IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 1

Supplementary Material: Probabilistic Principal Curves on Riemannian Manifolds

Seungwoo Kang and Hee-Seok Oh

S-I. M ATHEMATICAL P RELIMINARY analogous to vector addition and subtraction in the Euclidean
space.
A. Geometric Components of Riemannian Manifolds
This section briefly reviews the required background knowl- B. Riemannian Symmetric Space
edge of differential geometry used in this study. For formal
definitions and details, see [1] and [2]. The Riemannian symmetric space refers to a connected
Riemannian manifold is a smooth manifold M equipped Riemannian manifold that satisfies the following properties:
with a smoothly varying inner product ⟨·, ·⟩p on the tangent there exists an isometry σp : M → M for each p ∈ M such
space Tp M for each p ∈ M. This inner product is called that σp is involutive (σp ◦ σp is an identity map) and has p
Riemannian metric, and the subscript p is dropped when no as an isolated fixed point. Recall that the map σp is called
confusion arises. Using the inner product, the normpof the isometry if d(q1 , q2 ) = d(σp (q1 ), σp (q2 )).
tangent vector v ∈ Tp M is defined as ∥v∥ = ⟨v, v⟩. The following two important properties are derived from
Then, for any piecewise differentiable curve, its distance can the definition of Riemannian symmetric space. We state the
be computed by integrating the norm of the tangent vector results without proof.
(velocity) along the curve. The distance between two points Proposition 1. A Riemannian symmetric space is complete.
on the manifold can be defined as an infimum length of any
piecewise differentiable curve joining two points. We denote A complete Riemannian manifold M is a manifold where
the distance between p, q ∈ M as dM (p, q) and omit the the exponential map Expp (v) is defined for all v ∈ Tp M and
subscript when the space is clear. p ∈ M. The Hopf–Rinow theorem states that for complete
Riemannian manifolds, there exists a geodesic connecting any
Affine connections are a generalized notion of horizontal
two points on the manifold. Thus, the distance between two
components of directional derivatives on the surface of Rie-
points in this space can be defined as the shortest length of
mannian manifolds. Riemannian metric on the manifold deter-
the geodesic joining two points, namely geodesic distance.
mines a unique affine connection called Levi-Civita connection
that satisfies metric compatibility and symmetry. We hereafter Proposition 2. A Riemannian symmetric space is a homoge-
denote the Levi-Civita connection by ∇, and all connections neous space.
used in this supplement and the main article will be the Levi-
This result states that isometries on M are transitive. These
Civita connection. This connection is used to define derivatives
propositions are useful when extending statistical concepts in
of curves on the manifold denoted by ∇γ ′ , called covariant
Euclidean space to a manifold. For example, the normalizing
derivative along a curve γ : I ⊂ R → M. Note that
constant of Riemannian normal distributions appears in a
γ ′ = dγ/dt is the velocity of the curve that forms a vector
handy form for Riemannian symmetric spaces, as stated in
field along γ. From the covariant derivative, two important
the Proposition 1 of the main article.
concepts are derived. First, the notion of ‘acceleration’ of a
curve can be obtained by ∇γ ′ γ ′ . When ∇γ ′ γ ′ is zero along
the curve γ, we call γ a geodesic. Geodesics are a natural S-II. D ETAILS OF E QUATION (12)
analog of ‘straight lines’ in Euclidean space. Second, for a This section reproduces Section 2.1 of [3], but we repeat it
curve γ and a tangent vector v ∈ Tγ(t0 ) M where t0 ∈ I, here for self-contained material.
there exists a unique tangent vector field V along γ such The roughness penalty matrix K in equation (12) is defined
that (∇γ ′ V )(t) = 0 for all t ∈ I and V (t0 ) = v. We call as K = QR−1 Q⊤ , where Q ∈ Rn×(n−2) and R ∈
V (t) ∈ Tγ(t) M the parallel transport of v along γ. For a R(n−2)×(n−2) are band matrices defined in the following way.
special case that γ is a geodesic, the map is called parallel Denoting hi = ai+1 − ai for i = 1, . . . , n − 1, Q = (qij ) is
γ(t ) defined as
transport along a geodesic, and we denote it by Pγ(t)0 (v). In
fitting a probabilistic principal curve, we approximate a curve qj−1,j = h−1 qjj = −h−1 −1
qj+1,j = h−1
j−1 , j−1 − hj , j ,
such that a parallel transport along a geodesic is used. All
spaces chosen for simulation in Section IV and Section S- for j = 2, . . . , n − 1, and qij = 0 for |i − j| ≥ 2. Note that the
V possess a closed-form expression for the parallel transport index for the columns of Q starts from 2. Hence, the upper
along a geodesic. left element is q12 . R = (rij ) is a symmetric matrix defined
A geodesic γv (t) is defined as its initial point, p = γv (0) ∈ as
M and velocity, v = γv′ (0) ∈ Tp M. Then, the exponential 1
rii = (hi−1 + hi ), for i = 2, . . . , n − 1,
map at p, Expp : Tp M → M is defined by Expp (v) = γv (1). 3
The logarithmic maps, Logp , are the local inverses of the 1
ri,i+1 = ri+1,i = hi , for i = 2, . . . , n − 2,
exponential maps. The exponential and logarithmic maps are 6
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 2

