ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.

1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

Kernel density estimation and its application

Stanisław Węglarczyk1,*
1
Cracow University of Technology, Institute of Water Management and Water Engineering, Warszawska 24, 31-115 Kraków,
Poland

Abstract. Kernel density estimation is a technique for estimation of probability density function
that is a must-have enabling the user to better analyse the studied probability distribution than when
using a traditional histogram. Unlike the histogram, the kernel technique produces smooth estimate
of the pdf, uses all sample points' locations and more convincingly suggest multimodality. In its
two-dimensional applications, kernel estimation is even better as the 2D histogram requires
additionally to define the orientation of 2D bins. Two concepts play fundamental role in kernel
estimation: kernel function shape and coefficient of smoothness, of which the latter is crucial to the
method. Several real-life examples, both for univariate and bivariate applications, are shown.

1 Introduction Two concepts play fundamental role in kernel

estimation: the kernel function and the coefficient of
Out of all probability distribution functions, probability smoothness.
density function (pdf) best shows how the whole 100%
probability mass is distributed over the x-axis, i.e., over
the values of an X random variable. However, the oldest 2 Kernel density
pdf empirical representation  a histogram  is a highly
subjective structure as its shape depends on the Let the series {x1, x2,..., xn} be an independent and
subjective choice of the number (or widths) of class identically distributed (iid) sample of n observations
intervals (bins) to which the range of a sample is taken from a population X with an unknown probability
divided, and on the choice of the initial point (e.g., [1]). distribution function f(x). Kernel estimate fˆ ( x) of
To this aim several formulas have been proposed of original f(x) assigns each i-th sample data point xi
which most relate the number of intervals to the sample a function K(xi,t) called a kernel function in the
size only [2–3]; the other include additionally certain following way [11]:
sample characteristics as standard deviation [4],
interquartile range [5] or skewness [6]. 1 n
fˆ (t )   K ( xi , t ) (1)
Independently of the class selection method used, the n i1
histogram suffers from its original sin: data binning,
which depraves the data of their individual location K(x,t) is nonnegative and bounded for all x and t:
replacing their locations with a bin (interval) location.
This causes the histogram shape to become 0  K ( x, t )   for all real x, t (2)
discontinuous, and flat in each bin.
Kernel estimation of probability density function has and, for all real x,
not these drawbacks. It produces (in in most practical 
applications) a smooth empirical pdf based on individual
locations of all sample data. Such pdf estimate seems to
 K ( x, t )dt  1.

(3)

better represent the "true" pdf of a continuous variable.

Kernel estimation is not a quite new technique: it was Property (3) ensures the required normalization of
originated more than a half century ago by Rosenblatt kernel density estimate (1):
[7] and Parzen [8]. With the development of computer  
1 n
technology, this method has been developing rapidly and  fˆ (t )dt    K ( xi , t )dt  1. (4)
vastly [4, 9–17].  n i1 
The paper shows the advantages and disadvantages
of the method illustrating them with real-life examples In other words, kernel transforms the "sharp" (point)
for one- and two-dimensional applications. location of xi into an interval centred (symmetrically or
not) around xi.

*
Corresponding author: [email protected]
© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(http://creativecommons.org/licenses/by/4.0/).
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

In most common practical applications, the kernel

estimation uses symmetric kernel function, although
asymmetric functions have recently been increasingly
used [18–20]. Figs. 1 and 2 illustrate the idea of kernel
estimation for both cases.

Fig. 3. The value of the smoothing parameter h influences the

shape of the resulting kernel density. The 4-element sample
(vertical segments) are the same as in Figs. 1 and 2.

Many types of kernel function can be found in the

relevant literature. Examples of symmetric kernels are
presented in Table 1 and in Fig. 4, while Table 2 shows
Fig. 1. Construction of kernel density estimator (1) (continuous the asymmetric ones.
line) with a symmetric kernel (dashed lines) for a 4-element
sample (vertical segments). Table 1. Examples of symmetrical kernel functions [11].

