed in multiple rectangular grids gets the proper number of points. It
is clear that a finite number of points in [0, 1)d cannot be simultaneously stratified
with respect to every hyper-rectangular subset of [0, 1)d , yet it is interesting to ask
how far we might be able to go in that direction. This is a problem that has been
studied since Weyl [96] originated his theory of uniform distribution. Kuipers and
Niederreiter [44] summarize that theory.
77

Let a and c be points in [0, 1)d for which a < c holds componentwise, and then
let [a, c) denote the box of points x where a x < c holds componentwise. We
use |[a, c)| to denote the d-dimensional volume of this box.
An infinite sequence of points x1 , x2 , [0, 1)d is uniformly distributed
P
if limn (1/n) ni=1 1axi <c = |[a, c)| holds for all boxes. This means that
In I for every function f (x) of the form 1ax<c and so for any finite linear
combination of such indicators of boxes. Riemann integrable functions are well
approximated by linear combinations of indicators of boxes; if the sequence (xi )
is uniformly distributed then limn |In I| = 0 for any function f that is Riemann integrable. Thus uniformly distributed sequences can be used to provide a
deterministic law of large numbers.
To show that a sequence is uniformly distributed it is enough to show that

In I when f is the indicator of a suitable subset of boxes. Anchored boxes

take the form [0, a) for some point a [0, 1)d . If In I for all indicators of
anchored boxes, then the same holds for all boxes. For integers b 2 a b-adic box
is a Cartesian product of the form
d h
Y
`j `j + 1 
,
.
bk j bkj
j=1

(6.14)

for integers kj 0 and 0 `j < bkj . When b = 2 the box is called dyadic. An
arbitrary box can be approximated by b-ary boxes. If I I for all indicators of badic boxes then the sequence (xi ) is uniformly distributed. A mathematically more
interesting result is the Weyl condition. The sequence (xi ) is uniformly
distributed

2
1kx

where
if and only if In I for all trigonometric polynomials f (x) = e
d
kZ .
If xi are independent U [0, 1)d variables, then (xi ) is uniformly distributed with
probability one. Of course we hope to do better than random points. To that end,
we need a numerical measure of how uniformly distributed a sequence of points is.
These measures are called discrepancies, and there are a great many of them. One
of the simplest is the star discrepancy
Dn

Dn (x1 , . . . , xn )

n
1 X



= sup 
10xi <a [0, a)
a[0,1)d n

(6.15)

i=1

Figure 6.4 illustrates this discrepancy. It shows an anchored box [0, a) [0, 1)2 and
P
a list of n = 20 points. The anchored box has 5 of the 20 points so (1/n) ni=1 10xi <a =
78

Figure 6.4: Shown are 20 points in the unit square and an anchored box (shaded)
from (0, 0) to a = (.3, .7). The anchored box [0, a) has volume 0.21 and contains
a fraction 5/20 = 0.2 of the points.

0.20. The volume of the anchored box is 0.21, so the difference is |0.2 0.21| =
0.01. The star discrepancy Dn is found by maximizing this difference over all
anchored boxes [0, a).
For xi U [0, 1)d , Chung [9] showed that

2nDn
lim sup p
=1
(6.16)
n
log(log(n))
79

so Dn = O((log log(n)/n)1/2 ) with probability one. An iterated logarithm grows

slowly with n, so Dn may be only slightly larger than n1/2 for large n.
It is known that a deterministic choice of (xi ) can yield Dn much smaller
than (6.16). There are infinite sequences (xi ) in [0, 1)d with Dn (x1 , . . . , xn ) =
O(log(n)d /n). Such sequences are called low discrepancy sequences, and some
of them are described in chapter 6.5. It is suspected but not proven that infinite
sequences cannot be constructed with Dn = o(log(n)d /n); see Beck and Chen [5].
In an infinite sequence, the first m points of x1 , . . . , xn are the same for any
n m. If we knew in advance the value of n that we wanted then we might use
a sequence customized for that value of n, such as xn1 , . . . , xnn [0, 1)d , without
insisting that xni = xn+1 i . In this setting Dn (xn1 , . . . , xnn ) = O(log(n)d1 /n)
is possible. The effect is like reducing d by one, but the practical cost is that such
a sequence is not extensible to larger n.
There is a connection between better discrepancy and more accurate integration. Hlawka [29] proved the Koksma-Hlawka inequality
|I I| Dn (x1 , . . . , xn )VHK (f ).

(6.17)

The factor VHK (f ) is the total variation of f in the sense of Hardy and Krause.
Niederreiter [56] gives the definition.
Equation (6.17) shows that a deterministic law of large numbers can be much
better than the random one, for large enough n and a function f with finite variation
VHK (f ). One often does see QMC methods performing much better than MC, but
equation (6.17) is not good for predicting when this will happen. The problem is
that Dn is hard to compute, VHK (f ) is harder still, and that the bound (6.17) can
grossly overestimate the error. In some cases VHK is infinite while QMC still beats
MC. Schlier [66] reports that even for QMC the variance of f is more strongly
related to the error than is the variation.

6.5 Digital Nets and Related Methods

Niedereitter [56] presents a comprehensive account of digital nets and sequences.
We will define them below, but first we illustrate a construction for d = 1.
The simplest digital nets are the radical inverse sequences initiated by van der
Corput [85, 86]. Let b 2 be an integer base. The non-negative integer n can
P
k1 where n {0, 1, . . . , b 1} and only finitely many
be written as
k
k=1 nk b
P
k
nk are not zero. The base b radical inverse function is b (n) =
k=1 nk b
80

`
0
1
2
3
4
5
6
7

` base 2
0.
1.
10.
11.
100.
101.
110.
111.

2 (`)
0.000
0.100
0.010
0.110
0.001
0.101
0.011
0.111

0.000
0.500
0.250
0.750
0.125
0.625
0.375
0.875

Table 6.1: The first column shows integers ` from 0 to 7. The second column shows
` in base 2. The third column reflects the digits of ` through the binary point to
construct 2 (`). The final column is the decimal version of 2 (`).
[0, 1). A radical inverse sequence consists of b (i) for n consecutive values of i,
conventionally 0 through n 1.
Table 6.1 illustrates a radical inverse sequence, using b = 2 as van der Corput
did. Because consecutive integers alternate between even and odd, the van der
Corput sequence alternates between values in [0, 1/2) and [1/2, 1). Among any 4
consecutive van der Corput points there is exactly one in each interval [k/4, (k +
1)/4) for k = 0, 1, 2, 3. Similarly any bm consecutive points from the radical
inverse sequence in base b are stratified with respect to bm congruent intervals of
length 1/bm .
If d > 1 then it would be a serious mistake to simply replace a stream of
pseudo-random numbers by the van der Corput sequence. For example with d = 2
taking points xi = (2 (2i 2), 2 (2i 1)) [0, 1)2 we would find that all xi lie
on a diagonal line with slope 1 inside [0, 1/2) [1/2, 1).
For d > 1 we really need a stream of quasi-random d-vectors. There are
several ways to generalize the van der Corput sequence to d 1. The Halton [25]
sequence in [0, 1)d works with d relatively prime bases b1 , . . . , bd . Usually these
are the first d prime numbers. Then for i 1,
xi = (2 (i 1), 3 (i 1), 5 (i 1), . . . , bd (i 1)) [0, 1)d .
The Halton sequence has low discrepancy: Dn = O((log n)d /n).
The Halton sequence is extensible in both n and d. For small d the points
of the Halton sequence have a nearly uniform distribution. The left panel of Figure 6.5 shows a two dimensional portion of the Halton sequence using prime bases
81

Figure 6.5: The left panel shows the first 23 32 = 72 points of the Halton sequence
using bases 2 and 3. The middle panel shows the first 72 points for the 10th and
11th primes, 29 and 31 respectively. The right panel shows these 72 points after
Faures [20] permutation is applied.

2 and 3. The second panel shows the same points for bases 29 and 31 as would be
needed with d = 11. While they are nearly uniform in both one dimensional projections, their two dimensional uniformity is seriously lacking. When it is possible
to identify the more important components of x, these should be sampled using the
smaller prime bases.
The poorer distribution for larger primes can be mitigated using a permutation
of Faure [20]. Let be a permutation of {0, . . . , b 1}. Then the radical inverse
P
k
function can be generalized to b, (n) =
k=1 (nk )b . It still holds that any
consecutive bm values of b, (i) stratify into bm boxes of length 1/bm . Faures
transformation b of 0, . . . , b 1 is particularly simple. Let 2 = (0, 1). For even
b > 2 take b = (2b/2 , 2b/2 + 1), so 4 = (0, 2, 1, 3). For odd b > 2 put
k = (b 1)/2 and = b1 . Then add 1 to any member of greater than or equal
to k. Then b = ((0), . . . , (k 1), k, (k), . . . , (b 2)). For example with
b = 5 we get k = 2, and after the larger elements are incremented, = (0, 3, 1, 4).
Finally 5 = (0, 3, 2, 1, 4). The third plot in Figure 6.5 shows the effect of Faures
permutations on the Halton sequence.
Digital nets provide more satisfactory generalizations of radical inverse sequences to d 2. Recall the b-ary boxes in (6.14). The box there has volume
bK where K = k1 + + kd . Ideally we would like nbK points in every such
box. Digital nets do this, at least for small enough K.
82

Let b 2 be an integer base and let m t 0 be integers. A (t, m, d)net in

base b is a finite sequence x1 , . . . , xbm for which every b-ary box of volume btm
contains exactly bt points of the sequence.
Clearly t = 0 corresponds to better stratification. For given values of b, m, and
d, particularly for large d, there may not exist a net with t = 0, and so nets with
t > 0 are widely used.
Faure [19] provides a construction of (0, m, p)nets in base p where p is a
prime number. The first component of these nets is the radical inverse function in
base p applied to 0 through bm 1. Figure 6.6 shows 81 points of a (0, 4, 2)net
in base 3. There are 5 different shapes of 3-ary box with volume 1/81. The aspect
ratios are 1 1/81, 1/3 1/27, 1/9 1/9, 1/17 1/3, and 1/81 1. Latin
hypercube samples of 81 points balance the first and last of these, jittered sampling
balances the third, while multi-jittered sampling balances the first, third, and fifth.
A (0, 4, 2)net balances 81 different 3-ary boxes of each of these 5 aspect ratios.
If f is well approximated by a sum of the corresponding 405 indicator functions,
then |I I| will be small.
The extensible version of a digital net is a digital sequence. A (t, s)sequence
in base b is an infinite sequence (xi ) for i 1 such that for all integers r 0
and m t, the points xrbm +1 , . . . , x(r+1)bm form a (t, m, d)net in base b. This
sequence can be expressed as an infinite stream of (t, m, d)nets, simultaneously
for all m t. Faure [19] provided a construction of (0, p)-sequences in base p.
Niederreiter [55] showed that construction can be extended to (0, q)sequences in
base q where q = pr is a power of a prime p. The Faure net shown in Figure 6.6 is
in fact the first 81 points of the first two variables in a (0, 3)-sequence in base 3.
For m t and 1 < b, the first bm points in a (t, d)sequence are
balanced with respect to all b-ary boxes of volume btm or larger. If n is not of the
form bm , then the points do not necessarily balance any non-trivial b-ary boxes.
The Faure sequence and Niederreiters generalization of it, require b d.
When the dimension is large then it becomes necessary to use a large base b, and
then either bm is very large, or m is very small. Then the Sobol [75] sequences
become attractive. They are (t, d)sequences in base b = 2. The quality parameter t depends on d. Niederreiter [55] combined the methods of Sobol and Faure,
generating new sequences. Any (t, s)sequence is a low discrepancy sequence, as
shown in Niederreiter [56]. The smallest values of t, for given b and d, among
known (t, d)sequence constructions, are those of Niedereitter and Xing [54].
83

Figure 6.6: Shown are 81 points of a (0, 4)net in base 3. Reference lines are
included to make the 3-ary boxes more visible. There 5 different shapes of 3-ary
box balanced by these points. One box of each shape is highlighted.

6.6 Integration Lattices

In addition to digital nets and sequences, there is a second major QMC technique,
known as integration lattices. The simplest example of an integration lattice is a
rank one lattice. These take the form
xi =

i1
(g1 , . . . , gd )
n
84

mod n

(6.18)

Figure 6.7: Shown are the points of two integration lattices in the unit square. The
lattice on the right has much better uniformity, showing the importance of making
a good choice of lattice generator.

for i = 1, . . . , n. Usually g1 = 1. Figure 6.7 shows two integration lattices in

[0, 1)2 with n = 89. The first has g2 = 22 and the second one has g2 = 55.
It is clear that the second lattice in Figure 6.7 is more evenly distributed than
the first one. The method of good lattice points is the application of the rule (6.18)
with n and g carefully chosen to get good uniformity. Fang and Wang [18] and
Hua and Wang [31] describe construction and use of good lattice points, including
extensive tables of n and g.
Sloan and Joe [74] describe integration lattices in general, including lattices of
rank higher than 1. A lattice of rank r for 1 r d requires r vectors like g to
generate it. The origin of lattice methods is in Korobov [43]. Korobovs rules have
g = (1, h, h2 , . . . , hd1 ) so that the search for a good rule requires only a careful
choice of two numbers n and h.
Until recently, integration lattices were not extensible. Extensible integration
lattices are a research topic of current interest, following the publication of Hickernell, Hong, LEcuyer and Lemieux [28].
85

Integration lattices are not as widely used in computer graphics as digital nets.
Their periodic structure is likely to produce unwanted aliasing artifacts, at least in
some applications. Compared to digital nets, integration lattices are very good at
integrating smooth functions, especially smooth periodic functions.

6.7

Randomized Quasi-Monte Carlo

QMC methods may be thought of as derandomized MC. Randomized QMC (RQMC)

methods re-randomize them. The original motivation is to get sample based error
estimates.
In RQMC, one takes a QMC sequence (ai ) and transforms it into random points
(xi ) such that xi retain a QMC property and the expectation of I is I. The simplest
way to achieve the latter property is to have each xi U [0, 1)d . With RQMC
we can repeat a QMC integration R times independently getting I1 , . . . , IR . The
P

combined estimate I = (1/R) R

expected value I and an unbiased
r=1 Ir has
Pr
1

2
estimate of the RMSE of I is [R(R 1)]
r=1 (Ir I) .
Cranley and Patterson [15] proposed a rotation modulo one
xi = ai + U

mod 1

where U U [0, 1)d and both addition and remainder modulo one are interpreted
componentwise. It is easy to see that each xi U [0, 1)d . Cranley and Patterson
proposed rotations of integration lattices. Tuffin [83] considered applying such
rotations to digital nets. They dont remain nets, but they still look very uniform.
Owen [57] proposes a scrambling of the base b digits of ai . Suppose that ai
is the ith row of the matrix A with entries Aij for j = 1, . . . , d, and either i =
P
k
1, . . . , n for a finite sequence or i 1 for an infinite one. Let Aij =
k=1 b aijk
where aijk {0, 1, . . . , b 1}. Now let xijk = jaij1 ...aij k1 (aijk ) where
jaij1 ...aij k1 is a uniform random permutation of 0, . . . , b 1. All the permutations required are independent, and the permutation applied to the kth digits of
Aij depends on j and on the preceding k 1 digits.
Applying this scrambling to any point a [0, 1)d produces a point x
U [0, 1)d . If (ai ) is a (t, m, d)net in base b or a (t, d)sequence in base b, then
with probability 1, the same holds for the scrambled version (xi ). The scrambling
described above requires a great many permutations. Random linear scrambling
is a partial derandomization of scrambled nets, given by Matousek [49] and also
86

in Hong and Hickernell [30]. Random linear scrambling significantly reduces the
number of permutations required from O(dbm ) to O(dm2 ).
= o(1/n)
For integration over a scrambled digital sequence we have Var(I)
2
for any f with < . Thus for large enough n a better than MC result will be
obtained. For integration over a scrambled (0, m, d)-net Owen [58] shows that

Var(I)

 b min(d1,m) 2
2.72 2

.
b1
n
n

.
That is scrambled (0, m, d)nets cannot have more than e = exp(1) = 2.72 times
the Monte Carlo variance for finite n. For nets in base b = 2 and t 0, Owen [60]
shows that
2
2t 3d .
Var(I)
n
Compared to QMC, we expect RQMC to do no harm. After all, the resulting
xi still have a QMC structure, and so the RMSE should be O(n1 (log n)d ). Some
forms of RQMC reduce the RMSE to O(n3/2 (log n)(d1)/2 ) for smooth enough
f . This can be understood as random errors cancelling where deterministic ones
do not. Surveys of RQMC appear in Owen [61] and LEcuyer and Lemieux [46].