and otherwise 0. Similarly, indices for the rows and columns and the final one is from the definition of unrolled curve f † .
of R start from 2.
There are two remarks.
PD First, the roughness penalty in (11) Proof of Proposition 4. By Lemma 1, we have
can be written as λ j=1 (f †j )⊤ Kf †j by Theorem 2.1 of [3].
From this result, it can be shown that (12) is an explicit Q† (θ|θ0 ) − Q(θ|θ0 )
solution by differentiating Q† by f †j . Second, when evaluating n h i
wili log gY † |ali (yi† ) − log gY |ali (yi )
X
(12), we apply the Reinsch algorithm for a faster computation. =
That is, the band matrices Q and R are directly used for the i=1
" #
n
evaluation of (12) instead of explicitly computing K. Utilizing X ∥f † (ali ) − yi† ∥2 d(f (ali ), yi )2
that Q and R are band matrices, O(n) time computation of = − +
i=1
2σl2i 2σl2i
(12) becomes possible. Refer to Section 2.3.3 of [3] for details.
= 0,
S-III. P ROOFS FOR S ECTION III
since ∥f † (ali ) − yi† ∥2 = d(f (ali ), yi )2 with the aid of the
A. Proof of Equation (10) metric compatible connection (Levi-Civita connection).
Substituting σk2 as 1/τk in equation (8), we get
n X n S-IV. I NITIALIZATION OF THE S CALE PARAMETER
X  τk 
Q(θ|θ0 ) = wik − log C(τk ) − d(yi , f (ak ))2 We suggest a simple method for selecting the initial value
2
i=1 k=1
n n
c of the scale parameter σk2 . The gradient ascent algorithm
Z 1
XX optimizes scale parameters, so approximating the variance
+ wik log vk − λ ∥(∇f ′ f ′ )(t)∥2 dt.
0 locally at each point is sufficient. Thus, we provide the
i=1 k=1
following rule-of-thumb initializer for σk2 ,
The last two terms of the above equation are constants w.r.t. τk .
1) For each data point yi , (i = 1, . . . , n)
That is, these terms vanish when differentiating w.r.t. τk . Also,
on a D-dimensional manifold, calculate
C(τk ) is defined as equation (4) in the main paper. Hence, 1
P[0.1n]
σ̂i2 := [0.1n] 2
j=1 d(ỹj , yi ) , where ỹj s are data
Z i(M) 2 D
points that belong to the 0.1n-nearest neighborhood of
 2 Y
dC(τk ) r r
= −AD−1 exp − τk hk (r)dr. yi and [·] is the rounding operator.
dτk 0 2 2
k=2 2) Set the initial value of σk2 to c = mediani=1,...,n (σ̂i2 ) ×
Similar derivation also appears in [4]. Plugging in dC(τk )/dτk χ20.8 (D), where χ20.8 (D) is the quantile satisfying
derived above to dQ(θ|θ0 )/dτk , equation (10) is immediate. P(X ≤ χ20.8 (D)) = 0.8 and X ∼ χ2 (D).
Suppose that Y is a random vector following a multivariate
B. Proof of Proposition 3 normal distribution N (0, σ 2 ID ) in a D-dimensional Euclidean
Due to Proposition 2 in the main paper, we compare only the space. Then, P(∥Y ∥2 ≤ σ 2 χ20.8 (D)) = 0.8 holds. Intuitively
two first terms of equations (8) and (11) in the article. When speaking, we have chosen c by regarding each data point as
M = RD , ∥yi† − f † (ak )∥2 = ∥yi − f (ak )∥2 = d(yi , f (ak ))2 , a center of a normal distribution that generates nearby data
for each 1 ≤ i, k ≤ n. Thus, equality holds. points.

