Spectral Mesh Processing

SIGGRAPH Asia 2009 Course 32

Bruno Lévy (INRIA, France)

Hao (Richard) Zhang (Simon Fraser University, Canada)
About the instructors

Bruno Lévy
INRIA, centre Nancy Grand Est,
rue du Jardin Botanique,
54500 Vandoeuvre, France
Email: bruno.levy@inria.fr
http://alice.loria.fr/index.php/bruno-levy.html

Bruno Lévy is a researcher with INRIA. His main contribution concerns pa-
rameterization of triangulated surfaces (LSCM), and is now used by some 3D
modeling software (including Maya, Silo, Blender, Gocad and Catia). He ob-
tained his Ph.D in 1999, and was hired by INRIA in 2000. Since 2004, he has
been leading Project ALICE - Geometry and Light, that aims at developping
mathematical tools for next generation geometry processing. He served on the
committee of ACM SPM, IEEE SMI, ACM/EG SGP, IEEE Visualization, Eu-
rographics, PG, ACM Siggraph, and was program co-chair of ACM SPM in 2007
and 2008, and will be program co-chair of SGP 2010. He was recently awarded
a ”Starting Grant” from the European Research Council (3% acceptance, all
disciplines of science).

Hao (Richard) Zhang

School of Computing Science
Simon Fraser University, Burnaby, Canada V5A 1S6
Email: haoz@cs.sfu.ca
http://www.cs.sfu.ca/˜haoz

Hao (Richard) Zhang is an Associate Professor in the School of Computing

Science at Simon Fraser University, Canada, and he co-directs the Graphics,
Usability, and Visualization (GrUVi) Lab. He received his Ph.D. from the
University of Toronto in 2003 and M. Math. and B. Math. degrees from the
University of Waterloo. His research interests include geometry processing,
shape analysis, and computer graphics. He gave a Eurographics State-of-the-
art report on spectral methods for mesh processing and analysis in 2007 and
wrote the first comprehensive survey on this topic. Recently, he has served on
the program committees of Eurographics, ACM/EG SGP, ACM/SIAM GPM,
and IEEE SMI. He was a winner of the Best Paper Award from SGP 2008.
Course description

Summary statement: In this course, you will learn the basics of Fourier
analysis on meshes, how to implement it, and how to use it for filtering, remesh-
ing, matching, compressing, and segmenting meshes.

Abstract: Spectral mesh processing is an idea that was proposed at the be-
ginning of the 90’s, to port the “signal processing toolbox” to the setting of
3D mesh models. Recent advances in both computer horsepower and numerical
software make it possible to fully implement this vision. In the more classi-
cal context of sound and image processing, Fourier analysis was a corner stone
in the development of a wide spectrum of techniques, such as filtering, com-
pression, and recognition. In this course, attendees will learn how to transfer
the underlying concepts to the setting of a mesh model, how to implement the
“spectral mesh processing” toolbox and use it for real applications, including
filtering, shape matching, remeshing, segmentation, and parameterization.

Background: Some elements of this course appeared in the “Geometric Mod-

eling based on Polygon Meshes” course (SIGGRAPH 07, Eurographics 08), the
“Mesh Parameterization, Theory and Practice” course (SIGGRAPH 07), and
the “Mesh Processing” course (invited course, ECCV 08). Based on some feed-
back from the attendees, it seems that there is a need for a course focused on the
spectral aspects. Whereas these previous courses just mentioned some applica-
tions of spectral methods, this course details the notions relevant to spectral
mesh processing, from theory to implementation. Spectral methods for mesh
processing had been presented as a Eurographics State-of-the-art report in 2007
and a subsequent survey was completed in 2009. Some coverages in the current
notes, in particular those on applications, have been selected from the survey.
However, we start with a more basic setup this time and include updates from
some of the most recent developments in spectral mesh processing.

Intended audience: Researcher in the areas of geometry processing, shape

analysis, and computer graphics, as well as practitioners who want to learn more
about how to implement these new techniques and what are the applications.