Kernel Definition

 3 (1  15 t ) 2 for t  5
K (t )  
4 5
Epanechnikov
 0 for t  5

 1615 (1  t ) for t  1
2 2

Biweight K (t )  
0 for t  1


1  t for t  1
Triangular K (t )  

0 for t  1
1 t 2 / 2
Gaussian K (t )  e
2
Fig. 2. Construction of kernel density estimator (1) (continuous
 for t  1
1
line) with an asymmetric kernel (dashed lines) for the same 4-
Rectangular K (t )   2
element sample as in Fig. 1. 0 for t  1
Fig. 1 shows that the shape of a symmetric kernel is Table 2. Examples of asymmetrical kernel functions.
the same for all sample points while Fig. 2 reveals that Symbol b denotes the smoothing parameter.
the shape of an asymmetric kernel differs with the point
placement.
Symmetry property allows to write the kernel Kernel Definition
function in a form used most frequently: x / b t / b
t e
Gamma 1 [18] K GAM 1 ( x, b; t ) 
1  x t  b
x / b 1
( x / b  1)
K sym ( x, t )  K  (5)
h  h  t b ( x ) 1et / b
KGAM 2 ( b ( x), b; t )  b ( x )
b ( b ( x))
where parameter h, called smoothing parameter, window Gamma 2 [18]
x / b for x  2b
width or bandwidth, governs the amount of smoothing b ( x )   1
 4 ( x / b)  1 for x  [0,2b)
2
applied to the sample (Fig. 3).
For symmetrical kernel functions, the choice of the 
1 t x
Inverse 1  2 
2 bx  x t
shape of the kernel function K(.) has rather little effect K IG ( x, b; t )  e
Gaussian [19] 2 bt
on the shape of the estimator [11, 21], whereas  as Fig.
3

3 shows  the influence of the smoothing parameter h is Reciprocal

1    2
x b  t x b 

critical because it determines the amount of smoothing. Inverse K RIG ( x, b, t )  e 2b  x  b t 

Gaussian [19] 2 bt
Too small value of h may cause the estimator to show
 ln t  ln x 2
insignificant details while too large value of h causes Lognormal 1 
8 ln(1 b )
K LN ( x, b; t )  e
oversmoothing of the information contained in the [20] 8 ln(1  b)t
sample, which, in consequence, may mask some of
important characteristics, e.g. multimodality, of f(x) (cf.
Fig. 3). A certain compromise is needed.

2
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

Two versions of (6) are used in practice: the product

kernel estimator and the radial kernel estimator [24].
In its most popular form, the product kernel estimator
may be written as follows

1 n
 xi  x   y j  y 
fˆ ( x, y )  K   K   (7)
nhx hy i 1  hx   hy 

The radial kernel estimator is based on the Euclidean

Fig. 4. Shapes of symmetric kernels defined in Table 1. distance between an arbitrary point {x,y} and sample
point {xi,yi}, i = 1,2, ..., n:
Fig. 5 illustrates how the kernel type (cf. Table 1)
used to estimate pdf influences the kernel pdf estimate.
n   x  x 2  y  y 2 
Triangular and rectangular kernels (especially the latter) 1
produce many local maxima and thus they are rather not
fˆ ( x, y )   K   i
   h 
i  (8)

  x   y 
nhx hy i 1 h
recommended for application. The biweight kernel has 
shorter support than the Epanechnikov one, so reveals
In practice, the product kernel estimator is mostly
more details and more clearly suggests two basic modes.
used.
The Gaussian kernel, distributed over the whole x-axis,
produces the most smooth estimate, and this property The advantage of multivariate kernel pdf over
probably causes the kernel to be most frequently used. multivariate histogram is even greater than in an
univariate case. This is because of an additional
subjective requirement occurs: the user has to decide
about the orientation of a two-dimensional bin, which
may considerably influence the final shape of the
histogram.