6.8

Padding and Latin Supercube Sampling

In some applications d is so large that it becomes problematic to construct a meaningful QMC sequence. For example the number of random vectors needed to follow a single light path in a scene with many reflective objects can be very large
and may not have an a priori bound. As another example, if acceptance-rejection
sampling (Devroye [17]) is used to generate a random variable then a large number
of random variables may need to be generated in order to produce that variable.
Padding is a simple expedient solution to the problem. One uses a QMC or
RQMC sequence in dimension s for what one expects are the s most important input variables. Then one pads out the input with d s independent U [0, 1) random
variables. This technique was used in Spanier [76] for particle transport simulations. It is also possible to pad with a d s dimensional Latin hypercube sample
as described in Owen [59], even when d is conceptually infinite.
In Latin supercube sampling, the d input variables of xi are partitioned into
some number k of groups. The jth group has dimension dj 1 and of course
Pk
j=1 dj = d. A QMC or RQMC method is applied in each of the k groups.
87

Just as the van der Corput sequence cannot simply be substituted for a pseudorandom generator, care has to be taken in using multiple (R)QMC methods within
the same problem. It would not work to take k independent randomizations of
the same QMC sequence. The fix is to randomize the run order of the k groups
relative to each other, just as Latin hypercube sampling randomizes the run order
of d stratified samples.
To describe LSS, for j = 1, . . . , k and i = 1, . . . , n let aji [0, 1)dj . Suppose
that aj1 , . . . , ajn are a (R)QMC point set. For j = 1, . . . , k, let j (i) be independent uniform permutations of 1, . . . , n. Then let xji = ajj (i) . The LSS has rows
xi comprised of x1i , . . . , xki . Owen [59] shows that in Latin supercube sampling
the function f can be written as a sum of two parts. One, from within groups
of variables, is integrated with an (R)QMC error rate, while the other part, from
between groups of variables, is integrated at the Monte Carlo rate. Thus a good
grouping of variables is important as is a good choice of (R)QMC within groups.

88

Chapter 7

Monte Carlo Path Tracing

By Pat Hanrahan

This Chapter discusses Monte Carlo Path Tracing. Many of these ideas appeared
in James Kajiyas original paper on the Rendering Equation. Other good original
sources for this material is L. Carter and E. Cashwells book Particle-Transport
Simulation with the Monte Carlo Methods and J. Spanier and E. Gelbards book
Monte Carlo Principles and Neutron Transport Problems.

7.1

Solving the Rendering Equation

To derive the rendering equation, we begin with the Reflection Equation

Z
Lr (~x,
~ r) =
fr (~x,
~i
~ r ) Li (~x,
~ i ) cos i di .
i

The reflected radiance Lr is computed by integrating the incoming radiance over

a hemisphere centered at a point of the surface and oriented such that its north
pole is aligned with the surface normal. The BRDF fr is a probability distribution
function that describes the probability that an incoming ray of light is scattered in a
random outgoing direction. By convention, the two direction vectors in the BRDF
point outward from the surface.
One of the basic laws of geometric optics is that radiance does not change
as light propagates (assuming there is no scattering or absorption). In free space,
where light travels along straight lines, radiance does not change along a ray. Thus,
assuming that a point ~x on a receiving surface sees a point ~x0 on a source surface,
89

Li(x,i)

Lr(x,r)

fr(x,i,r)

Figure 7.1: Integrating over the upper hemisphere.

the incoming radiance is equal to the outgoing radiance:
Li (~x,
~ i (~x0 , ~x)) = Lo (~x0 ,
~ o0 (~x, ~x)).
Where we use the direction as a function of two points as

~ (~x1 , ~x2 ) =
~ (~x1 ~x2 ) =

~x2 ~x1
.
|~x2 ~x1 |

The standard convention is to parameterize the incoming radiance Li at ~x with

the direction from the receiver ~x to the source ~x0 . When using this convention,
the incoming radiance is defined on a ray pointing in the direction opposite to the
direction of light propagation.
It is also useful to introduce notation for the two-point radiance
L(~x, ~x0 ) = L(~x ~x0 ) = Lo (~x,
~ (~x, ~x0 )).
and the three-point BRDF
fr (~x, ~x0 , ~x00 ) = fr (~x ~x0 ~x00 ) = fr (~x0 ,
~ (~x0 , ~x),
~ (~x00 , ~x0 ))
(Note: the two-point radiance function defined here is different than the two-point
intensity function defined in Kajiyas original paper.)
90

dA(x')
'

x'

(x',x)

(x,x')

x
dA(x)
Figure 7.2: Two-point geometry.
Point to point functions are useful since they are intuitive and often clarify the
geometry and physics. For example, if ~x sees ~x0 , then ~x0 sees ~x. This mutual visibility is represented as the two-point visibility function, V (~x, ~x0 ), which is defined
to be 1 if a line segment connecting ~x to ~x0 does not intersect any opaque object,
and 0 otherwise.
The reflection equation involves an integral over the upper hemisphere. This
integral may be converted to an integral over other surfaces by changing of variables from solid angles to surface areas. This is easily done by relating the solid
angle subtended by the source to its surface area.
di =

cos o0
dA(~x0 )
|~x ~x0 |2

The projected solid angle then becomes

cos i di = G(~x, ~x0 ) dA(~x0 )
where

cos i cos o0
V (~x, ~x0 )
|~x ~x0 |2
In these equations we are making a distinction between the parameters used to
specify points on the surface (the ~xs) and the measure that we are using when perG(~x, ~x0 ) = G(~x0 , ~x) =

91

forming the integral (the differential surface area dA(~x)). Sometimes we will be
less rigorous and just use dA or dA0 when we mean dA(~x) and dA(~x0 ). The geometry factor G is related to the differential form factor by the following equation:
0
F (~x, ~x0 )dA0 = G(~x,~x ) dA0 .
Performing this change of variables in the reflection equation leads to the following integral
Z
fr (~x,
~ (~x, ~x0 ),
~ ) Lo (~x0 ,
~ (~x0 , ~x)) G(~x, ~x0 ) V (~x, ~x0 ) dA(~x0 )
Lr (~x,
~) =
A

In this equation, ~x and

~ are indendent variables and we are integrating over surface
area which is parameterized by ~x0 ; thus, the incoming direction
~ i and the direction
of Lo are functions of these positions and are indicated as such.
The final step in the derivation is to account for energy balance
Lo (~x,
~ ) = Le (~x,
~ ) + Lr (~x,
~ ).
This states that the outgoing radiance is the sum of the emitted and reflected radiances. Placing an emission function on each surface allows us to create area light
sources. Inserting the reflection equation into the energy balance equation results
in the Rendering Equation.
Z
L(~x,
~ ) = Le (~x,
~) +
fr (~x,
~ (~x, ~x0 ),
~ ) L(~x0 ,
~ (~x0 , ~x)) G(~x, ~x0 ) V (~x, ~x0 ) dA0
A

For notational simplicity, we will drop the subscript o on the outgoing radiance.
The rendering equation couples radiance at the receiving surfaces (the lefthand side) to the radiances of other surfaces (inside the integrand). This equation
applies at all points on all surfaces in the environment. It is important to recognize
the knowns and the unknowns. The emission function Le and the BRDF fr are
knowns since they depends on the scene geometry, the material characteristics, and
the light sources. The unknown is the radiance L on all surfaces. To compute the
radiance we must solve this equation. This equation is an example of an integral
equation, since the unknown L appears inside the integral. Solving this equation in
the main goal of Monte Carlo Path Tracing.
The rendering equation is sometimes written more compactly in operator form.
An operator is a method for mapping a function to another function. In our case,
the function is the radiance.
L = Le + K L
92

Sometimes it is useful to break the operator K into two operators, T and S. T is

the transfer operator and applied first; it takes outgoing light on one surface and
transfers it to another surface.
Li (~x,
~ (~x0 , ~x)) = T L(~x,
~ (~x, ~x0 ))
S is the scattering or reflection operator which takes the incoming light distribution
and computes the outgoing light distribution.
Lr (~x,
~ ) = S Li (~x,
~ 0)
Operator equations like the rendering equation may be solved using iteration.
L0 = Le
L1 = Le + K L0 = Le + K Le
...
Ln = Le + K Ln1 =

n
X

K n Le

i=0

Noting that K 0 = I, where I is the identity operator. This infinite sum is called the
Neumann Series and represents the formal solution (not the computed solution) of
the operator equation.
Another way to interpret the Neumann Series is to draw the analogy between
1
= (1 x)1 = 1 + x + x2 ...,
1x
and
(I K)1 = I + K + K 2 ....
The rendering equation
(I K) L = Le
then has the following solution
L = (I K)1 Le
Note that (I K)1 is just an operator acting on the emission function. This
operator spreads the emitted light over all the surfaces.
93

fr(x2,x3,x4)
dA(x2)
dA(x0)

G(x0,x1)

dA(x4)

G(x2,x3)
G(x1,x2)
dA(x3)

dA(x1)
fr(x1,x2,x3)

G(x3,x4)

fr(x2,x3,x4)

Figure 7.3: A path from point ~x1 to ~x5 .

It is useful to write out the formal solution
L(~x,
~) =

K i Le (~x0 ,
~ 0 ))

i=0

in all its gory detail. Lets consider one term:

Ln (~x,
~ ) = L(~xn , ~xn+1 ) = K n Le
Z
Z
=
...
Le (~x0 , ~x1 )G(~x0 , ~x1 )fr (~x0 , ~x1 , ~x2 )G(~x1 , ~x2 )...
A

~ x)n
G(~xn1 , ~xn )fr (~xn1 , ~xn , ~xn+1 ) dA0 dA1 ...dA(~
This integral has a very simple geometric and physical intuition. It represents
a family of light paths. Each path is characterized by the number of bounces or
length n. There are many possible paths of a given length. Paths themselves are
specified by a set vertices. The first vertex is a point on the light source and subsequent vertices are points on reflecting surfaces. The total contribution due to all
paths of a given length is formed by integrating over all possible light and surface
positions. This involves doing n integrals over surface areas. However, we must
weight paths properly when performing the integral. A integrand for a particular
path consists of alternating sequence of geometry and reflection terms. Finally, the
final solution to the equation is the sum of paths of all lengths; or more simply, all
possible light paths.
Note that these are very high dimensional integrals. Specifically, for a path of
length n the integral is over a 2n dimensional space. The integral also involves
very complicated integrands that include visibility terms and complex reflection
94

functions defined over arbitrary shapes. It turns out these complexities is what
makes Monte Carlo the method of choice for solving the rendering equation.
There is one more useful theoretical and practical step, and that is to relate the
solution of the rendering equation to the process of image formation in the camera.
The equation that governs this process is the Measurement Equation
Z Z Z
M=

R(~x,
~ , t)L(~x,
~ , t) dt d dA.
A

The response function R(~x,

~ , t) depends on the the pixel filter (the ~x dependence),
the aperture (the
~ dependence), and the shutter (the t dependence). Other factors
such as transformations of rays by the lens system and spectral sensitivities may
also be included, but we will ignore these factors to simplify the presentation.
As seen above, the pixels value in the image is a function that involves nested
integrals. These integrals are very complicated, but we can easily evaluate the
integrand which corresponds to sampling the function.
Sampling a pixel over (x, y) prefilters the image and reduces aliasing.
Sampling the camera aperture (u, v) produces depth of field.
Sampling in time t (the shutter) produces motion blur.
Sampling in wavelength simulates spectral effects such as dispersion
Sampling the reflection function produces blurred reflection.
Sampling the tranmission function produces blurred transmission.
Sampling the solid angle of the light sources produces penumbras and soft
shadows.
Sampling paths accounts for interreflection.
Sampling in x, y, u, v and t has been discussed previously. Sampling light sources
and performing hemispherical integration has also been discussed. What remains
is to sample paths.
95

7.2 Monte Carlo Path Tracing

First, lets introduce some notation for paths. Each path is terminated by the eye
and a light.
E - the eye.
L - the light.
Each bounce involves an interaction with a surface. We characterize the interaction as either reflection or tranmission. There are different types of reflection and
transmission functions. At a high-level, we characterize them as
D - diffuse reflection or transmission
G - glossy reflection or tranmission
S - specular reflection or refraction
Diffuse implies that light is equally likely to be scattered in any direction. Specular
implies that there is a single direction; that is, given an incoming direction there is
a unique outgoing direction. Finally, glossy is somewhere in between.
Particular ray-tracing techniques may be characterized by the paths that they
consider.
Appel Ray casting: E(D|G)L
Whitted Recursive ray tracing: E[S ](D|G)L
Kajiya Path Tracing: E[(D|G|S)+ (D|G)]L
Goral Radiosity: ED L
The set of traced paths are specified using regular expressions, as was first proposed
by Shirley. Since all paths must involve a light L, the eye E, and at least one
surface, all paths have length at least equal to 3.
A nice thing about this notation is that it is clear when certain types of paths
are not traced, and hence when certain types of light transport is not considered
by the algorithm. For example, Appels algorithm only traces paths of length 3,
ignoring longer paths; thus, only direct lighting is considered. Whitteds algorithm
traces paths of any length, but all paths begin with a sequence of 0 or more mirror
96

reflection and refraction steps. Thus, Whitteds technique ignores paths such as
the following EDSDSL or E(D|G) L. Distributed ray tracing and path tracing
includes multiple bounces involving non-specular scattering such as E(D|G) L.
However, even these methods ignore paths of the form E(D|G)S L; that is, multiple specular bounces from the light source as in a caustic. Obviously, any technique
that ignores whole classes of paths will not correctly compute the solution to the
rendering equation.
Lets now describe the basic Monte Carlo Path Tracing Algorithm:
Step 1. Choose a ray given (x,y,u,v,t)
weight = 1
Step 2. Trace ray to find point of intersection with the nearest surface.
Step 3. Randomly decide whether to compute emitted or reflected light.
Step 3A. If emitted,
return weight * Le
Step 3B. If reflected,
weight *= reflectance
Randomly scatter the ray according to the BRDF pdf
Go to Step 2.
This algorithm will terminate as long as a ray eventually hits a light source. For
simplicity, we assume all light sources are described by emission terms attached to
surfaces. Latter we will discuss how to handle light sources better.
A variation of this algorithm is to trace rays in the opposite direction, from light
sources to the camera. We will assume that reflective surface never absorb light,
and that the camera is a perfect absorber.
Step 1. Choose a light source according to the light source power distribution.
Generate a ray from that light source according to its intensity distribution.
weight = 1
Step 2. Trace ray to find point of intersection.
Step 3. Randomly decide whether to absorb or reflect the ray.
97

Step 3A. If scattered,

weight *= reflectance
Randomly scatter the ray according to the BRDF.
Go to Step 2.
Step 3B. If the ray is absorbed by the camera film,
Record weight at x, y
Go to Step 1.
The first algorithm is an example of forward ray tracing; in forward ray tracing
rays start from the eye and propagate towards the lights. Forward ray tracing is
also called eye ray tracing. In contrast, in backward ray tracing rays start at the
light and trace towards the eye. As we will discuss in a subsequent section, Both
methods are equivalent because the physics of light transport does not change if
paths are reversed. Both methods have there advantages and disadvantages, and in
fact may be coupled.
The above simple algorithms form the basis of Monte Carlo Path Tracing.
However, we must be more precise. In particular, there are two theoretical and
practical challenges:
Challenge 1 : Sampling an infinite sum of paths in an unbiased way.
Challenge 2 : Finding good estimators with low variance.

7.3

Random Walks and Markov Chains

To understand more about why path tracing works, lets consider a simpler problem: a discrete random walk. Instead of a physical system with continuous variables, such as position and direction, consider a discrete physical system comprised
of n states. Path tracing as described above is an example of a random walk where
we move from sample to sample, or from point to point, where the samples are
drawn from a continuous probability distribution. In a discrete random walk, we
move from state to state, and the samples are drawn from a discrete probability
distribution.
A random walk is characterized by three distributions:
1. Let p0i be the probability of starting in state i.
2. Let pi,j is the probability of moving from state i to state j.
98

Transition
pi,j

Creation p0i

p*i Termination

Figure 7.4: State transition diagram for a discrete random walk.

3. Let pi is the probability of being terminated in state i.
Because the probability of transition and termination must sum to one, pi = 1
P
j=0 pi,j ; that is, the probability of terminating in state i is equal to the probability
of not moving from state i to j.
A discrete random walk consists of the following steps.
Step 1. Create a random particle in state i with probability p0i .
Step 2. With probability pi , terminate in state i.
Score particle in state i by incrementing the counter for state i
Go to Step 1.
Step 3. Randomly select new state according to the transition probability distribution.
Set i to new j.
Go to Step 2.
Random walks are also called Markov Chains. A Markov Chain is a sequence
of states generated by a random process. We will return to Markov Chains when
we discuss the Metropolis Algorithm. Markov Chains also come up Bayesian reasoning and in learning theory (e.g. Hidden Markov Models). Keep your eyes open
for Markov Chains; you will see these techniques used more and more in computer
graphics.
Given a set of particles following random walks, the problem is to compute the
final probability of a particle being terminated in state i. To solve this problem,
we introduce another random variable Pin , which is the probability of being in
99

state i after n transitions. Since each state transition is independent of the previous
transitions, this probability may be computed using a simple recurrence
Pj0 = p0j
Pj1 = p0j +

pi,j Pi0

...
Pjn = p0j +

pi,j Pin1 .