C. Proof of Proposition 4 S-V. N UMERICAL E XPERIMENTS

Before proving Proposition 4, we present the following A. Simulation Study
lemma. In this section, we conduct a simulation study to evaluate
the empirical performance of the proposed method. To this
Lemma 1. For Rthe Levi-Civita connection ∇,
R1 1 end, we simulate several data sets on three different Rieman-
∥(∇f ′ f ′ )(t)∥2 dt = 0 ∥f¨† (t)∥2 dt.
0 nian symmetric spaces: hyperbolic space H2 , sphere S2 , and
Proof. It follows that the following equalities hold compact Lie group SO(3). For all experiments, we used the
Z 1 first principal geodesic as an initial curve.
∥(∇f ′ f ′ )(t)∥2 dt For comparison, we consider the following existing meth-
0 ods: (i) PGA: the first principal geodesic of [5], (ii) PNS: the
2
f (t+h)
Z 1 Pf (t) (f ′ (t + h)) − f ′ (t) principal nested sphere of [6], (iii) HPC: principal curves of
= lim dt [7], and (iv) SPC: spherical principal curve of [8]. Note that
0 h→0 h
PNS and SPC only apply to a spherical domain, and HPC and
Z 1 2 SPC use a biweight kernel.
d f (t) ′
= P (f (t)) dt 1) Poincaré Half-Plane: Poincaré half-plane H2 :=
0 dt f (0)
Z 1 {(x, y) ∈ R2 : y > 0} is a two-dimensional conformal hy-
= ∥f¨† (t)∥2 dt. perbolic space, having a negative constant Gaussian curvature
0 of −1. The metric tensor at (x, y) is given by I2 /y 2 , and
The first equality is from parallel transport [1], the second is geodesic induced by the metric is a straight vertical line or
from the metric compatibility of the Levi-Civita connection, a semicircle with its center on the x-axis. Exponential and
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 3