Prerequisites: Knowledge about mesh processing, programming, mesh data

structures, basic notions of linear algebra and signal processing.
Course syllabus

• (5 min) Introduction [Lévy]

• (45 min) “What is so spectral?” [Zhang]

Intuition and theory behind spectral methods
Different interpretations and motivating applications
• (55 min) Do your own Spectral Mesh processing at home [Lévy]
DEC Laplacian, numerics of spectral analysis
Tutorial on implementation with open source software

• (15 min) break

• (50 min) Applications - “What can we do with it ?” 1/2 [Zhang]

Segmentation, shape retrieval, non-rigid matching, symmetry detection,
...

• (40 min) Applications - “What can we do with it ?” 2/2 [Lévy]

Quadrangulation, parameterization, . . .
• (15 min) Wrapup, conclusion, Q&A [Zhang and Lévy]
 Spectral mesh processing involves the use of eigenvalues, eigenvectors, or
eigenspace projections derived from appropriately defined mesh operators to
carry out desired tasks. Early work in this area can be traced back to the
seminal paper by Taubin [Tau95a] in 1995, where spectral analysis of mesh
geometry based on a graph Laplacian allows us to understand the low-pass
filtering approach to mesh smoothing. Over the past fifteen years, the list of
geometry processing applications which utilize the eigenstructures of a variety
of mesh operators in different manners have been growing steadily.
Most spectral methods for mesh processing have a basic framework in com-
mon. First, a matrix representing a discrete linear operator based on the topo-
logical and/or geometric structure of the input mesh is constructed, typically as
a discretization of some continuous operator. This matrix can be seen as incor-
porating pairwise relations between mesh elements, adjacent ones or otherwise.
Then an eigendecomposition of the matrix is performed, that is, its eigenvalues
and eigenvectors are computed. Resulting structures from the decomposition
are employed in a problem-specific manner to obtain a solution.
We will look at the various applications of spectral mesh processing towards
the end of the course notes (Section 5), after we provide the intuition, moti-
vation, and some theory behind the use of the spectral approach. The main
motivation for developing spectral methods for mesh processing is the pursuit
of Fourier analysis in the manifold setting, in particular, for meshes which are
the dominant discrete representations of surfaces in the field of computer graph-
ics. There are other desirable characteristics of the spectral approach including
effective and information-preserving dimensionality reduction and its ability to
reveal global and intrinsic structures in geometric data [ZvKDar]. These should
become clear as we discuss the applications.
We shall start the course notes with a very gentle introduction in Section 1.
Instead of diving into 3D data immediately, we first look at the more classic case
of processing a 2D shape represented by a contour. The motivating application
is shape smoothing. We reduce the 2D shape processing problem to the study
of 1D functions or signals specifying the shape’s contour, naturally exposing the
problem in a signal-processing framework. Our choice to start the coverage in
the discrete setting is intended to not involve the heavy mathematical formu-
lations at the start so as to better provide an intuition. We present Laplacian
smoothing and show how spectral processing based on 1D discrete Laplace oper-
ators can perform smoothing as well as compression. The relationship between
this type of spectral processing and the classical Fourier transform is revealed
and a natural extension to the mesh setting is introduced.
Having provided an intuition, we then instill more rigor into our coverage and
more formally present spectral mesh processing as a means to perform Fourier
analysis on meshes. In particular, we deepen our examination on the connec-
tion between the continuous and the discrete settings, focusing on the Laplace
operator and its eigenfunctions. While Section 2 provides some theoretical
background in the continuous setting (eigenfunctions of the Laplace-Beltrami
operator) and establishes connections with other domains (machine learning
and spectral graph theory). Section 3 is concerned with ways to discretize the
Laplace operator. Since spectral mesh processing in practice often necessitates
the computation of eigenstructures of large matrices, Section 4 presents ways to
make that process efficient. Finally, we discuss various applications of spectral
mesh processing in Section 5. For readers who wish to obtain a more thorough
coverage on the topic, we refer to the survey of Zhang et al. [ZvKDar].