Defining a matrix M whose entries are Mi,j = pi,j , the above process can be
viewed as the following iterative product of a matrix times a vector
P 0 = p0
P 1 = p0 + M P 0
...
P n = p0 + M P n1 .
And this procedure may be recognized as the iterative solution of the following
matrix equation
(I M )P = p0
since then
P = (I M )1 p0 = p0 + M (p0 + M (p0 ... =

M i p0 .

i=0

This process will always converge assuming the matrices are probability distributions. The basic reason for this is that probabilities are always less than one, and so
a product of probabilities quickly tends towards zero.. Thus, the random walk provides a means for solving linear systems of equations, assuming that the matrices
are probability transition matrices. Note the similiarity of this discrete iteration of
matrices to the iterative application of the continuous operator when we solve the
rendering equation using Neumann Series.
This method for solving matrix equations using discrete random walks may be
directly applied to the radiosity problem. In the radiosity formulation,
Bi = Ei + i

X
j

100

Fi,j Bj

where the form factor equals

Z Z
1
Fi,j =
G(~x, ~x0 )V (~x, ~x0 ) dA(~x) dA(~x0 )
Ai Ai Aj
Recall that the form factors may be interpreted as a probabilities. The form factor
Fi,j is the percentage of outgoing light leaving surface element Ai that falls on
surface element Aj . In fact, the form factor is the probability that a random ray
leaving surface element Ai makes it to Aj (although one has to be careful about
how one defines a random ray). Thus, form factor matrices may be interprested
as transition matrices. In this equation, is the diffuse reflectance and is equal to
= B/E; The reflectance must be positive and less than 1. The absorption or
termination probability is thus equal to 1 .
These observations lead to a method for solving the matrix radiosity equation
using a discrete random walk. The major issue, which is always a case with radiosity solution techniques, is computing the form factor matrix. This process is
expensive and error prone because of the complexity of the environment and the
difficulty in doing exact visible surface determination. The form-factor matrix is
also very large. For example, a scene consisting of a million surface elements
would require a million squared matrix. Therefore, in practice, the form factor matrix is often calculated on-the-fly. Assuming a particle is on some surface element
i, an outgoing particle may be sent off in a random direction where the random
direction is chosen from a cosine-weighted distribution (here the cosine is with respect to the surface element normal). The particle is then ray-traced and the closent
point of intersection on surface element j is found. This random process is roughly
equivalent to one generated from a known form-factor matrix.
It is interesting to prove that random walks provide an unbiased estimate of
the solution of the linear system of equations. Although this proof is a bit formal,
it is worthwhile working it through to get a flavor of the mathematical techniques
involved.
The first step is to define a random variable on the space of all paths. Lets
signify a path of length k as k = (i1 , i2 , ..., ik ); this path involves a sequence of
transitions from state i1 to state i2 and ending up finally after k transitions in state
ik . The random variable without the subscript is the set of all paths of length one
to infinity.
The next step is to pick an estimator W () for each path. Then we can compute
the expected value of the estimator by weighting W by the probability that a given
101

path is sampled, p().

E[W ] =

p()W ()

X
X

p(k )W (k )

k=1 k

X
X

...

k=1 i1

p(i1 , ..., ik )W (i1 , ..., ik )

ik

In the last line we group all paths of the same length together. The sums on each
index i go from 1 to n - the number of states in the system. Thus, there are nk
paths of length k, and of course paths can have infinite length. There are a lot of
paths to consider!
What is the probability p(k ) of path k ending in state ik ? Assuming the discrete random walk process described above, this is the probability that the particle
is created in state i1 , times the probability of making the necessary transitions to
arrive at state ik , times the probability of being terminated in state ik
p(k ) = p0i1 pi1 ,i2 ...pik1 ,ik pk
With these careful definitions, the expected value may be computed
E[Wj ] =
=

X
X
k=1 k
X
X

p(k )Wj (k )
...

k=1 i1

p0i1 pi1 ,i2 ...pik1 ,ik pik Wj (k )

ik

Recall, that our estimator counts the number of counts the number of particles that
terminate in state j. Mathematically, we can describe this counting process with
a delta function, Wj (k ) = ik ,j /pj . This delta function only scores particles
terminating in ik = j. The expected value is then
E[Wj ] =

X
X
k=1 i1
X
X
k=1 i1

...

XX

p0i1 pi1 ,i2 ...pik1 ,ik pik ik ,j /pj

ik1 ik

...

p0i1 pi1 ,i2 ...pik1 ,j .

ik1

102

This sum may be recognized as the j component of the matrix equation

E[W ] = p0 + M p0 + M 2 p0 + ...
which is the desired solution of the linear system of equations.
Note that we had to be careful about the estimator. If we hadnt divided the
count of the particles by the probability of a termination event, the expected value
would not have equaled the right answer. Picking the wrong estimator - that is, an
estimator that results in the wrong expected value - for a complex sampling process
is one of the most common errors when uding Monte Carlo Techniques. Until you
have a lot of experience, it is worthwhile convincing yourself that your estimator
is unbiased.
This technique was originally developed by von Neumann and Ulam, the originators of the Monte Carlo Method. The estimator they used is often called the
absorption estimator, since only particles that are absorbed are counted. An interesting variation, developed by Wasow, is to count all the number of particles that
pass through state j (including those terminate as well as those that make a transition). This is called the collision estimator, since it counts all particles colliding
with a surface. It is an interesting exercise to show that the collision estimator also
provides an unbiased estimate of the solution to the linear equation. It is more challenging, but more interesting, to also derive the conditions when and if the collision
estimator works better than the absorption estimator.
This basic proof technique is easy to generalize to continuous distributions,
but the notation is messy. The details are described in Chapter 3 of the Spanier
and Gelbard book on neutron transport [77], the most authoritative source if you
wish to understand the theory behind Monte Carlo Techniques for solving integral
equations and tranport problems.

7.4

Adjoint Equations and Importance Sampling

Recall, that the pixel response is equal to the sum over paths of length n
Z
Z
Mn =
... S(~x0 , ~x1 )G(~x0 , ~x1 )fr (~x0 , ~x1 , ~x2 )G(~x1 , ~x2 )
0

~ x)n .
...fr (~xn2 , ~xn1 , ~xn )G(xn1 , xn )R(~xn1 , ~xn )dA0 dA1 ...dA(~
where we have switched notation and written the source term as S(~x, ~x0 ) = Le (~x, ~x0 ).
103

As noted above this equation is symmetric under the interchange of lights and
sensors. Switching Le with R, and noting that
Z
Z
Mn =
...
R(~x0 , ~x1 )G(~x0 , ~x1 )fr (~x0 , ~x1 , ~x2 )G(~x1 , ~x2 )
A

~ x)n
...fr (~xn2 , ~xn1 , ~xn )G(xn1 , xn )S(~xn , ~xn1 ) dA0 dA1 ...dA(~
Z
Z
=
...
S(~xn , ~xn1 )G(xn , xn1 )fr (~xn , ~xn1 , ~xn2 )G(~xn1 , ~xn2 )
A

~ x)n
...fr (~x2 , ~x1 , ~x0 )G(~x1 , ~x0 )R(~x1 , ~x0 ) dA0 dA1 ...dA(~
In the second step, we noted from the symmetry of the geometry that
G(~xi , ~xj ) = G(~xj , ~xi )
and because of the reciprocity principle the BRDF is also symmetric
fr (~xi , ~xj , ~xk ) = fr (~xk , ~xj , ~xi )
These symmetries implie that we may ray trace from either the light or the eye;
both methods will lead to the same integral.
Suppose now we break the path at some point k. The amount of light that
makes to k is
Z
Z
LS (~xk , ~xk+1 ) =
...
S(~x0 , ~x1 )G(~x0 , ~x1 )fr (~x0 , ~x1 , ~x2 )G(~x1 , ~x2 )
A

~ x)k1
...G(~xk1 , ~xk )fr (~xk1 , ~xk , ~xk+1 ) dA0 ...dA(~
In a similar way, treating the sensor as a virtual light source, we can compute the
amount of light coming from the sensor makes it to k.
Z
Z
LR (~xk , ~xk+1 ) =
... fr (~xk , ~xk+1 , ~xk+2 )G(~xk+1 , ~xk+2 )
k+2

~ x)n
...fr (~xn2 , ~xn1 , ~xn )G(xn1 , xn )R(~xn1 , ~xn ) dAk+2 ...dA(~
The measured response is then
Z Z
M=
LS (~xk , ~xk+1 )G(~xk , ~xk+1 )LR (~xk , ~xk+1 ) dAk dAk+1
A

Note the use of the notation LS and LR to indicate radiance cast from the source
vs. the receiver.
104

dA(x2)
LS(x2,x3)
S

G(x2,x3)

LR(x2,x3)
dA(x3)

Figure 7.5: A path with both the forward and backward (adjoint) solution of the
transport equation. The forward solution is generated from the source term S and
the backward solution is generated from the received term R. For physical situations where the transport equation is invariant under path reversal, the forward and
backward equations are the same.

We make two observations about this equation. First, this equation can be
considered the inner product of two radiance functions. If we consider radiance to
be a function on rays r = (~x,
~ ), then if we have functions f (r) and g(r), the inner
product of f and g is
Z
< f, g >= f (r)g(r)d(r)
where d(r) is the appropriate measure on rays. The natural way to measure
the rays between two surface elements A and A0 is d(r) = G(x, x0 ) dA dA0 .
Equivalently, considering r to be parameterized by position ~x and direction
~ , the
~
d(r) = d~
dA(~x)(~x).
Second, this integral naturally leads to a method for importance sampling a
path. Suppose we are tracing light and arrive at surface k. To compute the sensor
response, we need to integrate L against R. In this sense, R may be considered
an importance function for sampling the next directions, since we want a sampling
technique that is proportional to R to achieve low variance. But R is the solution
of the reversed transport equation that would be computed if we were to trace rays
from the sensor. R tells us how much light from the sensor would make it to this
point. Thus, the backward solution provides an importance function for the forward
solution, and vice versa. This is the key idea between bidirectional ray tracing.
Manipulating about adjoint equations is easy using the operator notation. Using
105

the operator notation, an integral equation is just

Z
K f = K(x, y)f (y) dy.
We want to estimate the integral given by the measurement equation, which is just
the inner product of two functions
Z

Z
M =< f, K g >= f (x)
K(x, y)g(y) dy dx.
This of f as the response of the sensor and K g as the solution of the rendering
equation. This equation may be rearranged
Z

Z
< f, K g > =
f (x)
K(x, y)g(y) dy dx
Z

=
f (x)K(x, y) dx g(y) dy
= < K + f, g > .
Note the difference between
Z
K f =

K(x, y)f (y) dy

and
K+ f =

Z
K(x, y)f (x) dx.

One integral is over the first variable, the other is over the second variable. Of
course, if K(x, y) = K(y, x) these two integrals are the same, in which case
K + = K and the operator is said to be self-adjoint.
This notation provides a succinct way of proving that the forward estimate is
equal to the backward estimate of the rendering equation. Recall
K LS = S
We can also write a symmetric equation in the other direction
K LR = R
Then,
< R, LS > = < K LR , LS >
= < LR , K + LS >
= < LR , K LS >
= < LR , S >
106

This result holds even if the operator is not self-adjoint. We will leave the demonstration of that fact as an exercise.
This is a beautiful result, but what does it mean in practice. Adjoint equations
have lots of applications in all areas of mathematical physics. What they allow you
to do is create output sensitive algorithms. Normally, when you solve an equation
you solve for the answer everywhere. Think of radiosity; when using the finite
element method you solve for the radiosity on all the surfaces. The same applies to
light ray tracing or the classic discrete random walk; you solve for the probability
of a particle landing in any state. However, in many problems you only want to
find the solution at a few points. In the case of image synthesis, we only need to
compute the radiance that we see, or that falls on the film. Computing the radiance
at other locations only needs to be done if its effects are observable.
We can model the selection of a subset of the solution as the inner product
of the response function times the radiance. If we only want to observe a small
subset of the solution, we make the response function zero in the locations we
dont care about. Now consider the case when all the surfaces act as sources and
only the film plane contributed a non-zero response. Running a particle tracing
algorithm forward from the sources would be very inefficient, since only rarely
is a particle terminated on the film plane. However, running the algorithm in the
reverse direction is very efficient, since all particles will terminate on sources. Thus
each particle provides useful information. Reversing the problem has led to a much
more efficient algorithm.
The ability to solve for only a subset of the solution is a big advantage of
the Monte Carlo Technique. In fact, in the early days of the development of the
algorithm, Monte Carlo Techniques were used to solve linear systems of equations.
It turns out they are very efficient if you want to solve for only one variable. But
be wary: more conventional techniques like Gaussian elimination are much more
effective if you want to solve for the complete solution.

107

108

Chapter 8

The Rendering Equation and

Path Tracing
by Philip Dutre

This chapter gives various formulations of the rendering equation, and outlines
several strategies for computing radiance values in a scene.

8.1

Formulations of the rendering equation

The global illumination problem is in essence a transport problem. Energy is emitted by light sources and transported through the scene by means of reflections (and
refractions) at surfaces. One is interested in the energy equilibrium of the illumination in the environment.
The transport equation that describes global illumination transport is called
the rendering equation. It is the integral equation formulation of the definition
of the BRDF, and adds the self-emittance of surface points at light sources as an
initialization function. The self-emitted energy of light sources is necessary to
provide the environment with some starting energy. The radiance leaving some
point x, in direction , can be expressed as an integral over all hemispherical
directions incident on the point x (figure 8.1):
Z
L(x ) = Le (x ) +

fr (x, )L(x )cos(Nx , )d

x

109

Nx
L(x)

L(x)
Le(x)

L(x)
L(x)
x
Figure 8.1: Rendering equation
One can transform the rendering equation from an integral over the hemisphere
to an integral over all surfaces in the scene. Also, radiance remains unchanged
along straight paths, so exitant radiance can be transformed to incident radiance and
vice-versa, thus obtaining new versions of the rendering equation. By combining
both options with a hemispheric or surface integration, four different formulations
of the rendering equation are obtained. All these formulations are mathematically
equivalent.
Exitant radiance, integration over the hemisphere
Z
L(x ) = Le (x ) +

fr (x, )L(y ) cos(Nx , )d

x

with
y = r(x, )
When designing an algorithm based on this formulation, integration over the
hemisphere is needed, and as part of the function evaluation for each point in the
integration domain, a ray has to be cast and the nearest intersection point located.
Exitant radiance, integration over surfaces
Z
L(x ) = Le (x ) +

(x, y)G(x, y)dA

fr (x, )L(y
yx)V
y

with
G(x, y) =

cos(Nx , )cos(Ny , )
2
rxy
110

Algorithms based on this formulation need to evaluate the visibility V (x, y)

between two points x and y, which is a different operation than casting a ray from
x in a direction .
Incident radiance, integration over the hemisphere
Z
L(x ) = Le (x ) +

fr (y, )L(y ) cos(Ny , )d

y

with
y = r(x, )
Incident radiance, integration over surfaces
Z
L(x ) = Le (x ) +

(y, z)G(y, z)dA

fr (y,
yz)L(y

yz)V
z

with
y = r(x, )

8.2

Importance function

In order to compute the average radiance value over the area of a pixel, one needs
to know the radiant flux over that pixel (and associated solid angle incident w.r.t.
the aperture of the camera). Radiant flux is expressed by integrating the radiance
distribution over all possible surface points and directions. Let S = Ap p denote
all surface points Ap and directions p visible through the pixel. The flux (S) is
written as:
Z Z
(S) =
L(x ) cos(Nx , )d dAx
Ap

When designing algorithms, it is often useful to express the flux as an integral over all possible points and directions in the scene. This can be achieved by
introducing the initial importance function We (x ):
Z Z
(S) =
L(x )We (x ) cos(Nx , )d dAx
A

We (x ) is appropriately defined by:

111

(
1
We (x ) =
0

if (x, ) S
if (x, )
/S

The average radiance value is then given by:

R R
Laverage =

)We (x ) cos(Nx , )d dAx

We (x ) cos(Nx , )d dAx

A RL(x
R
A

We now want to develop the notion of importance further, by considering the

possible influence of some energy value at each pair (x, ) on the value (S). Or:
if a single radiance value L(x ) is placed at (x, ), and if there are no other
sources of illumination present, how large would the resulting value of (S) be?
This influence value attributed to L(x ) is called the importance of (x, )
w.r.t. S, is written as W (x ), and depends only on the geometry and reflective
properties of the objects in the scene.
The equation expressing W (x ) can be derived by taking into account two
mechanisms in which L(x ) can contribute to (S):
Self-contribution If (x, ) S, then L(x ) fully contributes to (S). This
is called the self-importance of the set S, and corresponds to the above definition of We (x ).
Indirect contributions It is possible that some part of L(x ) contributes
to (S) through one or more reflections at several surfaces. The radiance
L(x ) travels along a straight path and reaches a surface point r(x, ).
Energy is reflected at this surface point according to the BRDF. Thus, there
is a hemisphere of directions at r(x, ), each emitting a differential radiance
value as a result of the reflection of the radiance L(r(x, ) ). By
integrating the importance values for all these new directions, we have a new
term for W (x ).
Both terms combined produces the following equation:
Z
W (x ) = We (x )+
fr (z, )W (z ) cos(Nr (x, ), )d
z

with
z = r(x, )
112

Mathematically, this equation is identical to the transport equation of incident

radiance, and thus, the notion incidence can be attributed to importance. The source
function We = 1 if x is visible through the pixel and is a direction pointing
through the pixel to the aperture of the virtual camera.
To enhance the analogy with radiance as a transport quantity, exitant importance can be defined as:
W (x ) = W ((r, ) )
and also:
Z
W (x ) = We (x ) +

fr (x, )W (x )cos(Nx , )d
x

An expression for the flux of through every pixel, based on the importance
function, can now be written. Only the importance of the light sources needs to be
considered when computing the flux:
Z Z
(S) =
Le (x )W (x )cos(Nx , )d dAx
A

It is also possible to write (S) in the following form:

Z Z
Le (x )W (x )cos(Nx , )d dAx
(S) =
A

and also:
Z Z
L(x )We (x ) cos(Nx , )d dAx

(S) =
A

Z Z
L(x )We (x ) cos(Nx , )d dAx

(S) =
A

There are two approaches to solve the global illumination problem: The first
approach starts from the pixel, and the radiance values are computed by solving
one of the transport equations describing radiance. A second approach computes
the flux starting from the light sources, and computes for each light source the
corresponding importance value. If one looks at various algorithms in some more
detail:
113

 Stochastic ray tracing propagates importance, the surface area visible through
each pixel being the source of importance. In a typical implementation, the
importance is never explicitly computes, but is implicitly done by tracing
rays through the scene and picking up illumination values from the light
sources.
Light tracing is the dual algorithm of ray tracing. It propagates radiance
from the light sources, and computes the flux values at the surfaces visible
through each pixel.
Bidirectional ray tracing propagates both transport quantities at the same
time, and in an advanced form, computes a weighted average of all possible
inner products at all possible interactions.