logarithmic maps and parallel transport along a geodesic are of HPC in that a greedy approach is used to project each point
explicitly obtained, which leads to a fast computation of our to knots.
algorithm. 2) The Unit Sphere S2 : The two-dimensional unit sphere
Hyperbolic space is a useful geometry for analyzing net- is a well-known model of the Earth, and any data on the Earth
work data. Many real-world network data resemble tree struc- can be analyzed with low distortion compared to the analysis
tures, and using a hyperbolic space as an underlying latent of data points on a flat map of the Earth. Spatial distributions
space is known to reflect such a structure [9], [10]. The half- of earthquakes [8], [13], [14] and the Polar wander paths [15]
plane space is also useful for analyzing univariate normal are example data that can be dealt with in a two-dimensional
distributions. When using the Fisher information matrix as sphere.
a metric tensor in the parameter space of univariate normal Note that the canonical metric is used P 2on the unit sphere
distribution, the geometry of H2 can be induced [11]. Mapping Sn := {(x1 , . . . , xn+1 ) ∈ Rn+1 : xi = 1}, which is
a normal distribution to a negative curvature space is widely induced by the usual Euclidean metric on Rn+1 . Exponential
adopted in information geometry [11], [12]. and logarithmic maps and parallel transport along a geodesic
For simulated data sets, we consider four types of curves, as can be explicitly obtained with the aid of ambient space Rn+1
in [7], and sample 100 data points from each true curve. Then, of Sn .
data points are generated by adding random noises using the For the simulation study, we consider two data sets on S2 ,
exponential map. The detailed forms are below. Wave and Arc in Section IV-A of the main paper. Figure 2
shows the fitting results by applying the proposed PPC and
1. V-shape: (xi , yi ) = Exp(ti ,2+3/4×|ti |) (ϵi ),
other existing methods to the two data sets. Black dots denote
where ϵi ∼ NR2 (0, (1/8)2 I2 ). the data points, and black curves are the true curves. For SPC
2. Straight: (xi , yi ) = Exp(ti ,10) (ϵi ), and HPC, we set T = n for a fair comparison. From the
where ϵi ∼ NR2 (0, (1/32)2 I2 ). results in the upper panels, we observe that the proposed PPC
3. Wave: (xi , yi ) = Exp(ti ,5−sin ti ) (ϵi ), outperforms SPC and HPC. As one can see, the proposed PPC
where ϵi ∼ NR2 (0, (1/16)2 I2 ). represents the Wave data as a smooth curve, while the fits by
4. Parabola: (xi , yi ) = Exp(ti ,10+t2i /2) (ϵi ), SPC and HPC fail to capture the continuous feature of the true
where ϵi ∼ NR2 (0, (1/16)2 I2 ). curve. The first two panels in the bottom row show the fitting
Here the design points ti ’s (i = 1, . . . , 100) are i.i.d. samples results by PNS and PGA, which do not represent the Wave
from Uniform[−3, 3], which are mapped on the true curve. The data well.
error vector ϵi = (ϵi1 , ϵi2 )⊤ is associated with each tangent We conducted a numerical evaluation of the PPC to assess
∂ ∂ its performance in reconstructing the true curve. Wave data set
vector at p, ϵi1 ∂x p
+ ϵi2 ∂y p
.
The first row of Figure 1 shows the first three simulated data on a two-dimensional sphere for various noise levels σ were
sets (dots), true curves (dashed curves), and fits (red curves) considered:
by the proposed PPC. The gray semi-circles denote the initial 5. Wave with noise: (θi , ϕi ) = Exp(θi ,sin(3θi )/3) (ϵi ), where
curves by PGA. From the results, we observe that the proposed ϵi ∼ NR2 (0, σ 2 I2 ), θi ∼ Uniform[0, 1.5π], (i =
PPC fits non-geodesic trends of the data well and reconstructs 1, . . . , 100).
the true curves effectively. On the other hand, the first principal We P employ a reconstruction error defined as
1 100
geodesic fails to represent data points. In the second row, we 100 i=1 d(ti , πf (ti )), where ti denote equidistant points on
compare the proposed method with HPC under a simulated the true curve and πf (ti ) denotes the projection of ti to the
data set from Parabola. The proposed method represents curve f . The reconstruction error represents a discrepancy
the curve with n discrete knots, where n is the number of data between the true curve and the fitted curve. Table I lists the
points. The smoothness of PPC is controlled by parameter λ, average values of reconstruction errors and their standard
while HPC controls the curve’s smoothness by two parameters, deviations over 50 simulation sets while changing the tuning
the number of knots T and the bandwidth q. A larger λ (T and parameters λ (PPC) and q (SPC and HPC). The dashed
q) leads to a smoother curve. To compare the two methods, we slots in the table are excluded because curves fitted in
have a fixed number of knots of HPC to T = n so that PPC the corresponding settings are too wiggly. As listed, PPC
and HPC are compared in the same condition. In panels (d) reconstructs the true curve more precisely when the noise
and (e), both methods are implemented by several smoothing level is low. Figure 3 shows PPCs fitted to Wave data
parameters; for PPC, λ = 0.005, 0.1, 2, and for HPC, q = generated from various σ values. We observe that PPC is
0.12, 0.13, 0.14, 0.15, 0.16. As one can see, the fits by quite robust to noisy data. When the noise level is high,
PPC are smoother than those of HPC. This is mainly due to PPC tends to smooth parts with high curvatures. For noise
the different smoothing schemes used in each method. HPC levels σ = 0.05, 0.1, and 0.2, PPC with λ = 0.003, SPC
uses kernel smoothing that only concerns locations of nearby with q = 0.06, and SPC with q = 0.08 have the lowest
points, while our method uses a roughness penalty smoother reconstruction error, respectively. Yet, there is no significant
and takes care of the first and second derivative of the curve as difference in performance between PPC and SPC. In addition,
well. Specifically, the proposed PPC fits twice differentiable at all noise levels, PPC performed better than HPC. As a
curve at the tangent space. Note that to make the HPC fit result, PPC reconstructs the true curve more accurately than
smoothly, we need to increase the number of knots. However, the parametric curve fitting methods, PGA and PNS. Also,
increasing the number of knots affects the computational time PPC performs similarly to SPC. However, PPC fits a smooth
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 4