1 A gentle introduction
Consider the seahorse shape depicted by a closed contour, shown in Figure 1.
The contour is represented as a sequence of 2D points (the contour vertices) that
are connected by straight line segments (the contour segments), as illustrated
by a zoomed-in view. Now suppose that we wish to remove the rough features
over the shape of the seahorse and obtain a smoothed version, such as the one
shown in the right of Figure 1.

Figure 1: A seahorse with rough features (left) and a smoothed version (right).

1.1 Laplacian smoothing

A simple procedure to accomplish this is to repeatedly connect the midpoints
of successive contour segments; we refer to this as midpoint smoothing. Fig-
ures 2(a) illustrates two such steps. As we can see, after two steps of midpoint
smoothing, each contour vertex vi is moved to the midpoint of the line seg-
ment connecting the midpoints of the original contour segments adjacent to vi .
Specifically, let vi−1 = (xi−1 , yi−1 ), vi = (xi , yi ), and vi+1 = (xi+1 , yi+1 ) be
three consecutive contour vertices. Then the new vertex v̂i after two steps of
midpoint smoothing is given by a local averaging,
   
1 1 1 1 1 1 1
v̂i = (vi−1 + vi ) + (vi + vi+1 ) = vi−1 + vi + vi+1 . (1)
2 2 2 2 4 2 4

The vector between vi and the midpoint of the line segment connecting the
two vertices adjacent to vi , shown as a red arrow in Figure 2(b), is called the
 midpoint midpoint
smoothing smoothing

(a) Two steps of midpoint smoothing.

vi
Laplacian
smoothing
vi ! 1
v i+1

(b) Laplacian smoothing. (c) 1D Laplacians (red).

Figure 2: One step of Laplacian smoothing (b) is equivalent to two steps of

midpoint smoothing (a). The 1D discrete Laplacian vectors (c) are in red.

1D discrete Laplacian at vi , denoted by δ(vi ),

1
δ(vi ) = (vi−1 + vi+1 ) − vi . (2)
2
As we can see, after two steps of midpoint smoothing, each contour vertex is
displaced by half of its 1D discrete Laplacian, as shown in Figure 2(c); this is
referred to as Laplacian smoothing. The smoothed version of the seahorse in
Figure 1 was obtained by applying 10 steps of Laplacian smoothing.

1.2 Signal representation and spectral transform

Let us denote the contour vertices by a coordinate vector V , which has n rows
and two columns, where n is the number of vertices along the contour and the
two columns correspond to the x and y components of the vertex coordinates.
Let us denote by xi (respectively, yi ) the x (respectively, y) coordinate of a vertex
vi , i = 1, . . . , n. For analysis purposes, let us only consider the x component of
V , denoted by X; the treatment of the y component is similar.
We treat the vector X as a discrete 1D signal. Since the contour is closed,
we can view X as a periodic 1D signal defined over uniformly spaced samples
along a circle, as illustrated in Figure 3(a). We sort the contour vertices in
counterclockwise order and plot it as a conventional 1D signal, designating an
arbitrary element in X to start the indexing. Figure 3(b) shows such a plot for
the x-coordinates of the seahorse shape (n = 401) from Figure 1. We can now
express the discrete 1D Laplacians (2) for all the vertices using an n × n matrix
 350

X 300

x−coordinates
x1 y1 250

x2 y2 xi!1
V= xi+1
......
......
xi 200

xn yn
150
0 100 200 300 400
i!1 i i+1 Contour vertex indexes

Figure 3: Left: The x-component of the contour coordinate vector V , X, can

be viewed as a 1D periodic signal defined over uniform samples along a circle.
Right: it is shown by a 1D plot for the seahorse contour from Figure 1.

L, called the 1D discrete Laplace operator , as follows:

1 − 21 − 12
 
0 ... ... 0
− 1 1 − 1
0 ... ... 0 
 2 2 
δ(X) = LX =  ... .. .. .. .. .. ..  X. (3)

 . . . . . . 
 0 ... ... 0 − 12 1 − 21 
− 12 0 ... ... 0 − 21 1

The new contour X̂ resulting from Laplacian smoothing (1) is then given by

  1 1 1
 
x̂1 2 4 0 ... ... 0 4 x1
 x̂2
 1 1 1
0 ... ... 0   x2
  4 2 4  
X̂ = 
  ..
.. .. .. .. .. .. ..   ..
 = .  = SX. (4)
 
 .
  . . . . . .