8.3

Path formulation

The above description of global illumination transport algorithms is based on the

notion of radiance and importance. One can also express global transport by considering path-space, and computing a transport measure over each individual path.
Path-space encompasses all possible paths of any length. Integrating a transport
measure in path-space then involves generating the correct paths (e.g. random
paths can be generated using an appropriate Monte Carlo sampling procedure),
and evaluating the throughput of energy over each generated path. This view was
developed by Spanier and Gelbard and introduced into rendering by Veach.
Z
(S) =

f (x)d(x)

in which is the path-space, x is a path of any length and d(x) is a measure

in path space . f (x) describes the throughput of energy and is a succession of
G(x, y), V (x, y) and BRDF evaluations, together with a Le and We evaluation at
the beginning and end of the path.
An advantage of the path formulation is that paths are now considered to be
the sample points for any integration procedure. Algorithms such as Metropolis
light transport or bidirectional ray tracing are often better described using the path
formulation.
114

8.4

Simple stochastic ray tracing

In any pixel-driven rendering algorithm we need to use the rendering equation to

evaluate the appropriate radiance values. The most simple algorithm to compute
this radiance value is to apply a basic and straightforward MC integration scheme
to the standard form of the rendering equation:
L(x ) = Le (x ) + Lr (x )
Z
= Le (x ) +
L(x )fr (x, ) cos(, Nx )d
x

The integral is evaluated using MC integration, by generating N random directions i over the hemisphere x , according to some pdf p(). The estimator for
Lr (x ) is given by:
N
1 X L(x i )fr (x, i ) cos(i , Nx )
hLr (x )i =
N
p(i )
i=1

L(x i ), the incident radiance at x, is unknown. It is now necessary to trace

the ray leaving x in direction i through the scene to find the closest intersection
point r(x, ). Here, another radiance evaluation is needed. The result is a recursive
procedure to evaluate L(x i ), and as a consequence, a path, or a tree of paths
if N > 1, is generated in the scene.
These radiance evaluations will only yield a non-zero value, if the path hits
a surface for which Le has a value different from 0. In other words, in order to
compute a contribution to the illumination of a pixel, the recursive path needs to
reach at least one of the light sources in the scene. If the light sources are small,
the resulting image will therefore mostly be black. This is expected, because the
algorithm generates paths, starting at a point visible through a pixel, and slowly
working towards the light sources in a very uncoordinated manner.

8.5

Russian Roulette

The recursive path generator described above needs a stopping condition to prevent
the paths being of infinite length. We want to cut off the generation of paths, but at
the same time, we have to be very careful about not introducing any bias into the
115

image generations process. Russian Roulette addresses the problem of keeping the
lengths of the paths manageable, but at the same time leaves room for exploring all
possible paths of any length. Thus, an unbiased image can still be produced.
The idea of Russian Roulette can best be explained by a simple example: suppose one wants to compute a value V . The computation of V might be computationally very expensive, so we introduce a random variable r, which is uniformly distributed over the interval [0, 1]. If r is larger than some threshold value
[0, 1], we proceed with computing V . However, if r , we do not compute
V , and assume V = 0. Thus, we have a random experiment, with an expected
value of (1 )V . By dividing this expected value by (1 ), an unbiased estimator for V is maintained.
If V requires recursive evaluations, one can use this mechanism to stop the
recursion. is called the absorption probability. If is small, the recursion will
continue many times, and the final computed value will be more accurate. If is
large, the recursion will stop sooner, and the estimator will have a higher variance.
In the context of our path tracing algorithm, this means that either accurate paths
of a long length are generated, or very short paths which provide a less accurate
estimate.
In principle any value for can be picked, thus controlling the recursive depth
and execution time of the algorithm. 1 is often set to be equal to the hemispherical reflectance of the material of the surface. Thus, dark surfaces will absorb
the path more easily, while lighter surfaces have a higher chance of reflecting the
path.

8.6

Indirect Illumination

In most path tracing algorithms, direct illumination is explicitly computed separately from all other forms of illumination (see previous chapter on direct illumination). This section outlines some strategies for computing the indirect illumination
in a scene. Computing the indirect illumination is usually a harder problem, since
one does not know where most important contributions are located. Indirect illumination consists of the light reaching a target point x after at least one reflection
at an intermediate surface between the light sources and x.
116

8.6.1

Hemisphere sampling

The rendering equation can be split in a direct and indirect illumination term. The
indirect illumination (i.e. not including any direct contributions from light sources
to the point x) contribution to L(x ) is written as:
Z
Lindirect (x ) =

Lr (r(x, ) )fr (x, ) cos(, Nx )d

x

The integrand contains the reflected terms Lr from other points in the scene,
which are themselves composed of a direct and indirect illumination part. In a
closed environment, Lr (r(x, ) ) usually has a non-zero value for all (x, )
pairs. As a consequence, the entire hemisphere around x needs to be considered as
the integration domain.
The most general MC procedure to evaluate indirect illumination, is to use any
hemispherical pdf p(), and generating N random directions i . This produces
the following estimator:

hLindirect (x )i =

N
1 X Lr (r(x, i ) i )fr (x, i ) cos(i , Nx )
N
p(i )
i=1

In order to evaluate this estimator, for each generated direction i , the BRDF
and the cosine term are to be evaluated, a ray from x in the direction of i needs
to be traced, and the reflected radiance Lr (r(x, i ) i ) at the closest intersection point r(x, i ) has to be evaluated. This last evaluation shows the recursive
nature of indirect illumination, since this reflected radiance at r(x, i ) can be split
again in a direct and indirect contribution.
The simplest choice for p() is p() = 1/2, such that directions are sampled
proportional to solid angle. Noise in the resulting picture will be caused by variations in the BRDF and cosine evaluations, and variations in the reflected radiance
Lr at the distant points.
The recursive evaluation can again be stopped using Russian Roulette, in the
same way as was done for simple stochastic ray tracing. Generally, the local hemispherical reflectance is used as an appropriate absorption probability. This choice
can be explained intuitively: One only wants to spend work (i.e. tracing rays and
evaluating Lindirect (x)) proportional to the amount of energy present in different
parts of the scene.
117

8.6.2

Importance sampling

Uniform sampling over the hemisphere does not use any knowledge about the integrand in the indirect illumination integral. However, this is necessary to reduce
noise in the final image, and thus, some form of importance sampling is needed.
Hemispherical pdfs proportional (or approximately proportional) to any of the following factors can be constructed:
Cosine sampling
Sampling directions proportional to the cosine lobe around the normal Nx prevents directions to be sampled near the horizon of the hemisphere where cos(, Nx )
yields a very low value, and thus possibly insignificant contributions to the computed radiance value.
BRDF sampling
BRDF sampling is a good noise-reducing technique when a glossy or highly
specular BRDFs is present. It diminishes the probability that directions are sampled where the BRDF has a low value or zero value. Only for a few selected BRDF
models, however, is it possible to sample exactly proportional to the BRDF. Even
better would be trying to sample proportional to the product of the BRDF and the
cosine term. Analytically, this is even more difficult to do, except in a few rare
cases where the BRDF model has been chosen carefully.
Incident radiance field sampling
A last technique that can be used to reduce variance when computing the indirect illumination is to sample a direction according to the incident radiance
values Lr (x ). Since this incident radiance is generally unknown, an adaptive
technique needs to be used, where an approximation of Lr (x ) is constructed
during the execution of the rendering algorithm.

8.6.3

Overview

It is now possible to build a full global illumination renderer using stochastic path
tracing. The efficiency, accuracy and overall performance of the complete algorithm will be determined by the choice of all of the following parameters. As is
usual in MC evaluations, the more samples or rays are generated, the less noisy the
final image will be.
Number of viewing rays per pixel The amount of viewing rays through each pixel
is responsible for effects such as aliasing at visible boundaries of objects or
shadows.
118

Direct Illumination:
The total number of shadow rays generated at each surface point x;
The selection of a single light source for each shadow ray;
The distribution of the shadow ray over the area of the selected light
source.
Indirect Illumination (hemisphere sampling):
Number of indirect illumination rays;
Exact distribution of these rays over the hemisphere (uniform, cosine,
...);
Absorption probabilities for Russian Roulette.
The better one makes use of importance sampling, the better the final image
and the less noise there will be. An interesting question is, given a maximum
amount of rays one can use per pixel, how should these rays best be distributed to
reach the highest possible accuracy for the full global illumination solution? This
is still an open problem. There are generally accepted default choices, but there
are no hard and fast choices. It generally is accepted that branching out equally at
all levels of the tree is less efficient. For indirect illumination, a branching factor
of 1 is often used after the first level. Many implementations even limit the indirect
rays to one per surface point, and compensate by generating more viewing rays.

119

120

Chapter 9

Metropolis Sampling
By Matt Pharr

A new approach to solving the light transport problem was recently developed by
Veach and Guibas, who applied the Metropolis sampling algorithm [91, 87] (first
introduced in Section 2.3.3 of these notes.)1 . The Metropolis algorithm generates a
series of samples from a non-negative function f that are distributed proportionally
to f s value [52]. Remarkably, it does this without requiring anything more than the
ability to evaluate f ; its not necessary to be able to integrate f , normalize it, and
invert the resulting pdf. Metropolis sampling is thus applicable to a wider variety
of sampling problems than many other techniques. Veach and Guibas recognized
that Metropolis could be applied to the image synthesis problem after it was appropriately reformulated; they used it to develop a general and unbiased Monte Carlo
rendering algorithm which they named Metropolis Light Transport (MLT).
MLT is notable for its robustness: while it is about as efficient as other unbiased techniques (e.g. bidirectional ray tracing) for relatively straightforward lighting problems, it distinguishes itself in more difficult settings where most of the
light transport happens along a small fraction of all of the possible paths through
the scene. Such settings were difficult for previous algorithms unless they had specialized advance-knowledge of the light transport paths in the scene (e.g. a lot of
light is coming through that doorway); they thus suffered from noisy images due
1

We will refer to the Monte Carlo sampling algorithm as the Metropolis algorithm here. Other
commonly-used shorthands for it include M(RT)2 , for the initials of the authors of the original paper,
and Metropolis-Hastings, which gives a nod to Hastings, who generalized the technique [22]. It is
also commonly known as Markov Chain Monte Carlo.

121

to high variance because most of they generated would have a low contribution, but
when they randomly sampled an important path, there would be a large spike in the
contribution to the image. In contrast, the Metropolis method leads to algorithms
that naturally and automatically adapt to the subtleties of the particular transport
problem being solved.
The basic idea behind MLT is that a sequence of light-carrying paths through
the scene is computed, with each path generated by mutating the previous path
in some manner. These mutations are done in a way that ensures that the overall
distribution of sampled paths in the scene is proportional to the contribution these
paths make to the image being generated. This places relatively few restrictions on
the types of mutations that can be applied; in general, it is possible to invent unusual sampling techniques that couldnt be applied to other MC algorithms without
introducing bias.
MLT has some important advantages compared to previous unbiased approaches
to image synthesis:
Path re-use: because paths are often constructed using some of the segments
of the previous one, the incremental cost (i.e. number of rays that need to be
traced) for generating a new path is much less than the cost of generating a
path from scratch.
Local exploration: when paths that make large contributions to the final image are found, its easy to sample other paths that are similar to that one by
making small perturbations to the path.
The first advantage increases overall efficiency by a relatively fixed amount
(and in fact, path re-use can be applied to some other light transport algorithms.)
The second advantage is the more crucial one: once an important transport path
has been found, paths that are similar to it (which are likely to be important) can
be sampled. When a function has a small value over most of its domain and a large
contribution in only a small subset of it, local exploration amortizes the expense
(in samples) of the search for the important region by letting us stay in that area for
a while.
In this chapter, we will introduce the Metropolis sampling algorithm and the
key ideas that it is built on. We will then show how it can be used for some lowdimensional sampling problems; this setting allows us to introduce some of the
important issues related to the full Metropolis Light Transport algorithm without
122

getting into all of the tricky details. We first show its use in one-dimension. We
then demonstrate how it can be used to compute images of motion-blurred objects;
this pushes up the domain of the problem to three dimensions and also provides
a simpler setting in which to lay more groundwork. Finally, we will build on
this basis to make connections with and describe the complete MLT algorithm.
We will not attempt to describe every detail of MLT here; however the full-blown
presentation in MLT paper [91] and the MLT chapter in Veachs thesis [87] should
be much more approachable with this groundwork.

9.1

Overview

The Metropolis algorithm generates a set of samples xi from a function f , which

is defined over a state space , f : R. (See Figure 9.1 for notation used
in this chapter.) After the first sample x0 is selected (more details on this later),
each subsequent sample xi is generated by proposing a random mutation to xi1
to compute a proposed sample x0 . The mutation may be accepted or rejected, and
xi is set to either x0 or xi1 , respectively. When these transitions from one state
to another are chosen subject to a few limitations (described shortly), the distribution of xi values that results eventually reaches equilibrium; this distribution is the
stationary distribution.
The way in which mutations are accepted or rejected guarantees that in the
limit, the distribution of the set of samples xi is proportional to f (x)s density
function. Thus, even if we are unable to integrate f analytically, normalize it, and
invert the integral so that we can sample from it, the Metropolis algorithm still
generates samples from f s normalized density function fpdf .

9.1.1

Detailed Balance

In order to generate this particular distribution of samples, we need to accept or

reject mutations in a particular way. Assume that we have a method of proposing
mutations that makes a given state x into a proposed state x0 (this might be done
by perturbing x in some way, or even by generating a completely new value.) We
also need to be able to compute a tentative transition function T(x x0 ) that
gives the probability density of the mutation techniques proposing a transition to
x0 , given that the current state is x. (In general, the need to be able to compute this
density is the only limitation on how we might choose to mutatetheres a lot of
123


f (x)

x
x0
xi
p(x)
I(f )
fpdf
f(x)
T(x x0 )
a(x x0 )
hj (u, v)
Ij

Uniform random number between 0 and 1

Function being sampled
State space over which f is defined
A sample value, x
A proposed new sample value, based on some mutation strategy
The ith sample in a Markov chain x0 , x1 , . . . , xn generated
by the Metropolis sampling algorithm
A probability density function
R
The integrated value of f (x) over all of , I(f ) = f (x)d
Normalized probability density of f s distribution, fpdf = f /I(f )
Reconstructed function that approximates fpdf
Density of proposed transition from one state x to another x0
Acceptance probability of mutating from one state x to another x0
Value of the jth pixels reconstruction filter for an image
sample at (u, v)
Image function value at pixel j
Figure 9.1: Notation used in this chapter

freedom left over!) We also define an acceptance probability a(x x0 ) that gives
the probability of accepting a proposed mutation from x to x0 .
The key to the Metropolis algorithm is the definition of a(x x0 ) such that
the distribution of samples is proportional to f (x). If the random walk is already
in equlibrium, the transition density between any two states must be equal:2
f (x) T(x x0 ) a(x x0 ) = f (x0 ) T(x0 x) a(x0 x).