(a) V−shape (b) Straight (c) Wave

6.5
4.5

10.5

6.0
4.0

5.5
3.5

10.0

5.0
y

y
3.0

4.5
2.5

9.5

4.0
2.0

3.5
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

x x x

(d) Parabola, PPC (e) Parabola, HPC (f) Parabola, PGA

15

15

15
λ=0.005 q=0.12
λ=0.1 q=0.13
λ=2 q=0.14
q=0.15
14

14

14
q=0.16
13

13

13
12

12

12
y

y
11

11

11
10

10

10
9

9
−2 0 2 4 −2 0 2 4 −2 0 2 4

x x x

Fig. 1. PPC and other methods applied to four simulated data sets on the Poincaré half-plane. Black dots are the data points, and black dashed curves denote
the true curve. (a-c) Fit (red curve) by applying PPC to V-shape, Straight, and Wave data sets, respectively. Gray semi-circles denote the initial curve
by PGA. (d) Fits by applying PPC with various λ’s to Parabola data set. (e) Fits by applying HPC with different bandwidth q’s to Parabola data set.
(f) Fit by applying PGA to Parabola data set.

curve effectively, whereas SPC and HPC fail to capture TABLE I

the continuous feature of the true curve. Furthermore, as AVERAGES OF R ECONSTRUCTION E RRORS (T HEIR S TANDARD
D EVIATIONS ) BY A PPLYING PPC AND OTHERS TO Wave DATA ON S2 .
pointed out in Section III of the main paper, our method
does not have a model bias with a probabilistic formulation σ = 0.05 σ = 0.1 σ = 0.2
of the curve. In contrast, self-consistent curves such as SPC λ = 0.0003 0.0250 (0.0041) 0.0424 (0.0069) 0.0761 (0.0167)
and HPC inherently have a model bias. This is the primary PPC λ = 0.003 0.0218 (0.0044) 0.0364 (0.0077) 0.0714 (0.0175)
λ = 0.03 0.0349 (0.0077) 0.0676 (0.0144) 0.1310 (0.0163)
motivation for the proposed method. However, both methods q = 0.01 0.0276 (0.0038) — —
have estimation biases from smoothers or roughness penalties q = 0.02 0.0255 (0.0040) — —
adopted when estimating principal curves for a given data q = 0.03 0.0233 (0.0042) 0.0433 (0.0071) —
q = 0.04 0.0220 (0.0041) 0.0396 (0.0064) —
set, as listed in Table I. SPC
q = 0.05 0.0223 (0.0045) 0.0368 (0.0064) 0.0771 (0.0103)
3) Special Orthogonal Group SO(3): We now consider a q = 0.06 0.0251 (0.0067) 0.0355 (0.0067) 0.0727 (0.0123)
three-dimensional Riemannian symmetric space, the special q = 0.07 0.0313 (0.0118) 0.0362 (0.0072) 0.0682 (0.0122)
q = 0.08 0.0378 (0.0124) 0.0382 (0.0078) 0.0648 (0.0126)
orthogonal group SO(3) := {R ∈ GL(3, R) : R⊤ R = q = 0.01 0.0271 (0.0038) — —
I, det(R) = 1}, where GL(3, R) is the general linear group, q = 0.02 0.0274 (0.0050) — —
3 × 3 matrices of nonzero determinants. SO(3) is the space q = 0.03 0.0290 (0.0044) 0.0465 (0.0083) —
q = 0.04 0.0378 (0.0080) 0.0473 (0.0098) —
of rotations in R3 and is used to express an object’s position HPC
q = 0.05 0.4915 (0.0118) 0.0522 (0.0126) 0.0801 (0.0145)
within a three-dimensional frame. For example, human move- q = 0.06 0.0601 (0.0141) 0.0583 (0.0171) 0.0785 (0.0147)
ment data can be analyzed on the space of rotation matrices q = 0.07 0.0714 (0.0143) 0.0634 (0.0213) 0.0786 (0.0167)
q = 0.08 0.0808 (0.0151) 0.7326 (0.0245) 0.0843 (0.0192)
[7], [16]. By the bi-invariant metric ⟨H1 , H2 ⟩R = tr(H1⊤ H2 ), PNS 0.1952 (0.0128) 0.1946 (0.0122) 0.1939 (0.0116)
the exponential and logarithmic maps and parallel transport PGA 0.7949 (0.0468) 0.8033 (0.0522) 0.8041 (0.0656)
along a geodesic are explicitly derived, using the matrix expo-
nential and matrix logarithm.PNote that the matrix exponential
k
is defined by exp(A) := k≥0 A /k! = ExpI (A), where of the matrix exponential. Also, the tangent space at the
A0 = I and the matrix logarithm is defined as the inverse identity I is the space of real skew-symmetric matrices, i.e.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 5