3 Measures of discrepancy between the

kernel density estimator fˆ and the true
density f
Each estimator fˆ ( x) differs from its original f(x) with
100% probability. In order to build a method producing
an estimator fˆ ( x) which will be as close to f(x) as
possible, certain measures should be defined to evaluate
this discrepancy.
For each single x, a difference between the "true"
density function f(x) and its estimator fˆ ( x) can be
estimated with the mean squared error, MSEx, [11]:

   
Fig. 5. Different symmetrical kernel functions applied to
MSE x fˆ  E  fˆ  x   f  x  
2
a sample of 45 standardized annual maximum (9)
 
flows (1961–1995) of Odra river recorded at the
Racibórz-Miedonia gauge station (data source: [22]).
which, after simple transformations, can be presented as
The univariate case can be easily formally extended follows:
to the multivariate case [23]. However, its illustrative
   
MSE x fˆ  Efˆ  x   f  x   var fˆ  x 
2
(graphical) power works well for bivariate case only.
The most frequently used bivariate kernel function is (10)
  bias fˆ  x    var fˆ  x 
2
symmetric
 
1 n  xi  x y j  y 
fˆ ( x, y)   K  ,
hy 
 (6) that is, MSEx is the sum of the square bias and the
nhx hy i 1  hx variance of fˆ ( x) at x. Reducing the bias causes variance
to increase and vice versa, so a trade-off between these
where {xi, yi}, i = 1,2,...,n, is a sample, and hx and hy are terms is needed.
smoothing coefficients. Available are multivariate MSEx is a local measure. Integration of MSEx over
counterparts of univariate kernel functions listed in
Table 1, e.g., multivariate Epanechnikov kernel or all x gives a global measure of conformity of fˆ ( x) with
multivariate Gaussian kernel [11]. f(x), called the mean integrated square error, MISE, [11]:

3
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

 
 The value (15) is widely used in practice and referred to
MISE(fˆ )   MSE x fˆ dx
- as the Silverman’s bandwidth or (Silverman’s) rule of
 
(11)
thumb, and will be used in most of the remainder of the
   bias fˆ  x   dx   var fˆ  x  dx
2

-   - paper.

MISE is one of measures used to estimate the smoothing

parameter. 4.2 Least squares cross validation method
In practice, an approximate version of MISE, called (LSCV)
AMISE (asymptotic MISE) is also used, developed by The least squares cross validation method (LSCV) of
expanding MISE into a Taylor series and taking only the selecting the smoothing parameter is a very popular
most important parts [25, 26]. technique [11, 30, 33–38].
Integrated square error, ISE, is an intermediate LSCV uses the integrated square error, ISE (12),
measure, between MISE and MSE: which can be expressed in the following form 11:


  fˆ  x   f  x  dx
2 

  fˆ  x   f  x  
ISE(fˆ )  (12) ISE (h) 
2
dx


 
which is also a discrepancy measure used to estimate the   fˆ 2  x dx  2  fˆ  x   f  x  dx (16)
magnitude of the smoothing parameter.  

  f 2  x dx
4 Methods for calculating optimum 
value of smoothing parameter
The last part of the expression (16) does not depend
The choice of the optimal smoothing parameter is based,
on the estimator fˆ ( x) (it is a constant), therefore the
i.a., on formulas that minimize the criterion functions
discussed above, mainly ISE [27], MISE [28] and choice of the smoothing parameter (in the sense of
AMISE [11, 15, 29–32]. minimizing ISE) will correspond to the choice of
Many other methods for calculating the smoothing h which minimizes the function
parameter are available in the relevant literature; many  
of them are available also through statistical software.
Two methods are described below  one for the
   fˆ
R fˆ 

2
 x  dx  2  fˆ  x   f  x  dx

(17)

symmetrical kernel function (Gaussian), the other for

any kernel function. To estimate the second part of (17) a leave-one-out
density estimator, fˆi  x  , is used:
4.1 Rule-of-thumb method
 K  x, x j 
1
The rule-of-thumb method is based on the asymptotic fˆi  x   (18)
n  1 j i
mean integrated square error, AMISE, when the kernel
function and true distribution are assumed normal. which is an estimate of the density function calculated
Silverman [11] got then the values of the smoothing using all sample values except xi. The resulting form of
parameter h as follows: the LSCV criterion function is
h  1.06  ˆ  n1/5 (13) 
2
LSCV  h    fˆ 2  x  dx   fˆi  xi  (19)
where ˆ is the sample standard deviation and n is the 
n i
sample size.
In order to have an estimator more robust against The optimal smoothing parameter hLSCV is the value for
outliers the sample interquartile range IRQ may be which the LSCV(h) function achieves the minimum. The
final form of LSCV function (19), applicable to both
applied [11]: symmetrical and asymmetrical kernels, is:

h  0.79  IQR  n1/5 (14) 