(9.1)

This property is called detailed balance. Since f and T are set, Equation 9.1 tells
us how a must be defined. In particular, a definition of a that maximizes the rate at
which equilibrium is reached is


f (x0 ) T(x0 x)
0
a(x x ) = min 1,
(9.2)
f (x) T(x x0 )
One thing to notice from Equation 9.2 is that if the transition probability density
is the same in both directions, the acceptance probability simplifies to


f (x0 )
a(x x0 ) = min 1,
(9.3)
f (x)
2

See Kalos and Whitlock [38] or Veachs thesis [87] for a rigorous derivation.

124

x = x0
for i = 1 to n
x = mutate(x)
a = accept(x, x)
if (random() < a)
x = x
record(x)
Figure 9.2: Pseudocode for the basic Metropolis sampling algorithm. We generate
n samples by mutating the previous sample and computing acceptance probabilities
as in Equation 9.2. Each sample xi is then recorded in a data structure.
For some of the basic mutations that well use, this condition on T will be met,
which simplifies the implementation.
Put together, this gives us the basic Metropolis sampling algorithm shown in
pseudo-code in Figure 9.2. We can apply the algorithm to estimating integrals such
R
as f (x)g(x)d. The standard Monte Carlo estimator, Equation 2.6, says that
Z
f (x)g(x) d

N
1 X f (xi )g(xi )
N
p(xi )

(9.4)

i=1

where xi are sampled from a density function p(x). Thus, if we apply Metropolis
sampling and generate a set of samples x1 , . . . , xN , from a density function fpdf (x)
proportional to f (x), we have
"
#
Z
N
1 X
f (x)g(x) d
g(xi ) I(f )
(9.5)
N

i=1

where I(f ) is the value of f (x) integrated over all of .

9.1.2

Expected Values

Because the Metropolis algorithm naturally avoids parts of where f (x)s value
is relatively low, few samples will be accumulated there. In order to get some
information about f (x)s behavior in such regions, the expected values technique
can be used to enhance the basic Metropolis algorithm.
At each mutation step, we record a sample at both the current sample x and the
proposed sample x0 , regardless of which one is selected by the acceptance criteria.
125

x = x0
for i = 1 to n
x = mutate(x)
a = accept(x, x)
record(x, (1-a) * weight)
record(x, a * weight)
if (random() < a)
x = x
Figure 9.3: The basic Metropolis algorithm can be improved using expected values.
We still decide which state to transition into as before, but we record a sample at
each of x and x0 , proportional to the acceptance probability. This gives smoother
results, particularly in areas where f s value is small, where otherwise few samples
would be recorded.
Each of these recorded samples has a weight associated with it, where the weights
are the probabilities (1a) for x and a for x0 , where a is the acceptance probability.
Comparing the pseudocode in Figures 9.2 and 9.3, we can see that in the limit, the
same weight distribution will be accumulated for x and x0 . Expected values more
quickly gives us more information about the areas where f (x) is low, however.
Expected values doesnt change the way we decide which state, x or x0 to use
at the next step; that part of the computation remains the same.

9.2 One Dimensional Setting

Consider using Metropolis to sample the following one-dimensional function, defined over = [0, 1] and zero everywhere else (see Figure 9.4).

(x .5)2 : 0 x 1
1
(9.6)
f (x) =
0 : otherwise
For this example, assume that we dont actually know the exact form of f 1 its
just a black box that we can evaluate at particular x values. (Clearly, if we knew
that f 1 was just Equation 9.6, thered be no need for Monte Carlo sampling!)
Wed like to generate a set of samples of f 1 using Metropolis sampling. These
samples will be used to reconstruct a new function f1 that approximates f 1 s pdf.
A random choice between the strategies will be made each time a mutation is proposed, according to a desired distribution of how frequently each is to be used.
126

0.3

0.2

0.1

0.0
0.00

0.25

0.50

0.75

1.00

Figure 9.4: Graph of the function f 1 , which is used to demonstrate Metropolis

sampling in this section.

9.2.1

Mutations and Transition Functions

We will first describe two basic mutation strategies, each of which depends on a
uniform random number between zero and one. Our first mutation, mutate1 ,
discards the current sample x and uniformly samples a new one x0 from the entire
state space [0, 1]. Mutations like this one that sample from scratch are important to
make sure that we dont get stuck in one part of state space and never sample the
rest of it (an example of the problems that ensue when this happens will be shown
below.) The transition function for this mutation is straightforward. For mutate1 ,
since we are uniformly sampling over [0, 1], the probability density is uniform over
the entire domain; in this case, the density is just one everywhere. We have

mutate1 (x)
T1 (x x0 )

The second mutation adds a random offset between .05 to the current sample
x in an effort to sample repeatedly in the parts of f that make a high contribution
to the overall distribution. The transition probability density is zero if x and x0 are
far enough away that mutate2 will never mutate from one to the other; otherwise
the density is constant. Normalizing the density so that it integrates to one over its
1
domain gives the value 0.1
.
127

mutate2 (x) x + .1 ( .5)

 1
: |x x0 | .05
0
0.1
T2 (x x ) =
0 : otherwise
The second mutation is important for overall efficiency: when we find a sample
xi where f 1 (xi ) is larger than most other values of f 1 (x), we would like to examine
samples close to xi in the state space, since f (x) is also likely to be large there.
Furthermore, this mutation has the important property that it makes us more likely
to accept proposed mutations: if we only used the first mutation strategy, we would
find it difficult to mutate away from a sample xi where f 1 (xi ) had a relatively large
value; proposed transitions to other states where f 1 (x0 )  f (xi ) are rejected with
high probability, so many samples in a row would be accumulated at xi . (Recall
the definition of the acceptance probability a(x x0 ) in Equation 9.3.) Staying in
the same state for many samples in a row leads to increased varianceintuitively,
it makes sense that the more we move around , the better the overall results will
be. Adding the second mutation to the mix increases the probability of being able
to accept mutations around such samples where f 1 (x) is relatively high, giving a
better distribution of samples in the end.

9.2.2

Start-up bias

Before we can go ahead and use the Metropolis algorithm, one other issue must
be addressed: start-up bias. The transition and acceptance methods above tell us
how to generate new samples xi+1 , but all presuppose that the current sample xi
has itself already been sampled with probability proportional to f . A commonly
used solution to this problem is to run the Metropolis sampling algorithm for some
number of iterations from an arbitrary starting state, discard the samples that are
generated, and then start the process for real, assuming that that has brought us to
an appropriately sampled x value. This is unsatisfying for two reasons: first, the
expense of taking the samples that were then discarded may be high, and second,
we can only guess at how many initial samples must be taken in order to remove
start-up bias.
Veach proposes another approach which is unbiased and straightforward. If an
alternative sampling method is available, we sample an initial value x0 using any
density function x0 p(x). We start the Markov chain from the state x0 , but we
128

weight the contributions of all of the samples that we generate by the weight
w=

f (x0 )
.
p(x0 )

This method eliminates start-up bias completely and does so in a predictable manner.
The only potential problem comes if f (x0 ) = 0 for the x0 we chose; in this
case, all samples will have a weight of zero, leading to a rather boring result. This
doesnt mean that the algorithm is biased, however; the expected value of the result
still converges to the correct distribution (see [87] for further discussion and for a
proof of the correctness of this technique.)
To reduce variance from this step, we can instead sample a set of N candidate
sample values, y1 , . . . , yN , defining a weight for each by
wi =

f (yi )
.
p(yi )

(9.7)

We then choose the starting x0 sample for the Metropolis algorithm from the yi
with probability proportional to their relative weights and compute a sample weight
w as the average of all of the wi weights. All subsequent samples xi that are
generated by the Metropolis algorithm are then weighted by the sample weight w.
For our particular f 1 example above, we only need to take a single sample with
a uniform pdf over , since f 1 (x) > 0 except for a single point in which there
is zero probability of sampling.
x0 =
The sample weight w is then just f 1 (x0 ).

9.2.3

Initial Results

We can now run the Metropolis algorithm and generate samples xi of f 1 . At each
transition, we have two weighted samples to record (recall Figure 9.3.) A simple
approach for reconstructing the approximation to f 1 s probability distribution f1 is
just to store sums of the weights in a set of buckets of uniform width; each sample
falls in a single bucket and contributes to it. Figure 9.5 shows some results. For
both graphs, we followed a chain of 10,000 mutations, storing the sample weights
in fifty buckets over [0, 1]. The weighting method for eliminating start-up bias was
used.
129

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0.0

0.0
0.00

0.25

0.50

0.75


1.00

0.00

0.25

0.50

0.75


1.00

Figure 9.5: On the left, we always mutate by randomly selecting a completely new
x value. Convergence is slow, but the algorithm is finding the right distribution.
On the right, we perturb the current sample by .05 90% of the time, and pick a
completely new x value the remaining 10%.

On the left graph, we used only mutate1 when a new x0 value is to be proposed.
This alone isnt a very useful mutation, since it doesnt let us take advantage of the
times when we find ourselves in a region of where f has a relatively large value
and generate many samples in that neighborhood. However, the graph does suggest
that the algorithm is converging to the correct distribution.
On the right, we randomly chose between mutate1 and mutate2 with probabilities of 10% and 90%, respectively. We see that for the same number of samples
taken, we converge to f s distribution with less variance. This is because we are
more effectively able to concentrate our work in areas where f s value is large, and
propose fewer mutations to parts of state space where f s value is low. For example, if x = .8 and the second mutation proposes x0 = .75, this will be accepted
f (.75)/f (.8) 69% of the time, while mutations from .75 to .8 will be accepted
min(1, 1.44) = 100% of the time. Thus, we see how the algorithm naturally tends
to try to avoid spending time sampling around dip in the middle of the curve.
One important thing to note about these graphs is that the y axis has units that
are different than those in Figure 9.4, where f 1 is graphed. Recall that we just
1 ; as such
have a set of samples distributed according to the probability density fpdf
(for example), we would get the same sample distribution for another function
g = 2f 1 . If we wish to reconstruct an approximation to f 1 directly, we must
1 . We explain this process in
compute a normalization factor and use it to scale fpdf
130

more detail in Section 9.3.2.

9.2.4

Ergodicity

Figure 9.6 shows the surprising result of what happens if we only use mutate2 to
suggest sample values. On the left, we have taken 10,000 samples using just that
mutation. Clearly, things have gone awrywe didnt generate any samples xi > .5
and the result doesnt bear much resemblance to f 1 .
Thinking about the acceptance probability again, we can see that it would take
a large number of mutations, each with low probability of acceptance, to move xi
down close enough to .5 such that mutate2 s short span would be enough to get us
to the other side. Since the Metropolis algorithm tends to keep us away from the
lower-valued regions of f (recall the comparison of probabilities for moving from
.8 to .75, versus moving from .75 to .8), this happens quite rarely. The right side
of Figure 9.6 shows what happens if we take 300,000 samples. This was enough
to make us jump from one side of .5 to the other a few times, but not enough to get
us close to the correct distribution.
This problem is an example of a more general issue that must be addressed
with Metropolis sampling: its necessary that it be possible to reach all states
x where f (x) > 0 with non-zero probability. In particular, it suffices that
T(x x0 ) > 0 for all x and x0 where f (x) > 0 and f (x0 ) > 0. Although the
first condition is in fact met when we use only mutate2 , many samples would be
necessary in practice to converge to an accurate distribution. Periodically using
mutate1 ensures sufficiently better coverage of such that that this problem goes
away.

9.2.5

Mutations via PDFs

If we have a pdf that is similar to some component of f , then we can use that
to derive one of our mutation strategies as well. Note that if we had a pdf that
was exactly proportional to f , all this Metropolis sampling wouldnt be necessary,
but lacking that we often can still find pdfs that approximate some part of the
function being sampled. Adding such an extra mutation strategy to the mix can
improve overall robustness of the Metropolis algorithm, by ensuring good coverage
of important parts of state space.
If we can generate random samples from a probability density function p, x
131

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0.0

0.0
0.00

0.25

0.50

0.75


1.00

0.00

0.25

0.50

0.75


1.00

Figure 9.6: Two examples that show why it is important to periodically pick a
completely new sample value. On the left, we ran 10,000 iterations using only
mutate2 , and on the right, 300,000 iterations. It is very unlikely that a series of
mutations will be able to move from one side of the curve, across 0.5, to the other
side, since mutations to areas where f 1 s value is low will usually be rejected. As
such, the results are inaccurate for these numbers of iterations. (Its small solace
that they would be correct in the limit.)

p, the transition function is straightforward:

T(x x0 ) = p(x0 ).
i.e. the current state x doesnt matter for computing the transition density: we
propose a transition into a state x0 with a density that depends only on the newly
proposed state x0 and not at all on the current state.
To apply this to the one-dimensional f 1 example, we add a mutation based on
a linear probability density function p1 that is somewhat similar to the shape of f 1 .
(see Figure 9.7).

1.2 : x 1/3
.6 : 1/3 < x 2/3
p1 (x) =

1.2 : 2/3 < x

Note that p1 (x) is a valid probability density function; it integrates to one over
the domain [0, 1]. We can apply the function inversion technique described in SecRx
tion 2.3.1 to derive a sampling method. Inverting the integral = 0 p1 (x)dx
132

1.0

0.5

0.0
0.00

0.25

0.50

0.75


1.00

Figure 9.7: Linear pdf used to sample f 1 . We can also develop mutation strategies based on pdfs that approximate some component of the function that were
sampling. Here, were using a simple linear function that is roughly similar to the
shape of f 1 .
gives us:
x=

1
3
2
3

+
+

1
3
1
3

1
3 .4
(.4)
.2
(.6)
.4

: .4
: .4 < .6
: > .6

(9.8)

Our third mutation strategy, mutate3 , just generates a uniform random number
and proposes a mutation to a new state x0 according to Equation 9.8. Results
of using this mutation alone are shown in Figure 9.8; the graph is not particularly
better than the previous results, but in any case, it is helpful to have a variety of
methods with which to develop mutations.

9.3

Motion Blur

We will now show how Metropolis sampling can be used to solve a tricky problem with rendering motion-blurred objects. For objects that are moving at a high
speed across the image plane, the standard distribution ray tracing approach can be
inefficient; if a pixel is covered by an object for only a short fraction of the overall
time, most of the samples taken in that pixel will be black, so that a large number
are needed to accurately compute the pixels color. We will show in this section
how Metropolis can be applied to solve this problem more efficiently. In addition
to being an interesting application of Metropolis, this also gives us an opportunity
133

0.8

0.6

0.4

0.2

0.0
0.00

0.25

0.50

0.75


1.00

Figure 9.8: Result of reconstructing f 1 s pdf using the new mutate3 .

to lay further groundwork before moving onto the complete MLT algorithm.
The basic setting of the motion blur problem is that we have a two-dimensional
image plane, with some number of pixels, pu and pv , in the u and v directions. The
time range is from zero to one. We define the scenes radiance function L(u, v, t),
which gives the radiance visible along the ray through the scene from the position
(u, v) on the image plane, at time t. (L can be computed with standard ray tracing
methods.) The state space for this problem is thus the triples (u, v, t) R3 ,
where Ls value is zero for (u, v) outside the image and where t < 0 or t > 1. The
measure is just the product measure du dv dt.
The key step to being able to apply Metropolis sampling to image synthesis
problems is the definition of the image contribution function [87, Chapter 11, Section 3]. For an image with j pixels, each pixel Ij has a value that is the product of
the pixels image reconstruction filter hj and the radiance L that contributes to the
image:
Z
Ij =

hj (u, v) L(u, v, t) du dv dt

The standard Monte Carlo estimate of Ij is3

Ij

N
1 X hj (xi ) L(xi )
,
N
p(xi )

(9.9)

i=1

where p is a probability density function used for sampling xi and where we

3
Because the filter support of hj is usually only a few pixels wide, a small number of samples xi
will contribute to each pixel.

134

have written hj and L as functions of samples xi R3 (even though hj typically

only depends on the u and v sample location of xi .)
In order to be able to apply Metropolis sampling to the problem of estimating
pixel values Ij , we can apply Equation 9.5 to rewrite this as
Ij

Z

N
1 X
hj (xi )
L(x) d ,
N

(9.10)

i=1

since the Metropolis sampling algorithm generates a set of samples xi according

to the distribution that exactly matches Ls probability density function.
In the remainder of this section, we will describe an application of Metropolis sampling to the motion blur problem and compare the results to standard approaches. Our example scene is a series of motion-blurred balls, moving across the
screen at increasing velocities from top-to-bottom (see Figure 9.9).