Fig. 2. PPC and other methods applied to the simulated data on the unit sphere S2 . Black dots are the data points, and black curves are the true curve. From
upper left, fits by applying PPC with different λ to Wave data; fits by SPC and HPC with five varying bandwidths, respectively; the first principal nested
sphere; first principal geodesic; and fit by applying PPC to Arc data with λ = 0.1.

6. Arc on SO(3):
 
0 − cos ti −1
Ti = ExpI cos ti 0 sin ti  ,
1 − sin ti 0
  
0 ϵi1 ϵi2
Di = ExpTi Ti × −ϵi1 0 ϵi3  .
−ϵi2 −ϵi3 0
7. Wave on SO(3):
 
0 − cos ti − cos(3ti )/2
Ti = ExpI  cos ti 0 − sin ti  ,
cos(3ti )/2 sin ti 0
  
0 ϵi1 ϵi2
Di = ExpTi Ti × −ϵi1 0 ϵi3  .
−ϵi2 −ϵi3 0
Here ϵi = (ϵi1 , ϵi2 , ϵi3 )⊤ ∼ NR3 (0, (0.05)2 I3 ).
Figure 4 shows the fitting results by the proposed PPC for
the two data sets. Each column of the rotation matrix in Figure
Fig. 3. PPC derived when λ = 0.003, for various noise levels in the Wave
data. 4 is drawn as a line segment on R3 , starting from the origin.
Three orthogonal vectors form a rectangular parallelepiped on
R3 . Note that the lengths of vectors are adjusted differently to
distinguish them from each other. The first and second rows
are trajectories of the true curve, and the PPC fit for the Arc
data set. The third and fourth rows correspond to the true
if A ∈ TI SO(3), A = −A⊤ , and the tangent space at R is curve and the PPC fit for the Wave data set. As one can
{RS|S ∈ TI SO(3)}. see, the proposed PPC represents the two data sets on SO(3)
effectively. We further provide a multidimensional scaling plot
Two data sets, Arc, Wave on SO(3) are used for the simu- for the two data sets. Figure 5 shows two data points (dots),
lation study. Each data point is generated using the exponential two true curves (black), and the corresponding fits by PGA
map. Ti denote the points on the true curve, and data points (gray) and PPC (red) in a two-dimensional multidimensional
Di are generated by adding noise ϵi to Ti . The design points scaling domain. We can also see why the names of the two
ti are sampled from Uniform[0, π] (i = 1, . . . , 100). data sets are Arc and Wave.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 6