  K  x, x  K  x, x dx
1
LSCV  h   i j
n2 i , j 
Silverman [11] believes that the value (13) smoothes (20)
non-unimodal distributions too much, and  as one of the
 K  xi , x j 
2

remedies  proposes a slightly reduced value of the n(n  1) i j  i
smoothing parameter (13):
Least squares cross-validation is also referred to as
 IQR  1/5 unbiased cross-validation 26.
h  0.9  min  ˆ , n
 1.34 
(15) Unfortunately, the LSCV method also has
  drawbacks: the variance of the obtained smoothing

4
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

parameters calculated for samples drawn from the same

distribution is large [30]. It happens that the LSCV(h)
function has several minimums, often false and far on
the side of too small smoothing [39]; sometimes
LSCV(h) does not have any minima at all [14, 30].
There are other versions of the cross-validation
method, e.g. biased cross-validation (BCV) or smoothed
cross-validation (SCV), and other methods to obtain
optimum smoothing coefficient (e.g., [40]). Some
resulting examples are shown in Fig. 6.

Fig. 7. Kernel density estimates for four 45-year time series of

standardized annual maximum flows (1961–1995)
of given River/Gauging station (data source: [22]).

Fig. 8. Kernel density estimates for four 32-year time series of

standardized annual minimum flows (1983–2015)
of given River/Gauging station (data source: [41]).

Fig. 6. Different methods for kernel smoothing coefficient

estimation available in Wolfram Mathematica 11.1 applied to
the 1961–1995 series of standardized annual maximum flows
of Odra river recorded at the Racibórz-Miedonia
gauge station (data source: [22]).

5 Kernel density in practice

5.1 The univariate case

Figs. 7 through 9 contain several kernel pdf estimates
obtained for maximum and minimum annual flows of
certain rivers and maximum annual precipitations in
Poland. Apart from the nice smoothness contrasting with Fig. 9. Kernel density estimates for four 30-year time series of
a histogram shape, a very attractive characteristic of standardized annual maximum precipitation (1984–
kernel estimation is shown: its ability to suggest 2013) at given Precipitation station/River basin/ (data source:
[42]).
multimodality in a more convincing way than the
histogram does. Of course, the multimodal shape of a pdf estimate
does not prove the existence of the real multimodality. It
is, however, a sign of possible non-homogeneity that
should be considered through the analysis of the
mechanism generating the data. Some attempts to
statistical testing multimodality are described in [11];

5
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

however, as Silverman ([11], p. 141) conclude: "It may If the amount of the probability leakage cannot be
be futile to expect very high power from procedures disregarded, one of the remedies is to logarithmize the
aimed at such broad hypotheses as unimodality and data and apply the kernel estimation to such data. If pdf
multimodality". Nevertheless, the kernel estimation is of logarithmized data is gˆ( x) the following recalculation
a good method for an initial stage of the planned study should be used:
on probability distribution.
When the variable under study is nonnegative, it may 1
happen that kernel estimate exhibits an undesirable case: fˆ ( x)  gˆ (ln( x)) (21)
x
probability leakage below zero. It occurs when a part of
the sample lies near zero and the magnitude of the Fig. 11(b) shows the result. The leakage has been
smoothing coefficient enables such crossing in removed; unfortunately, the second mode disappeared
a considerable amount. Four such cases are presented in although certain suggestion of non-unimodality has
Fig. 10. remained visible in the heaviness of the right tail.
Another remedy is to use an asymmetric kernel
shown in Fig. 11(c). This approach shows the bimodality
revealed in Fig. 11(a). In terms of cumulative
distribution function (Fig. 11(d)), log transformation and
asymmetric kernel approach are almost equivalent.