  .
x̂n−1   0 1 1 1 
... ... 0 4 2 4 xn−1

1 1 1
x̂n 4 0 ... ... 0 4 2
x n

The smoothing operator S is related to the Laplace operator L by

1
S = I − L.
2
To analyze the behavior of Laplacian smoothing, in particular what happens
in the limit, we rely on the set of basis vectors formed by the eigenvectors of
L. This leads to a framework for spectral analysis of geometry. From linear
algebra, we know that since L is symmetric, it has real eigenvalues and a set of
real and orthogonal set of eigenvectors which form a basis. Any vector of size
n can be expressed as a linear sum of these basis vectors. We are particularly
interested in such an expression for the coordinate vector X. Denote by e1 ,
e2 ,. . .,en the normalized eigenvectors of L, corresponding to eigenvalues λ1 , λ2 ,
. . ., λn , and let E be the matrix whose columns are the eigenvectors. Then we
 Eigenvalue ! =0.000000 Eigenvalue ! =0.000123 Eigenvalue ! =0.000123 Eigenvalue ! =0.000491
0.1 0.1 0.1 0.1

0.05 0.05 0.05 0.05

0 0 0 0

−0.05 −0.05 −0.05 −0.05

−0.1 −0.1 −0.1 −0.1

100 200 300 400 100 200 300 400 100 200 300 400 100 200 300 400

Eigenvalue ! =0.000491 Eigenvalue ! =0.001105 Eigenvalue ! =0.001105 Eigenvalue ! =0.001963

0.1 0.1 0.1 0.1

0.05 0.05 0.05 0.05

0 0 0 0

−0.05 −0.05 −0.05 −0.05

−0.1 −0.1 −0.1 −0.1

100 200 300 400 100 200 300 400 100 200 300 400 100 200 300 400

Figure 4: Plots of first 8 eigenvectors of the 1D discrete Laplace operator (n =

401) given in equation (3). They are sorted by the eigenvalue λ.

can express X as a linear sum of the eigenvectors,

E11 E1n E11 ... E1n x̃1

      
n
X  E21   E2n   E21 ... E2n   x̃2 
X=  ...  x̃1 + . . . +  ...  x̃n =  ...
ei x̃i =      ..   ...  = E X̃.
..   
i=1 . .
En1 Enn En1 ... Enn x̃n
(5)
The above expression represents a transform from the original signal X to a
new signal X̃ in terms of a new basis, the basis given by the eigenvectors of L.
We call this a spectral transform, whose coefficients X̃ can be obtained by

X̃ = E T X, where E T is the transpose of E,

and for each i, the spectral transform coefficient

x̃i = eT
i · X. (6)

That is, the spectral coefficient x̃i is obtained as a projection of the signal X
along the direction of the i-th eigenvector ei . In Figure 4, we plot the first 8
eigenvectors of L, sorted by increasing eigenvalues, where n = 401, maching
the size of the seahorse shape from Figure 1. The indexing of elements in
each eigenvector follows the same contour vertex indexing as X, the coordinate
vector; it was plotted in Figure 3(b) for the seahorse. It is worth noting that
aside from an agreement on indexing, the Laplace operator L and the eigenbasis
vectors do not depend on X, which specifies the geometry of the contour. L,
as defined in equation (3), is completely determined by n, the size of the input
contour, and a vertex ordering.
As we can see, the eigenvector corresponding to the zero eigenvalue is a
constant vector. As eigenvalue increases, the eigenvectors start to oscillate as
sinusoidal curves at higher and higher frequencies. Note that the eigenvalues of
 k = 3. k = 5. k = 10. k = 20. k = 30. k ≈ 12 n. Original.

Figure 5: Shape reconstruction via Laplacian-based spectral analysis.

L repeat (multiplicity 2) after the first one, hence the corresponding eigenvec-
tors of these repeated eigenvalues are not unique. One particular choice of the
eigenvectors reveals a connection of our spectral analysis to the classical Fourier
analysis; this will be discussed in Section 1.5.