[7] S. Hauberg, “Principal curves on Riemannian manifolds,” IEEE Trans-

actions on Pattern Analysis and Machine Intelligence, vol. 38, no. 9,
pp. 1915–1921, 2016.
[8] J. Lee, J.-H. Kim, and H.-S. Oh, “Spherical principal curves,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 43,
no. 6, pp. 2165–2171, 2021.
[9] D. Krioukov, F. Papadopoulos, M. Kitsak, A. Vahdat, and M. Bo-
guná, “Hyperbolic geometry of complex networks,” Physical Review
E, vol. 82, no. 3, p. 036106, 2010.
[10] A. L. Smith, D. M. Asta, and C. A. Calder, “The geometry of continuous
latent space models for network data,” Statistical Science, vol. 34, no. 3,
pp. 428–453, 2019.
[11] O. Calin and C. Udrişte, Geometric Modeling in Probability and
Fig. 4. Three-dimensional plots of PPC fits to the data points on SO(3). Statistics. Cham, Switzerland: Springer, 2014.
Trajectories of the Arc’s true curve and the PPC fit with λ = 0.1 (first two [12] S.-I. Amari, Information Geometry and Its Applications. Tokyo, Japan:
rows), and trajectories of the Wave’s true curve and the PPC fit with λ = 0.05 Springer, 2016, vol. 194.
(last two rows). [13] V. M. Panaretos, T. Pham, and Z. Yao, “Principal flows,” Journal of the
American Statistical Association, vol. 109, no. 505, pp. 424–436, 2014.
B. Real Data Analysis: Seismological Data [14] Z. Yao and Z. Zhang, “Principal boundary on Riemannian manifolds,”
Journal of the American Statistical Association, vol. 115, no. 531, pp.
In this section, we analyze a seismological data set from 1435–1448, 2020.
the United States Geological Survey (https://earthquake.usgs. [15] P. E. Jupp and J. T. Kent, “Fitting smooth paths to spherical data,”
Journal of the Royal Statistical Society: Series C (Applied Statistics),
gov/earthquakes). A similar data set was used in [13] as a vol. 36, no. 1, pp. 34–46, 1987.
real-world application of their curve-fitting method. This data [16] D. Rancourt, L.-P. Rivest, and J. Asselin, “Using orientation statistics
set consists of the epicenter locations of 143 earthquakes that to investigate variations in human kinematics,” Journal of the Royal
Statistical Society: Series C (Applied Statistics), vol. 49, no. 1, pp. 81–
occurred in 2022, distributed over a sizable rectangular grid 94, 2000.
(longitudes 10° to 80° and latitudes −26° to −70°). Because
this region covers a large portion of the Earth, the curvature
of the Earth should be considered during the study. That is,
the sample space should be a two-dimensional sphere.
Earthquakes occur near the boundary of tectonic plates,
which can be viewed as a one-dimensional trend of epicenters.
As the boundaries of the plates do not follow a geodesic curve
in this region, a principal curve should be used to describe the
variation in the data.
We initially eliminated epicenters recorded on the African
continent that appear to be outlying observations, i.e., obser-
vations that cannot be described by a curve, to fit PPC to the
data set. So, we used 123 observations and obtained PPCs with
various tuning parameter values of λ. Figure 6 shows three
PPCs with λ = 0.1, 1, 10. Note that the curve is fitted on S2
and is mapped to a flat map for the purpose of illustration. This
data set can be used to show a multiscale view of the variation
in the pattern. For example, for a relatively small value of
λ = 0.1, the PPC fit closely follows fine-scale variations in
the data, resembling the tectonic plate boundaries. For a large
value of λ = 10, the resulting PPC describes the large-scale
variation of the data, the east-to-west variation.

R EFERENCES
[1] M. P. do Carmo, Riemannian Geometry. Boston, MA, USA: Birkhäuser,
1992.
[2] L. W. Tu, Differential Geometry: Connections, Curvature, and Charac-
teristic Classes. Cham, Switzerland: Springer, 2017, vol. 275.
[3] P. J. Green and B. W. Silverman, Nonparametric Regression and
Generalized Linear Models: A Roughness Penalty Approach. New
York, NY, USA: Chapman and Hall/CRC, 1993.
[4] M. Zhang and P. T. Fletcher, “Probabilistic principal geodesic analysis,”
Advances in Neural Information Processing Systems, vol. 26, pp. 1178–
1186, 2013.
[5] P. T. Fletcher, C. Lu, S. M. Pizer, and S. Joshi, “Principal geodesic
analysis for the study of nonlinear statistics of shape,” IEEE Transactions
on Medical Imaging, vol. 23, no. 8, pp. 995–1005, 2004.
[6] S. Jung, I. L. Dryden, and J. S. Marron, “Analysis of principal nested
spheres,” Biometrika, vol. 99, no. 3, pp. 551–568, 2012.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 7

1.0

0.5
0.5
MDS 2

MDS 2

0.0
0.0

−0.5
−0.5

−1.0 −0.5 0.0 0.5 1.0 1.5 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

MDS 1 MDS 1

Fig. 5. Two-dimensional multidimensional scaling plot of PPC fits to the data points on SO(3). The PPC fit with λ = 0.1 to the Arc data, and the PPC fit
with λ = 0.05 to the Wave data. Black curves denote the true curves, red curves are the PPC, and the gray lines are the initial curve by PGA.

0°

10°S
1.0

λ=10
λ=1
λ=0.1
20°S
Latitude

30°S
0.5

40°S
c(0, 0)

50°S
0.0

60°S
0° 20°E 40°E 60°E 80°E
Longitude

Fig. 6. Epicenters of earthquakes observed in 2022 in a specific area below Africa (black dots) and PPC fitted by various λ values.
−0.5