9.3.1

Basic Mutation Strategies

We will start with two basic mutation strategies for this problem. First, to ensure
ergodicity, we generate a completely new sample 10% of the time. This is just like
the one dimensional case of sampling f 1 . Here, we choose three random numbers
from the range of valid image and time sample values.
Our second mutation is a pixel and time perturbation that helps with local exploration of the space. Each time it is selected, we randomly move the pixel sample
location up to 8 pixels in each direction (for a 512 by 512 image), and up to .01
in time. If we propose a mutation that takes us off of the image or out of the valid
time range, the transition is immediately rejected. The performance of the algorithm isnt too sensitive to these values, though see below for the results of some
experiments where they were pushed to extremes.
The transition probabilities for each of these mutations are straightforward,
analogous to the one dimensional examples.
Figure 9.9 shows some results. The top image was rendered with distribution
ray tracing (with a stratified sampling pattern), and the bottom the Metropolis sampling approach with these two mutations was used. The sample total number of
samples was taken for each. Note that Metropolis does equally well regardless of
the velocity of the balls, while fast moving objects are difficult for distribution ray
tracing to handle well. Because Metropolis sampling can locally explore the path
space after it has found a sample that hits one of the balls, it is likely to find other
135

samples that hit them as well, thus being more efficientthe small time perturbation is particularly effective for this. Note, however, that Metropolis doesnt do as
well with the ball that is barely moving at all, while this is a relatively easy case
for stratified sampling to handle well.
Its interesting to see the effect of varying the parameters to mutate2 . First, we
tried greatly increasing the amount we can move in , up to 80 pixels in each
direction and .5 in time. Figure 9.10 (top) shows the result. Because we are no
longer doing a good job of local exploration of the space, the image is quite noisy.
(One would expect it to degenerate to something like distribution ray tracing, but
without the advantages of stratified sampling patterns that are easily applied in that
setting.)
We then dialed down the parameters, to .5 pixels of motion and .001 in
time, for a single mutation; see Figure 9.10 (bottom). Here the artifacts in the
image are more clumpythis happens because we find a region of state space with
a large contribution but then have trouble leaving it and finding other important
regions. As such, we dont do a good job of sampling the entire image.
As a final experiment, we replaced mutate2 with a mutation that sampled the
pixel and time deltas from an exponential distribution, rather than a uniform distribution. Given minimum and maximum pixel offsets rmax and rmin and time offsets
tmax and tmin , we computed
r = rmax e log(rmax /rmin )
dt = tmax e log(tmax /tmin )
Given these offsets, a new pixel location was computed by uniformly sampling
an angle = 2, and the new image (u, v) coordinates were computed by
(u, v) = (r sin , r cos )
The new time value was computed by offsetting the old one by dt, where addition
and subtraction were chosen with equal probability.
We rendered an image using this mutation; the range of pixel offsets was
(.5, 40), and the range of time deltas was (.0001, .03). Figure 9.11 shows the result. The improvement is not major, but is noticeable. In particular, see how noise
is reduced along the edges of the fast-moving ball.
The advantage of sampling with an exponential distribution like this is that it
naturally tries a variety of mutation sizes. It preferentially makes small mutations,
136

Figure 9.9: Basic motion blur results. On the top, we have applied distribution
ray tracing with a stratified sampling pattern, and on the bottom, we have applied
Metropolis sampling. The images are 512x512 pixels, with an average of 9 samples
per pixel.

137

Figure 9.10: The effect of changing the parameters to the second motion blur mutation strategy. Top: if the mutations are too large, we get a noisy image. Bottom:
too small a mutation leads to clumpy artifacts, since were not doing a good job of
exploring the entire state space.

138

close to the minimum magnitudes specified, which help locally explore the path
space in small areas of high contribution where large mutations would tend to be
rejected. On the other hand, because it also can make larger mutations, it also
avoids spending too much time in a small part of path space, in cases where larger
mutations have a good likelihood of acceptance.

9.3.2

Re-normalization

Recall that the set of samples generated by the Metropolis algorithm is from the
normalized distribution fpdf of the function that we apply it to. When we are
computing an image, as in the motion blur example, this means that the images
pixel values need to be rescaled in order to be correct.
This problem can be solved with an additional short pre-processing step. We
take a small number of random samples (e.g. 10,000) of the function f and estimate
its integral over . After applying Metropolis sampling, we have an image of
sample densities. We then scale each pixel by the precomputed total image integral
divided by the total number of Metropolis samples taken. It can be shown that the
expected value is the original function f :
N
1 X
f (x)
f (xi ) I(f )
N
i=1

9.3.3

Adapting for large variation in f

When the image function I has regions with very large values, Metropolis sampling
may not quite do what we want. For example, if the image has a very bright light
source directly visible in some pixels, most of the Metropolis samples will naturally
be clustered around those pixels. As a result, the other regions of the image will
have high variance due to under-sampling. Figure 9.12 (top) shows an example
of this problem. The bottom ball has been made a very bright red, 1,000 times
brighter than the others; most of the samples concentrate on it, so the other balls
are under-sampled.
Veach introduced two-stage Metropolis to deal with this problem [87]. Twostage Metropolis attempts to keep the relative error at all pixels the same, rather
than trying to just minimize absolute error. We do this by renormalizing the function L to get a new function L0 :
L0 (x) =
139

L(x)
n(x)

Figure 9.11: Comparison of Metropolis sampling with the first two mutation strategies (top) versus Metropolis sampling where the second strategy is replaced a mutation based on sampling pixel and time offsets from an exponential distribution
(bottom). Note how noise is reduced along the edges of the fast-moving ball.

140

where n is a normalization function that roughly approximates L(x)s magnitude, such that the range of values taken on by L0 (x) is much more limited. The
Metropolis algorithm progresses as usual, just evaluating L0 (x) where it otherwise would have evaluated L(x). The result is that samples are distributed more
uniformly, resulting in a better image. We correct for the normalization when accumulating weights in pixels; by multiplying each weight by n(x), the final image
pixels have the correct magnitude.
For our example, we computed a normalization function by computing a lowresolution image (32 by 32 pixels) with distribution ray tracing and then blurring
it. We then made sure that all pixels of this image had a non-zero value (we dont
want to spend all of our sampling budget in areas where we inadvertently underestimated n(x), such that L0 (x) = L(x)/n(x) is large) and so we also set pixels
in the normalization image with very low values to a fixed minimum. Applying
Metropolis as before, we computed the image on the bottom of Figure 9.12. Here
all of the balls have been sampled well, resulting in a visually more appealing result (even though absolute error is higher, due to the red ball being sampled with
fewer samples.)

9.3.4

Color

For scenes that arent just black-and-white, the radiance function L(u, v, t) returns
a spectral distribution in some form. This distribution must be converted to a single
real value in order to compute acceptance probabilities (Equation 9.2). One option
is to compute computing the luminance of the spectrum and use that; the resulting
image (which is still generated by storing spectral values) is still correct, and there
is the added advantage that the image is sampled according to its perceived visual
importance.

141

Figure 9.12: Two stage sampling helps when the image has a large variation in
pixel brightness. Top: the red ball is much brighter than the others, resulting in
too few samples being taken over the rest of the image. Bottom: by renormalizing
the image function I, we distribute samples more evenly and generate a less noisy
image.
142

x1
x0
x3
x2

Figure 9.13: A path with three edges through a simple scene. The first vertex v0 is
on a light source, and the last, v3 is at the eye.

9.4

Metropolis Light Transport

Using the groundwork of the last few sections, the Metropolis Light Transport
algorithm can now be described. We will not explain all of its subtleties or all of
the details of how to implement it efficiently; see Chapter 11 of Veachs thesis for
both of these in great detail [87]. Rather, we will try to give a flavor for how it
all works and will make connections between MLT and the sampling problems we
have described in the previous few sections.
In MLT, the samples x from are now sequences v0 v1 . . . vk , k 1, of vertices
on scene surfaces. The first vertex, v0 , is on a light source, and the last, vk is at
the eye (see Figure 9.13). This marks a big change from our previous examples:
the state space is now an infinite-dimensional space (paths with two vertices, paths
with three vertices, ...). As long as there is non-zero probability of sampling any
particular path length, however, this doesnt cause any problems theoretically, but
its is another idea that needs to be juggled when studying the MLT algorithm.
As before, the basic strategy is to propose mutations x0 to paths x, accepting
or rejecting mutations according to the detailed balance condition. The function
f (x) represents the differential radiance contribution carried along the path x, and
the set of paths sampled will be distributed according to the image contribution
function 9.10. (See [87] for a more precise definition of f (x).) Expected values are
also used as described previously to accumulate contributions at both the current
path xs image location, as well as the image location for the proposed path x0 .
A set of n starting paths xi are generated with bidirectional path tracing, in a
manner that eliminates startup bias. N candidate paths are sampled (recall Section 9.2.2), where n  N and we select n of them along with appropriate weights.
143

v6
v5

v4

v5'

v6
v5

v4'
v1

v1

v1
v0

v0

v3'

v0

v3
v2'

v2

Figure 9.14: Example of a bidirectional mutation to a path. Left: we have a path

of length 6 from the eye to the light source. Each vertex of the path represents a
scattering event at a surface in the scene. Middle: the bidirectional mutation has
decided to remove a sub-path from the middle of the current path; we are left with
two short paths from the eye and from the light. Right: we have decided to add one
vertex to each of the paths; their locations are computed by sampling the BRDF at
the preceding vertices. We then trace a shadow ray between the new vertices v20
and v30 to complete the path.
We start with n separate paths, rather than just one, primarily to be able to improve
stratification of samples within pixels (see Section 9.4.2 below as well as [87, Section 11.3.2].)

9.4.1

Path Mutations

A small set of mutations is applied to the paths in an effort to efficiently explore

the space of all possible paths through the scene. The most important of the mutations is the bidirectional mutation. The bidirectional mutation is conceptually quite
straightforward; a subpath from the middle of the current path is removed, and the
two remaining ends are extended with zero or more new vertices (see Figure 9.14).
The resulting two paths are then connected with a shadow ray; if it is occluded,
the mutation is rejected, while otherwise the standard acceptance computation is
performed.
Computation of the acceptance probability for bidirectional mutations requires
that we figure out the values of the path contribution function f (x) and f (x0 ) and
the pair of T(x x0 ) densities. Recall from the introduction of the path integral
that that f (x) is a product of emitted importance from the eye, BRDFs values
along the path, geometry terms G(p, p0 ) = cos i cos o /r2 , and the differential
irradiance from the light source; as such, computation of two the f (x) values can
be simplified by ignoring the terms of each of them that are shared between the two
144

paths, since they just cancel out when the acceptance probability is computed.
Computation of the proposed transition densities is more difficult; see [87, Section 11.4.2.1] for a full discussion. The basic issue is that it is necessary to consider
all of the possible ways that one could have sampled the path you ended up with,
given the starting path. (For example, for the path in Figure 9.14, we might have
generated the same path by sampling no new vertices from the light source, but
two new vertices along the eye path.) This computation is quite similar in spirit to
how importance sampling is applied to bidirectional path tracing.
The bidirectional mutation by itself is enough to ensure ergodicity; because
there is some probability of throwing away the entire current path, we are guaranteed to not get stuck in some subset of path space.
Bidirectional mutations can be ineffective when a very small part of path space
is where the most important light transport is happeningalmost all of the proposed mutations will cause it to leave the interesting set of paths (e.g. those causing
a caustic to be cast from a small specular object.) This problem can be ameliorated
by adding perturbations to the mix; these perturbations try to offset some of the
vertices of the current path from their current location, while still leaving the path
mostly the same (e.g. preserving the mode of scatteringspecular or non-specular,
reflection or transmission, at each scattering event.)
One such perturbation is the caustic perturbation (see Figure 9.15). If the current path hits one or more specular surfaces before hitting a single diffuse surface
and then the eye, then its a caustic path. For such paths, we can make a slight
shift to the outgoing direction from the light source and then trace the resulting
path through the scene. If it hits all specular surfaces again and if the final diffuse
surface hit is visible to the eye, we have a new caustic sample at a different image location. The caustic perturbation thus amortizes the possibly high expense of
finding caustic paths.
Lens perturbations are based on a similar idea to caustic perturbations, where
the direction of outgoing ray from the camera is shifted slightly, and then followed
through the same set of types of scattering at scene surfaces. This perturbation
is particularly nice since it keeps us moving over the image plane, and the more
differently-located image samples we have, the better the quality of the final image.
145

non-specular

specular

Figure 9.15: The caustic perturbation in action. Given an old path that leaves the
light source, hits a specular surface, and then hits a non-specular surface before
reaching the eye, we perturb the direction leaving the light source. We then trace
rays to generate a new path through the scene and to the eye. (The key here is
that because the last surface is non-specular, we arent required to pick a particular
outgoing direction to the eyewe can pick whatever direction is needed.)

9.4.2

Pixel Stratification

Another problem with using Metropolis to generate images is that random mutations wont do a good job of ensuring that pixel samples are well-stratified over
the image plane. In particular, it doesnt even ensure that all of the pixels have any
samples taken within them. While the resulting image is still unbiased, it may be
perceptually less pleasing than an image computed with alternative techniques.
Veach has suggested a lens subpath mutation to address this problem. A set
of pixel samples that must be taken is generated (e.g. via a Poisson-disk process,
a stratified sampling pattern, etc.) As such, each pixel has some number of required sample locations associated with it. The new mutation type first sees if the
pixel corresponding to the current sample path has any precomputed sample positions that havent been used. If so, it mutates to that sample and traces a new
path into the scene, following as many specular bounces as are found until a nonspecular surface is found. This lens subpath is then connected with a path to a light
source. If the randomly selected pixel does have its quota of lens subpath mutations already, the other pixels are examined in a pseudo-random ordering until one
is found with remaining samples.
By ensuring a minimum set of well-distributed samples that are always taken,
overall image quality improves and we do a better job of (for example) anti-aliasing
geometric edges than we would to otherwise. Remember that this process just sets
a minimum number of samples that are taken per pixelif the number of samples
allocated to ensuring pixel stratification is 10% of the total number of samples
146

(for example), then most of our samples will still be taken in regions with high
contribution to the final image.

9.4.3

Direct Lighting

Veach notes that MLT (as described so far) often doesnt do as well with direct
lighting as standard methods, such as those described in Chapter 3 of these notes,
or in Shirley et als TOG paper [70]. The root of the problem is that the Metropolis
samples over the light sources arent as well stratified as they can be with standard
methods.
A relatively straightforward solution can be applied: when a lens subpath is
generated (recall that lens subpaths are paths from the eye that follow zero or more
specular bounces before hitting a diffuse surface), standard direct lighting techniques are used at the diffuse surface to compute a contribution to the image for
the lens subpath. Then, whenever a MLT mutation is proposed that includes direct
lighting, it is immediately rejected since direct lighting was already included in the
solution.
Veach notes, however, that this optimization may not always be more effective.
For example, if the direct lighting cannot be sampled efficiently by standard techniques (e.g. due to complex visibility, most of the light sources being completely
occluded, etc.), then MLT would probably be more effective.

9.4.4

Participating Media

Pauly et al have described an extension of MLT to the case of participating media [62]. The state space and path measure are extended to include points in the
volume in addition to points on the surfaces. Each path vertex may be on a surface
or at a point in the scattering volume. The algorithm proceeds largely the same
way as standard MLT, but places some path vertices on scene surfaces and others
at points in the volume. As such, it robustly samples the space of all contributing
transport paths through the medium.
They also describe a new type of mutation, tailored toward sampling scattering
in participating mediathe propagation perturbation. This perturbation randomly
offsets a path vertex along one of the two incident edges (see Figure 9.16). Like
other perturbations, this mutation helps concentrate work in important regions of
the path space.
147

Figure 9.16: For path vertices that are in the scattering volume, rather than at
a surface, the scattering propagation perturbation moves a vertex (shown here as
a black circle) a random distance along the two path edges that are incident the
vertex. Here we have chosen the dashed edge. If the connecting ray to the eye
is occluded, the mutation is immediately rejected; otherwise the usual acceptance
probability is computed.

Acknowledgements
Eric Veach was kind enough to read these notes through a series of drafts and offer
extensive comments and many helpful suggestions for clearer presentation.

148

Chapter 10

Biased Techniques
By Henrik Wann Jensen
In the previous chapters we have seen examples of several unbiased Monte Carlo
ray tracing (MCRT) techniques. These techniques use pure Monte Carlo sampling
to compute the various estimates of radiance, irradiance etc. The only problem
with pure MCRT is variance seen as noise in the rendered images. The only
way to eliminate this noise (and still have an unbiased algorithm) is to sample
more efficiently and/or to use more samples.
In this chapter we will discuss several approaches for removing noise by introducing bias. We will discuss techniques that uses interpolation of irradiance to
exploit the smoothness of the irradiance field. This approach makes it possible to
use more samples at selected locations in the model. We will also discuss photon
mapping, which stores information about the flux in a scene and performs a local
evaluation of the statistics of the stored flux in order to speedup the simulation of
global illumination.