5.2 The bivariate case and some general

remarks on the multivariate case

Formally, the univariate case can be easily extended to

the multivariate one, which has been exemplified by
equations (7) and (8) for the bivariate kernel. Fig. 12
illustrates with the use of equation (7) how the relation
between the two variables studied evolves over the year.
2D kernel pdf graphics may help the user in
differentiating the sample into subsamples, for which
a non-statistical (cause-and-effect) confirmation may be
Fig. 10. Probability leakage below zero (marked dark blue) in
found. Such graphics is informative when a sample
kernel density estimates for time series of standardized annual
maximum flow (1961–1995), top two graphs, and
contains many identical data, which are not visible in an
annual minimum flows (1983–2015), bottom two x-y plot.
graphs, of given River/Gauging station (data source: [22]). The Unfortunately, graphical illustration or interpretation
numbers within the graphs show the magnitude of probability for more than two-variate case is at least difficult if not
leakage. impossible. Moreover, sample size necessary for
preserving similar accuracy as that for one-dimensional
case grows rapidly with growing dimension  the
problem known as the 'curse of dimensionality'.
Minimization of the effect of the curse of
dimensionality requires not only sufficient data, but also
careful data preparation [23]. This may involve proper
transformation of marginal variable in order to reduce
the large skewness or heavy tails, determination if the
data are of full rank, and even  if the data do not have
many significant digits  carefully blurring the data [23].

6 Summary and conclusions

When compared with the commonly used histogram, the
kernel density estimator shows several advantages.
1. It is a smooth curve and thus it better exhibits the
details of the pdf, suggesting in some cases non-
unimodality.
Fig. 11. Removing probability leakage below zero (4.9%, 2. It uses all sample points' locations, so, therefore, it
marked dark blue) in kernel density estimate (a) of the 1984– better reveal the information contained in the sample.
2015 time series of standardized annual maximum flows 3. It more convincingly suggests multimodality.
; (b) logarithmized pdf added; (c) asymmetric 4. The bias of the kernel estimator is of one order
gamma kernel pdf added (cf. Table 2, kernel KGAM1, b = 0.06); better than that of a histogram estimator [26].
(d) three cumulative distribution functions (data source: [41]). 5. Compared with 1D application, 2D kernel
applications are even more better as the 2D histogram

6
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

requires additionally the specification of the orientation References

of the bins which enhances the subjectivity of histogram.
It should be remembered, however, that the value of 1. S. Węglarczyk., M. Kulig, Wiad. IMGW
smoothing coefficient is to some extent a subjective XIV(XLV), Z.2, 59–69 (2001)
estimate. 2. H.A. Sturges, J. Amer. Statist. Assoc. 21, 65–66
(1926)
3. C. E. P. Brooks, N. Carruthers, Handbook of
statistical methods in meteorology (HM Stationary
Office, London, 1953)
4. D.W. Scott, Biometrica 66, 605–610 (1979)
5. D. Freedman, P. Diaconis, Zeit. Wahr. ver. Geb.
57(4), 453–476 (1981)
6. D. P. Doane, American Statistician 30(4), 181–183
(1976)
7. M. Rosenblatt, Annals of Mathematical Statistics 27,
832–837 (1956)
8. E. Parzen, Annals of Mathematical Statistics 33,
1065–1076 (1962)
9. A. Bowman, Journal of Stat. Comp. Simul. 21, 313–
327 (1985)
10. G.R. Terrel, D.W. Scott, Journal of the American
Statistical Association 80(389), 209–214 (1985)
11. B.W. Silverman, Density estimation for statistics and
data analysis (Chapman and Hall, London, 1986)
12. L. Devroye, Annales de l’Institut Henri Poincaré 25,
533–580 (1989)
13. G.R. Terrel, Journal of the American Statistical
Association 85, 470–477 (1990)
14. S.J. Sheather, Computational Statistics 7, 225–250
(1992)
15. J.S. Marron, M. P. Wand, Annals of Statistics 20,
712–736 (1992)
16. L. Devroye, Statistics and Probability Letters 20,
183–188 (1994)
17. L. Devroye, A. Krzyżak, Journal of Multivariate
Analysis 82, 88–110 (2002)
18. S.X. Chen, Annals Of The Institute of Statistical
Mathematics 52(3), 471–480 (2000)
19. O. Scaillet, Density estimation using inverse and
reciprocal inverse gaussian kernels (IRES
Discussion Paper 17, Université Catolique de
Louvain, 2001)
20. X. Jin, J. Kawczak, Annals of Economics and
Finance 4, 103–124 (2003)
21. P. Hall, J.S. Marron, The Annals of Statistics 15(1),
163–181 (1987)
22. B. Fal, E. Bogdanowicz, W. Czernuszenko,
I. Dobrzyńska, A. Koczyńska, Przepływy
charakterystyczne głównych rzek polskich w latach
Fig. 12. Bivariate kernel density estimates for two-dimensional
1951–1995 (in Polish: Characteristics flows of main
random variable (monthly maximum temperature tmx, and rivers in Poland in 1951–1995) (Materiały
monthly sunshine duration, S), in Oxford, UK, 1853–2017 Badawcze, Seria: Hydrologia i Oceanologia 26,
(data source:[43]). Instytut Meteorologii i Gospodarki Wodnej,
Warszawa, 2000)