1.3 Signal reconstruction and compression

With the spectral transform of a coordinate signal defined as in equation (5),
we can now look at compression and filtering of a 2D shape represented by a
contour. Analogous to image compression using JPEG, we can obtain compact
representations of a contour by retaining the leading (low-frequency) spectral
transform coefficients and eliminating the rest. Given a signal X as in equation
(5), the signal reconstructed by using the k leading coefficients is
k
X
X (k) = ei x̃i , k ≤ n. (7)
i=1

This represents a compression of the contour geometry since only k out of n

coefficients need to be stored to approximate the original shape. We can quantify
the information loss by measuring the L2 error
v
n
u n
X u X
(k)
||X − X || = || ei x̃i || = t x̃2i ,
i=k+1 i=k+1

which is simply the Euclidean distance between X and X (k) . The last equality
is easy to obtain if we note the orthogonality of the eigenvectors, i.e., eT
i · ej = 0
whenever i 6= j. Also, since the ei ’s are normalized, eTi · ei = 1.
In Figure 5, we show some results of this type of shape reconstruction (7),
with varying k for the seahorse and a bird shape. As more and more high-
frequency spectral coefficients are removed, i.e., with decreasing k, we obtain
 2 2

Spectral transform coefficients for X

Spectral transform coefficients for X

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 50 100 150 200 250 300 350 400 0 10 20 30 40 50 60 70
Eigenvalue indexes Eigenvalue indexes

Figure 6: Plot of spectral transform coefficients for the x component of a

contour. Left: seahorse. Right: bird. The models are shown in Figure 5.

smoother and smoother reconstructed contours. How effectively a 2D shape can

be compressed this way may be visualized by plotting the spectral transform
coefficients, the x̃i ’s in (6), as done in Figure 6. In the plot, the horizontal
axis represents eigenvalue indexes, i = 1, . . . , n, which roughly correspond to
frequencies. One can view the magnitude of the x̃i ’s as the energy of the input
signal X at different frequencies.
A signal whose energies are sharply concentrated in the low-frequency end
can be effectively compressed at a high compression rate, since as a consequence,
the total energy at the high-frequency end, representing the reconstruction error,
is very low. Such signals will exhibit fast decay in its spectral coefficients. Both
the seahorse and the bird models contain noisy or sharp features so they are
not as highly compressible as a shape with smoother boundaries. This can be
observed from the plots in Figure 6. Nevertheless, at 2:1 compression ratio, we
can obtain a fairly good approximation, as one can see from Figure 5.

1.4 Filtering and Laplacian smoothing

Compression by truncating the vector X̃ of spectral transform coefficients can
be seen as a filtering process. When a discrete filter function f is applied to X̃,
we obtain a new coefficient vector X̃ 0 , where X̃ 0 (i) = f (i) · X̃(i), for all i. The
filtered signal X 0 is then reconstructed from X̃ 0 by X 0 = E X̃ 0 , where E is the
matrix of eigenvectors as defined in equation (5). We next show that Laplacian
smoothing is one particular filtering process. Specifically, when we apply the
Laplacian smoothing operator S to a coordinate vector m times, the resulting
coordinate vector becomes,
n n
1 X 1 X 1
X (m) = S m X = (I − L)m X = (I − L)m ei x̃i = ei (1 − λi )m x̃i . (8)
2 i=1
2 i=1
2

Equation (8) provides a characterization of Laplacian smoothing in the spec-

tral domain via filtering and the filter function is given by f (λ) = (1 − 21 λ)m .
A few such filters with different m are plotted in the first row of Figure 7. The
 1 1 1 1

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2 0 0.5 1 1.5 2 0 0.5 1 1.5 2

Figure 7: First row: filter plots, (1 − 12 λ)m with m = 1, 5, 10, 50. Second row:
corresponding results of Laplacian smoothing on the seahorse.