10.1

Biased vs. Unbiased

An unbiased Monte Carlo technique does not have any systematic error. It can be
stopped after any number of samples and the expected value of the estimator will
be the correct value. This does not mean that all biased methods give the wrong
result. A method can converge to the correct result as more samples are used and
still be biased, such methods are consistent.
In chapter 2 it was shown how the integral of a function g(x) can be expressed
149

as the expected value of an estimator where is:

N
1 X g(x)
.
=
N
p(x)

(10.1)

i=1

where p(x) is a p.d.f. distributed according to x such that p(x) > 0 when g(x) > 0.
The expected value of is the value of the integral:
Z
E{} = I =
g(x) d .
(10.2)
xS

Since the expected value of the is our integral is an unbiased estimate of I.

In contrast a biased estimator have some other source of error such that:
E{ } = I +  ,

(10.3)

where  is some error term. For a consistent estimator the error term will diminish
as the number of samples is increased:
lim  = 0 .

(10.4)

Why should we be interested in anything, but unbiased methods? To answer this

problem recall the slow convergence of pure Monte Carlo methods. To render
images without noise it can be necessary to use a very high number of samples. In
addition the eye is very sensitive to the high frequency noise that is typical with
unbiased MCRT methods.
Going into the domain of biased methods we give up the ability to classify
the error on the estimate by looking only at the variance. However the variance
only gives us a probability the error is within a certain range and as such it is not
a precise way of controlling the error. In addition the requirements for unbiased
algorithms are quite restrictive; we cannot easily use adaptive sampling or stop the
sampling once the estimate looks good enough such simple decisions lead to
biased methods [42].
With biased methods other sources of error are introduced to reduce the variance. This can lead to artifacts if the error is uncontrollable, so naturally we want a
consistent method that will converge to the correct result as we use more samples.
It is important to have the right balance between bias and variance. We want to
eliminate the noise, but not introduce other strange artifacts. As we will see in the
following the most common way of reducing noise is to blur the estimates; the eye
150

Figure 10.1: A path traced box scene with 10 paths/pixel. On the left the box has
two diffuse spheres, and on the right box box has a mirror and a glass sphere. Note
how the simple change of the sphere material causes a significant increase in the
noise in the right image.
is fairly insensitive to slowly changing illumination. The trick is to avoid blurring
edges and sharp features in the image and to have control over the amount of blur.
Consider the simple box scene in Figure 10.1. The figure contains two images
of the box scene: one in which the box contains two diffuse spheres, and one in
which the box contains a glass and a mirror sphere. Both images have been rendered with 10 paths per pixel. Even though the illumination of the walls is almost
the same in the two images the introduction of the specular spheres is enough to
make the path tracing image very noisy. In the following sections we will look at a
number of techniques for eliminating this noise without using more samples.

10.2 Filtering Techniques

An obvious idea for reducing the noise in a rendered image is to postprocess the
image and try to filter out the high frequency noise. This can be done to some degree with a low pass filter or with a median filter [23, 47]. The problem with these
filters is that they remove other features than just the noise. Low pass filtering in
particular will blur sharp edges of objects shadows etc. and in general low-pass filtered images look too blurry. Median filtering is much better at removing noise, but
it still introduces artifacts along edges and other places since it cannot distinguish
between noisy pixels and features of the illumination (such as small highlights).
In addition median filtering is not energy preserving. By simply removing outliers
in the image plane it is very likely that we take away energy or add energy to the
151

Figure 10.2: Filtered versions of the image of the box with specular spheres. On
the left is the result of a 3x3 low-pass filter, and on the right the result of a 3x3
median filter.
rendered image. The effect of low-pass and median filtering on the box scene is
shown in Figure 10.2
Several other more sophisticated filtering techniques have been used.
Jensen and Christensen [34] applied a median filter to the indirect illumination
on diffuse surfaces based on the assumption that most of the noise in path tracing is
found in this irradiance estimate. By filtering only this estimate before using it they
removed a large fraction of the noise without blurring the edges, and features such
as highlights and noisy textures. The problem with this approach is that it softens
the indirect illumination and therefore blurs features due to indirect lighting. Also
the technique is not energy-preserving.
Rushmeier and Ward [65] used a energy-preserving non-linear filter to remove
noisy pixels by distributing the extra energy in the pixel over several neighboring
pixels.
McCool [50] used an anisotropic diffusion filter which also preserves energy
and allows for additional input about edges and textures in order to preserve these
features.
Suykens and Willems [79] used information gathered during rendering (they
used bidirectional path tracing) about the probabilities of various events. This enabled them to predict the occurrence of spikes in the illumination. By distributing
the power of the spikes over a larger region in the image plane their method removes noise and preserves energy. The nice thing about this approach is that it
allows for progressive rendering with initially blurry results that converges to the
correct result as more samples are used.
152

10.3

Adaptive Sampling

Another way of eliminating noise in Monte Carlo ray traced images is to use adaptive sampling. Here the idea is to use more samples for the problematic (i.e. noisy)
pixels.
One techniques for doing this is to compute the variance of the estimate based
on the samples used for the pixel [48, 63]. Instead of using a fixed number of
samples per pixel each pixel is sampled until the variance is below a given threshold. An estimate, s2 , of the variance for a given set of samples can be found using
standard techniques from statistics:

N
1 1 X 2
s2 =
Li
N 1 N
i=1

N
1 X
Li
N

!2

(10.5)

i=1

This estimate is based on the assumption that the samples Li are distributed according to a normal distribution. This assumption is often reasonable, but it fails
when for example the distribution within a pixel is bimodal (this happens when a
light source edge passing through a pixel).
Using the variance as a stopping criteria can be effective in reducing noise, but
it introduces bias as shown by Kirk and Arvo [42]. They suggested using a pilot
sample to estimate the number of samples necessary. To eliminate bias the pilot
sample must be thrown away.
The amount of bias introduced by adaptive sampling is, however, usually very
small as shown by Tamstorf and Jensen [80]. They used bootstrapping to estimate
the bias due to adaptive sampling and found that it is insignificant for most pixels
in the rendered image (a notable exception is the edges of light sources).

10.4

Irradiance Caching

Irradiance caching is a technique that exploits the fact that the irradiance field often is smooth [95]. Instead of just filtering the estimate the idea is to cache and
interpolate the irradiance estimates. This is done by examining the samples of the
estimate more carefully to, loosely speaking, compute the expected smoothness of
the irradiance around a given sample location. If the irradiance is determined to be
sufficiently smooth then the estimate is re-used for this region.
153

To understand in more detail how this works let us first consider the evaluation
of the irradiance, E, at a given location, x, using Monte Carlo ray tracing.
M N
XX
Li,j (j , i ) ,
E(x) =
MN

(10.6)

j=1 i=1

where
r
j = sin1

j 1
M

!
and i = 2

i 2
.
N

(10.7)

Here (j , i ) specify a direction on the hemisphere above x in spherical coordinates. 1 [0, 1] and 2 [0, 1] are uniformly distributed random numbers, and
M and N specify the subdivision of the hemisphere. Li,j (j , i ) is evaluated by
tracing a ray in the (j , i ) direction. Note that the formula uses stratification of
the hemisphere to obtain a better estimate than pure random sampling.
To estimate the smoothness of the local irradiance on the surface around the
sample location Ward et al. [95] looked at the distances to the surfaces intersected
by the rays as well as the local changes in the surface normal. This resulted in
an estimate of the local relative change, i , in irradiance as the surface location is
changed away from sample i:
i (x, ~n) =

||xi x|| p
+ 1 1~n ~ni .
R0

(10.8)

Here xi is the original sample location and x is the new sample location (for which
we want to compute the change), R0 is the harmonic distance to the intersected
surfaces, ~ni is the sample normal, and ~n is the new normal.
Given this estimate of the local variation in irradiance Ward et al. developed a
caching method where previously stored samples re-used whenever possible. All
samples are stored in an octree-tree this structure makes it possible to quickly
locate previous samples. When a new sample is requested the octree is queried first
for previous samples near the new location. For these nearby samples the change
in irradiance, , is computed. If samples with a sufficiently low  is found then
these samples are blended using weights inversely proportional to :
P
wi (x, ~n)Ei (xi )
E(x, ~n)

i,wi >1/a

wi (x, ~n)

i,wi >1/a

Here wi =

1
i ,

a is the desired accuracy and Ei is the irradiance of sample i.

154

(10.9)

Figure 10.3: The box scene rendered using irradiance caching.

If no previously computed sample has a sufficiently high weight then a new

sample is computed.
To further improve this estimate Ward and Heckbert added estimates of the
gradients of the irradiance [94]. Their approach looks not only at the distances
to the nearest surfaces, but also in the relative change of the incoming radiance
from different directions in order to more accurately estimate the gradient due to
a change in position as well as orientation. The great thing about this approach is
that it does not require any further samples, but simply uses a more sophisticated
analysis of the samples in the irradiance estimate.
With a few minor modifications this interpolation scheme works quite well as
can be seen in Figure 10.3. The areas where it fails are in the case where the
overall smoothness assumption for the irradiance field is no longer true. This is
particularly the case for caustics (e.g. the light focused through the glass sphere;
the caustic on the wall is missed completely).
Another issue with irradiance caching is that the method can be quite costly
in scenes where multiple diffuse light reflections are important, since the costly
irradiance estimates are computed recursively for each sample ray. The recursive
evaluation also results in a global error (a global bias). Each light bounce has some
approximation and multiple light bounces distributes this error all over the scene.
This global aspect can make the error hard to control.
155

10.5 Photon Mapping

Photon mapping [32] is a method that uses biasing to reduce variance in many
places. This is in part done by aggressively storing and re-using information whenever possible. This section contains a short overview of the photon mapping technique (for all the details consult [33]).
From a light transport point of view photon mapping exploits that:
The irradiance field is mostly smooth, but has important focusing effects
such as caustics
There are two important sources of light paths in a scene: the light sources
and the observer
In addition photon mapping uses a point sampling data-structure that is independent of the scene geometry, and it therefore works with complex geometry as well
as non-Lambertian reflection models.
Photon mapping is a two-pass method in which the first pass is building the
photon map. This is done using photon tracing in which photons are emitted from
the light sources and traced through the scene. When a photon intersects a diffuse
surface it is stored in the photon map. The second pass is rendering in which
the photon map is a static structure with information about the illumination in the
model. The renderer computes various statistics from the photon map to make the
rendering faster.

10.5.1

Pass 1: Photon Tracing

The first step is building the photon map. This is done by emitting photons from the
light sources and tracing them through the scene using photon tracing. The photons
are emitted according to the power distribution of the light source. As an example,
a diffuse point light emits photons with equal probability in all directions. When
a photon intersects a surface it is stored in the photon map (if the surface material
has a diffuse component). In addition the photon is either scattered or absorbed
based on the albedo of the material. For this purpose Russian roulette [4] is used to
decide if a photon path is terminated (i.e. the photon is absorbed), or if the photon
is scattered (reflected or transmitted). This is shown in Figure 10.4.
The photon tracing algorithm is unbiased. No approximations or sources of
systematic error is introduced in the photon tracing step. In contrast many other
156

Figure 10.4: The photon map is build using photon tracing. From [33].

light propagation algorithms, such as progressive refinement radiosity [11], introduces a systematic error at each light bounce (due to the approximate representation of the illumination).
For efficiency reasons several photon maps are constructed in the first pass. A
caustics photon map that stores only the photons that correspond to a caustic light
path, a global photon map that stores all photon hits at diffuse surfaces (including
the caustics), and a volume photon map that stores multiple scattered photons in
participating media. In the following we will ignore the case of participating media
(see [35, 33] for details).

10.5.2

The Radiance Estimate

The photon map represents incoming flux in the scene. For rendering purposes we
want radiance. Using the expression for reflected radiance we find that:
Z
Lr (x,
~) =

fr (x,
~ 0,
~ )Li (x,
~ 0 )(~nx
~ 0 ) d~
0 ,

(10.10)

where Lr is the reflected radiance at x in direction

~ . x is the hemisphere of
incoming directions, fr is the BRDF and Li is the incoming radiance. To use the
157

Figure 10.5: The radiance estimate is computed from the nearest photons in the
photon map. From [33].

information in the photon map we can rewrite this integral as follows:

Z
d2 i (x,
~ 0)
Lr (x,
~) =
fr (x,
~ 0,
~)
(~nx
~ 0 ) d~
i0
(~nx
~ 0 ) d~
i0 dAi
x

Z
=

fr (x,
~ 0,
~)

d2 i (x,
~ 0)
.
dAi

(10.11)

Here we have used the relationship between radiance and flux to rewrite the incoming radiance as incoming flux instead. By using the nearest n photons around
x from the photon map to estimate the incoming flux, we get:
Lr (x,
~)

n
X

fr (x,
~ p,
~)

p=1

p (x,
~ p)
.
A

(10.12)

This procedure can be seen as expanding a sphere around x until it contains enough
photons. The last unknown piece of information is A. This is the area covered
by the photons, and it is used to computed the photon density (the flux density). A
simple approximation for A is the projected area of the sphere used to locate the
photons. The radius of this sphere is r (where r is the distance to the nth nearest
photon), and we get A = r2 . This is equivalent to a nearest neighbor density
estimate [73]. The radiance estimate is illustrated in Figure 10.5.
The radiance estimate is biased. There are two approximations in the estimate.
It assumes that the nearest photons represents the illumination at x, and it uses
a nearest neighbor density estimate. Both of these approximations can introduce
artifacts in the rendered image. In particular the nearest neighbor density estimate
is often somewhat blurry.
158

However, the radiance estimate is also consistent. As more photons are used in
the photon map as well as the estimate it will converge to the correct value.
A useful property of the radiance estimate is that the bias is purely local (the
error is a function of the local geometry and the local photon density).

10.5.3

Pass 2: Rendering

For rendering the radiance through each pixel is computed by averaging the result
of several sample rays. The radiance for each ray is computed using distribution
ray tracing. This ray tracer is using the photon map both to guide the sampling
(importance sampling) as well as limit the recursion.
There are several strategies by which the photon map can be used for rendering.
One can visualize the radiance estimate directly at every diffuse surface intersected
by a ray. This approach will work (a very similar strategy is used by [69]), but it
requires a large number of photons in both the photon map as well as the radiance
estimate. To reduce the number of photons the two-pass photon mapping approach
uses a mix of several techniques to compute the various components of the reflected
radiance at a given surface location.
We distinguish between specular and diffuse reflection. Here specular means
perfect specular or highly glossy, and diffuse reflection is the remaining part of the
reflection (not only Lambertian).
For all specular surface components the two-pass method uses recursive ray
tracing to evaluate the incoming radiance from the reflected direction. Ray tracing
is pretty efficient at handling specular and highly glossy reflections.
Two different techniques are used for the diffuse surface component. The first
diffuse surface seen either directly through a pixel or via a few specular reflections
is evaluated accurately using Monte Carlo ray tracing. The direct illumination
is computed using standard ray tracing techniques and similarly the irradiance is
evaluated using Monte Carlo ray tracing or gathering (this sampling is improved
by using the information in the photon map to importance sample in the directions
where the local photons originated). Whenever a sample ray from the gathering
step reaches another diffuse surface the estimate from the global photon map is
used). The use of a gathering step means that the radiance estimate from the global
photon map can be fairly coarse without affecting the quality of the final rendered
image. The final component of the two-pass method is caustics, to reduce noise
in the gathering step the caustics component is extracted and caustics are instead
159

Figure 10.6: The rendering step uses a gathering step to compute the first diffuse
bounce more accurately. From [33].

rendered by directly visualizing a caustics photon map naturally this photon

map should have a higher density than the global photon map used in the gathering
step. Fortunately, most noticeable caustics are often caused by focusing of light,
so this happens automatically. The gathering approach is illustrated in Figure 10.6
To make the gathering step faster it pays to use the irradiance caching method
for Lambertian surfaces. Photon mapping works very well with irradiance caching
since photon mapping handles the caustics which irradiance caching cannot handle very well. In addition photon mapping does not need the expensive recursive
evaluation of irradiance values.
Figure 10.7 shows the rendering of the box scene using photon mapping. Here
the global photon map has 200,000 photons and the caustics photon map has 50,000
photons. Notice the smooth overall appearance as well as the caustic on the right
wall due to light reflected of the mirror sphere and focused through the glass sphere.
The bias in the photon mapping method is difficult to quantify. It is a mix of
a local bias (the radiance estimate) and global bias (irradiance caching), and it is
a function of the number of photons in the photon map, the number of photons in
the radiance estimate and the number of sample rays. In general the bias from the
photon map tends to appear as blurry illumination. Features gets washed out if too
few photons are used. However, since the photon map is not visualized directly in
the two-pass method it is somewhat harder to predict what the potential errors of
using too few photons may be.

160

Figure 10.7: The box scene rendered using photon mapping.