7
ITM Web of Conferences 23, 00037 (2018) https://doi.org/10.1051/itmconf/20182300037
XLVIII Seminar of Applied Mathematics

23. D. W. Scott, Multivariate Density Estimation,

Theory, Practice, and Visualization (John Wiley and
Sons, Inc., 1992)
24. W. Härdle, M. Müller, S. Sperlich, A. Werwatz,
Nonparametric and Semiparametric Models
(Springer, 2004)
25. T. Ledl, Austrian Journal of Statistics 33(3), 267–279
(2004)
26. S.J. Sheather, Statist. Sci. 19(4), 588–597 (2004)
27. S.R Sain, Adaptive kernel density estimation, (PhD
diss., Rice University, http://hdl.handle.net/1911/
16743, 1994, accessed June 2018)
28. J.S. Marron, D. Nolan, Statistics and Probability
Letters 7, 195–199 (1989)
29. A. Bowman, P. Hall, T. Prvan, Biometrika 85(4),
799–808 (1998)
30. B.A. Turlach, Bandwidth selection in kernel density
estimation: A Review (Discussion Paper, C.O.R.E.
and Institut de Statistique, Université Catolique de
Louvain-la-Neuve, Belgium, 1993)
31. W. Feluch, Wybrane metody jądrowej estymacji
funkcji gęstości prawdopodobieństwa i regresji w
hydrologii (in Polish: Selected methods for kernel
estimation of probability density function and
regression in hydrology) (Prace Naukowe Poli-
techniki Warszawskiej 15, Oficyna Wydawnicza
Politechniki Warszawskiej, Warszawa, 1994)
32. E. Choi, P. Hall, Biometrika 86(4), 941–947 (1999)
33. P. Hall, S.J. Sheather, M.C. Jones, J.S. Marron,
Biometrika 78(2), 263–269 (1991)
34. S.T. Chiu, Statistica Sinica 6, 129–145 (1996)
35. S.T. Chiu, The Annals of Statistics 19(4), 1883–1905
(1991)
36. S.T. Chiu, Biometrika 79(4), 771–782 (1992)
37. M. Rudemo, Scand. Journal of Statistics 9, 65–78
(1982)
38. A. Bowman, Biometrica 71, 353–360 (1984)
39. P. Hall, J.S. Marron, Journal of the Royal Statistical
Society B(53), 245–252 (1991)
40. A. Michalski, Meteorology Hydrology and Water
Management 4(1), 40–46 (2016)
41. Roczniki Hydrologiczne 1984–2015, IMGW-PIB
(Institute of Meteorology and Water Management -
National Research Institute), CD-ROM
42. IMGW-PIB (Institute of Meteorology and Water
Management - National Research Institute)
43. www.metoffice.gov.uk/pub/data/weather/uk/
climate/stationdata/oxforddata.txt (accessed May
2018)