161

162

Bibliography
[1] James Arvo. Analytic Methods for Simulated Light Transport. PhD thesis,
Yale University, December 1995.
[2] James Arvo. Applications of irradiance tensors to the simulation of nonLambertian phenomena. In Computer Graphics Proceedings, Annual Conference Series, ACM SIGGRAPH, pages 335342, August 1995.
[3] James Arvo. Stratified sampling of spherical triangles. In Computer Graphics
Proceedings, Annual Conference Series, ACM SIGGRAPH, pages 437438,
August 1995.
[4] James Arvo and David B. Kirk. Particle transport and image synthesis. In
Forest Baskett, editor, Computer Graphics (SIGGRAPH 90 Proceedings),
volume 24, pages 6366, August 1990.
[5] J. Beck and W. W. L. Chen. Irregularities of Distribution. Cambridge University Press, New York, 1987.
[6] Marcel Berger. Geometry, volume II. Springer-Verlag, New York, 1987.
Translated by M. Cole and S. Levy.
[7] Kenneth Chiu, Peter Shirley, and Changyaw Wang. Multi-jittered sampling.
In Paul Heckbert, editor, Graphics Gems IV, pages 370374. Academic Press,
Boston, 1994.
[8] Kenneth Chiu, Peter Shirley, and Changyaw Wang. Multi-jittered sampling.
In Paul Heckbert, editor, Graphics Gems IV, pages 370374. Academic Press,
Boston, 1994.
[9] K.-L. Chung. An estimate concerning the Kolmogoroff limit distribution.
Transactions of the American Mathematical Society, 67:3650, 1949.
163

[10] William G. Cochran. Sampling Techniques (3rd Ed). John Wiley & Sons,
1977.
[11] Michael F. Cohen, Shenchang Eric Chen, John R. Wallace, and Donald P.
Greenberg. A progressive refinement approach to fast radiosity image generation. In John Dill, editor, Computer Graphics (SIGGRAPH 88 Proceedings),
volume 22, pages 7584, August 1988.
[12] Michael F. Cohen and John R. Wallace. Radiosity and Realistic Image Synthesis. Academic Press Professional, San Diego, CA, 1993. Excellent book
on radiosity algorithms.
[13] Robert L. Cook. Stochastic sampling in computer graphics. ACM Transactions on Graphics, 5(1):5172, January 1986.
[14] Robert L. Cook, Thomas Porter, and Loren Carpenter. Distributed ray tracing. Computer Graphics, 18(4):165174, July 1984. ACM Siggraph 84
Conference Proceedings.
[15] R. Cranley and T.N.L. Patterson. Randomization of number theoretic methods for multiple integration. SIAM Journal of Numerical Analysis, 13:904
914, 1976.
[16] P. J. Davis and P. Rabinowitz. Methods of Numerical Integration (2nd Ed.).
Academic Press, San Diego, 1984.
[17] Luc Devroye. Non-uniform Random Variate Generation. Springer, 1986.
[18] Kai-Tai Fang and Yuan Wang. Number Theoretic Methods in Statistics. Chapman and Hall, London, 1994.
[19] Henri Faure. Discrepance de suites associees a` un syst`eme de numeration (en
dimension s). Acta Arithmetica, 41:337351, 1982.
[20] Henri Faure. Good permutations for extreme discrepancy. Journal of Number
Theory, 42:4756, 1992.
[21] G. Fishman. Monte Carlo: Concepts, Algorithms, and Applications.
Springer-Verlag, 1995.
[22] George S. Fishman. Monte Carlo: concepts, algorithms, and applications.
Springer Verlag, New York, NY, 1996.
164

[23] Rafael C. Gonzalez and Paul Wintz. Digital Image Processing (2nd Ed.).
Addison-Wesley, Reading, MA, 1987.
[24] S. Haber. A modified Monte Carlo quadrature. Mathematics of Computation,
20:361368, 1966.
[25] J.H. Halton. On the efficiency of certain quasi-random sequences of points
in evaluating multi-dimensional integrals. Numerische Mathematik, 2:8490,
1960.
[26] John H. Halton. A retrospective and prospective of the Monte Carlo method.
SIAM Review, 12(1):163, January 1970.
[27] J. M. Hammersley and D. C. Handscomb. Monte Carlo Methods. Wiley, New
York, N.Y., 1964.
[28] F. J. Hickernell, H. S. Hong, P. LEcuyer, and C. Lemieux. Extensible lattice sequences for quasi-Monte Carlo quadrature. SIAM Journal on Scientific
Computing, 22(3):11171138, 2000.
[29] E. Hlawka. Funktionen von beschrankter Variation in der Theorie der Gleichverteilung. Annali di Matematica Pura ed Applicata, 54:325333, 1961.
[30] H. S. Hong and F. J. Hickernell. Implementing scrambled digital sequences.
AMS Transactions on Mathematical Software, 2003. To appear.
[31] L.K. Hua and Y. Wang. Applications of number theory to numerical analysis.
Springer, Berlin, 1981.
[32] Henrik Wann Jensen. Global illumination using photon maps. In Xavier
Pueyo and Peter Schroder, editors, Eurographics Rendering Workshop 1996,
pages 2130, New York City, NY, June 1996. Eurographics, Springer Wien.
ISBN 3-211-82883-4.
[33] Henrik Wann Jensen. Realistic Image Synthesis using Photon Mapping. AK
Peters, 2001.
[34] Henrik Wann Jensen and Niels J. Christensen. Optimizing path tracing using
noise reduction filters. In Winter School of Computer Graphics 1995, February 1995. held at University of West Bohemia, Plzen, Czech Republic, 14-18
February 1995.
165

[35] Henrik Wann Jensen and Per H. Christensen. Efficient simulation of light
transport in scenes with participating media using photon maps. In Michael
Cohen, editor, SIGGRAPH 98 Conference Proceedings, Annual Conference
Series, pages 311320. ACM SIGGRAPH, Addison Wesley, July 1998. ISBN
0-89791-999-8.
[36] James T. Kajiya. The rendering equation. Computer Graphics, 20(4):143
150, August 1986. ACM Siggraph 86 Conference Proceedings.
[37] M. H. Kalos and Paula A. Whitlock. Monte Carlo Methods, volume I, Basics.
John Wiley & Sons, New York, 1986.
[38] Malvin H. Kalos and Paula A. Whitlock. Monte Carlo Methods: Volume I:
Basics. John Wiley & Sons, New York, 1986.
[39] Alexander Keller. A quasi-Monte Carlo algorithm for the global illumination
problem in a radiosity setting. In Harald Niederreiter and Peter Jau-Shyong
Shiue, editors, Monte Carlo and Quasi-Monte Carlo Methods in Scientific
Computing, pages 239251, New York, 1995. Springer-Verlag.
[40] David Kirk and James Arvo. Unbiased sampling techniques for image synthesis. Computer Graphics, 25(4):153156, July 1991.
[41] David Kirk and James Arvo. Unbiased variance reduction for global illumination. In Proceedings of the Second Eurographics Workshop on Rendering,
Barcelona, May 1991.
[42] David B. Kirk and James Arvo. Unbiased sampling techniques for image
synthesis. In Thomas W. Sederberg, editor, Computer Graphics (SIGGRAPH
91 Proceedings), volume 25, pages 153156, July 1991.
[43] N. M. Korobov. The approximate computation of multiple integrals. Dokl.
Akad. Nauk SSSR, 124:12071210, 1959.
[44] L. Kuipers and H. Niederreiter. Uniform Distribution of Sequences. John
Wiley and Son, New York, 1976.
[45] Brigitta Lange. The simulation of radiant light transfer with stochastic raytracing. In Proceedings of the Second Eurographics Workshop on Rendering
(Barcelona, May 1991), 1991.
166

[46] P. LEcuyer and C. Lemieux. A survey of randomized quasi-Monte Carlo

methods. In M. Dror, P. LEcuyer, and F. Szidarovszki, editors, Modeling
Uncertainty: An Examination of Stochastic Theory, Methods, and Applications, pages 419474. Kluwer Academic Publishers, 2002.
[47] Mark E. Lee and Richard A. Redner. A note on the use of nonlinear filtering
in computer graphics. IEEE Computer Graphics and Applications, 10(3):23
29, May 1990.
[48] Mark E. Lee, Richard A. Redner, and Samuel P. Uselton. Statistically optimized sampling for distributed ray tracing. In B. A. Barsky, editor, Computer
Graphics (SIGGRAPH 85 Proceedings), volume 19, pages 6167, July 1985.
[49] J. Matousek. On the L2 discrepancy for anchored boxes. Journal of Complexity, 14:527556, 1998.
[50] Michael McCool. Anisotropic diffusion for monte carlo noise reduction.
ACM Transactions on Graphics, pages 171194, April 1999.
[51] M. D. McKay, R. J. Beckman, and W. J. Conover. A comparison of three
methods for selecting values of input variables in the analysis of output from
a computer code. Technometrics, 21(2):23945, 1979.
[52] Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth, Augusta H. Teller, and Edward Teller. Equation of state calculations by fast
computing machines. Journal of Chemical Physics, 21(6):10871092, 1953.
[53] Don P. Mitchell. Spectrally optimal sampling for distributed ray tracing. In
Thomas W. Sederberg, editor, Computer Graphics (SIGGRAPH 91 Proceedings), volume 25, pages 157164, July 1991.
[54] H. Niederreiter and C. Xing. Low-discrepancy sequences and global function
fields with many rational places. Finite Fields and Their Applications, 2:241
273, 1996.
[55] Harald Niederreiter. Point sets and sequences with small discrepancy. Monatshefte fur mathematik, 104:273337, 1987.
[56] Harald Niederreiter. Random Number Generation and Quasi-Monte Carlo
Methods. S.I.A.M., Philadelphia, PA, 1992.
167

[57] A. B. Owen. Randomly permuted (t, m, s)-nets and (t, s)-sequences. In

Harald Niederreiter and Peter Jau-Shyong Shiue, editors, Monte Carlo and
Quasi-Monte Carlo Methods in Scientific Computing, pages 299317, New
York, 1995. Springer-Verlag.
[58] A. B. Owen. Monte Carlo variance of scrambled equidistribution quadrature.
SIAM Journal of Numerical Analysis, 34(5):18841910, 1997.
[59] A. B. Owen. Latin supercube sampling for very high dimensional simulations. ACM Transactions on Modeling and Computer Simulation, 8(2):71
102, 1998.
[60] Art B. Owen. Scrambling Sobol and Niederreiter-Xing points. Journal of
Complexity, 14(4):466489, December 1998.
[61] Art B. Owen. Monte Carlo quasi-Monte Carlo and randomized quasi-Monte
Carlo. In H. Niederreiter and J. Spanier, editors, Monte Carlo and quasiMonte Carlo Methods 1998, pages 8697, 1999.
[62] Mark Pauly, Tomas Kollig, and Alexander Keller. Metropolis light transport
for participating media. In Rendering Techniques 2000 (Proceedings of the
Eurographics Rendering Workshop), pages 1122, New York, June 2000. Eurographics, Springer Wien.
[63] Werner Purgathofer. A statistical method for adaptive stochastic sampling.
Computers and Graphics, 11(2):157162, 1987.
[64] R. Y. Rubinstein. Simulation and the Monte Carlo Method. John Wiley &
Sons, New York, 1981.
[65] Holly E. Rushmeier and Gregory J. Ward. Energy preserving nonlinear filters. In Andrew Glassner, editor, Proceedings of SIGGRAPH 94 (Orlando,
Florida, July 2429, 1994), Computer Graphics Proceedings, Annual Conference Series, pages 131138. ACM SIGGRAPH, ACM Press, July 1994.
ISBN 0-89791-667-0.
[66] Ch. Schlier. A practitioners view on qmc integration. Mathematics and
Computers in Simulation, 2002.
168

[67] P. Shirley. Discrepancy as a quality measure for sample distributions.

In Werner Purgathofer, editor, Eurographics 91, pages 183194. NorthHolland, September 1991.
[68] Peter Shirley. A ray tracing method for illumination calculation in diffusespecular scenes. In Proceedings of Graphics Interface 90, pages 205212,
May 1990.
[69] Peter Shirley, Bretton Wade, Philip Hubbard, David Zareski, Bruce Walter,
and Donald P. Greenberg. Global illumination via density estimation. In
Eurographics Rendering Workshop 1995. Eurographics, June 1995.
[70] Peter Shirley, Chang Yaw Wang, and Kurt Zimmerman. Monte carlo techniques for direct lighting calculations. ACM Transactions on Graphics,
15(1):136, January 1996. ISSN 0730-0301.
[71] Peter Shirley, Changyaw Wang, and Kurt Zimmerman. Monte Carlo methods
for direct lighting calculations. ACM Transactions on Graphics, 15(1):136,
January 1996.
[72] Y. A. Shreider. The Monte Carlo Method. Pergamon Press, New York, N.Y.,
1966.
[73] B. W. Silverman. Density Estimation for Statistics and Data Analysis. Chapman and Hall, London, 1986.
[74] Ian H. Sloan and S. Joe. Lattice Methods for Multiple Integration. Oxford
Science Publications, Oxford, 1994.
[75] I. M. Sobol. The distribution of points in a cube and the accurate evaluation
of integrals (in Russian). Zh. Vychisl. Mat. i Mat. Phys., 7:784802, 1967.
[76] J. Spanier. Quasi-Monte Carlo Methods for Particle Transport Problems. In
Harald Niederreiter and Peter Jau-Shyong Shiue, editors, Monte Carlo and
Quasi-Monte Carlo Methods in Scientific Computing, pages 121148, New
York, 1995. Springer-Verlag.
[77] Jerome Spanier and Ely M. Gelbard. Monte Carlo Principles and Neutron
Transport Problems. Addison-Wesley, Reading, Massachusetts, 1969.
169

[78] Michael Stein. Large sample properties of simulations using Latin hypercube
sampling. Technometrics, 29(2):14351, 1987.
[79] Frank Suykens and Yves Willems. Adaptive filtering for progressive monte
carlo image rendering. 2000.
[80] Rasmus Tamstorf and Henrik Wann Jensen. Adaptive sampling and bias estimation in path @acing. In Julie Dorsey and Philipp Slusallek, editors, Eurographics Rendering Workshop 1997, pages 285296, New York City, NY,
June 1997. Eurographics, Springer Wien. ISBN 3-211-83001-4.
[81] Boxin Tang. Orthogonal array-based Latin hypercubes. Journal of the American Statistical Association, 88:13921397, 1993.
[82] D. M. Titterington, A. F. M. Smith, and U. E. Makov. The Statistical Analysis
of Finite Mixture Distributions. John Wiley & Sons, New York, NY, 1985.
[83] Bruno Tuffin. On the use of low discrepancy sequences in Monte Carlo methods. Technical Report 1060, I.R.I.S.A., Rennes, France, 1996.
[84] Greg Turk. Generating random points in triangles. In Andrew S. Glassner,
editor, Graphics Gems, pages 2428. Academic Press, New York, 1990.
[85] J. G. van der Corput. Verteilungsfunktionen I. Nederl. Akad. Wetensch. Proc.,
38:813821, 1935.
[86] J. G. van der Corput. Verteilungsfunktionen II. Nederl. Akad. Wetensch.
Proc., 38:10581066, 1935.
[87] Eric Veach. Robust Monte Carlo Methods for Light Transport Simulation.
PhD thesis, Stanford University, December 1997.
[88] Eric Veach and Leonidas Guibas. Bidirectional estimators for light transport.
In 5th Annual Eurographics Workshop on Rendering, pages 147162, June
1315 1994.
[89] Eric Veach and Leonidas Guibas. Optimally combining sampling techniques
for Monte Carlo rendering. In SIGGRAPH 95 Conference Proceedings,
pages 419428. Addison-Wesley, August 1995.
170

[90] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Computer Graphics Proceedings, Annual Conference Series, ACM SIGGRAPH, pages 419428, August 1995.
[91] Eric Veach and Leonidas J. Guibas. Metropolis light transport. In Turner
Whitted, editor, SIGGRAPH 97 Conference Proceedings, Annual Conference Series, pages 6576. ACM SIGGRAPH, Addison Wesley, August 1997.
ISBN 0-89791-896-7.
[92] Changyaw Wang. Physically correct direct lighting for distribution ray tracing. In David Kirk, editor, Graphics Gems 3. Academic Press, New York,
NY, 1992.
[93] Greg Ward. Adaptive shadow testing for ray tracing. In Proceedings of the
Second Eurographics Workshop on Rendering (Barcelona, May 1991), 1991.
[94] Gregory J. Ward and Paul Heckbert. Irradiance gradients. Third Eurographics
Workshop on Rendering, pages 8598, May 1992.
[95] Gregory J. Ward, Francis M. Rubinstein, and Robert D. Clear. A ray tracing
solution for diffuse interreflection. In John Dill, editor, Computer Graphics
(SIGGRAPH 88 Proceedings), volume 22, pages 8592, August 1988.

[96] H. Weyl. Uber

die gleichverteilung von zahlen mod. eins. Mathematische
Annalen, 77:313352, 1916.
[97] H. Wozniakowski. Average case complexity of multivariate integration. Bulletin (New Series) of the American Mathematical Society, 24(1):185193,
January 1991.
[98] Sidney J. Yakowitz. Computational Probability and Simulation. AddisonWesley, New York, N.Y., 1977.
[99] S. K. Zaremba. The mathematical basis of Monte Carlo and quasi-Monte
Carlo methods. SIAM Review, 10(3):303314, July 1968.

171