corresponding Laplacian smoothing leads to attenuation of the high-frequency

content of the signal and hence achieves denoising or smoothing.
To examine the limit behavior of Laplacian smoothing, let us look at equation
(8). Note that it can be shown (via the Gerschgorin’s Theorem [TB97]) that
the eigenvalues of the Laplace operator are in the interval [0, 2] and the smallest
eigenvalue λ1 = 0. Since λ ∈ [0, 2], the filter function f (λ) = (1 − 12 λ)m
is bounded by the unit interval [0, 1] and attains the maximum f (0) = 1 at
λ = 0. As m → ∞, all the terms in the right-hand side of equation (8) will
vanish except for the first, which is given by e1 x̃1 . Since e1 , the eigenvector
corresponding to the zero eigenvalue is a normalized, constant vector, we have
e1 = [ √1n √1n . . . √1n ]T . Now taking the y-component into consideration, we
get the limit point for Laplacian smoothing as √1n [x̃1 ỹ1 ]. Finally, noting that
x̃1 = eT T
1 · X and ỹ1 = e1 · Y , we conclude that the limit point of Laplacian
smoothing is the centroid of the set of original contour vertices.

1.5 Spectral analysis vs. discrete Fourier transform

The sinosoidal behavior of the eigenvectors of the 1D discrete Laplace operator
(see plots in Figure 4) leads one to believe that there must be a connection
between the discrete Fourier transform (DFT) and the spectral transform we
have defined so far. We now make that connection explicit. Typically, one
introduces the DFT in the context of Fourier series expansion. Given a discrete
signal X = [x1 x2 . . . xn ]T , its DFT is given by
n
1X
X̃(k) = X(k)e−i2π(k−1)(j−1)/n , k = 1, . . . , n.
n j=1
And the corresponding inverse DFT is given by
n
X
X(j) = X̃(k)ei2π(j−1)(k−1)/n , j = 1, . . . , n,
k=1

or
n
X
X= X̃(k)gk , where gk (j) = ei2π(j−1)(k−1)/n , k = 1, . . . , n.
k=1

We see that in the context of DFT, the signal X is expressed as a linear combina-
tion of the complex exponential DFT basis functions, the gk ’s. The coefficients
are given by the X̃(k)’s, which form the DFT of X. Fourier analysis, provided
by the DFT in the discrete setting, is one of the most important topics in math-
ematics and has wide-ranging applications in many scientific and engineering
disciplines. For a systematic study of the subject, we refer the reader to the
classic text by Bracewell [Bra99].
The connection we seek, between DFT and spectral analysis with respect to
the Laplace operator, is that the DFT basis functions, the gk ’s, form a set of
eigenvectors of the 1D discrete Laplace operator L, as defined in (3). A proof of
this fact can be found in Jain’s classic text on image processing [Jai89], where a
stronger claim with respect to circulant matrices was made. Note that a matrix
is circulant if each row can be obtained as a shift (with circular wrap-around)
of the previous row. It is clear that L is circulant.
Specifically, if we sort the eigenvalues of L in ascending order, then they are

πbk/2c
λk = 2 sin2 , k = 2, . . . , n. (9)
n
The first eigenvalue λ1 is always 0. We can see that every eigenvalue of L, except
for the first, and possibly the last, has a multiplicity of 2. That is, it corresponds
to an eigensubspace spanned by two eigenvectors. If we define the matrix G of
DFT basis as Gkj = ei2π(j−1)(k−1)/n , 1 ≤ k, j ≤ n, then the first column of G
is an eigenvector corresponding to λ1 and the k-th and (n + 2 − k)-th columns
of G are two eigenvectors corresponding to λk , for k = 2, . . . , n. The set of
eigenvectors of L is not unique. In particular, it has a set of real eigenvectors;
some of these eigenvectors are plotted in Figure 4.

1.6 Towards spectral mesh transform

Based on the above observation, one way to extend the notion of Fourier analysis
to the manifold or surface setting, where our signal will represent the geometry
of the surfaces, is to define appropriate discrete Laplace operators for meshes and
rely on the eigenvectors of the Laplace operators to perform Fourier analysis.
This was already observed in Taubin’s seminal paper [Tau95b].
To extend spectral analysis of 1D signals presented in Section 1.2 to surfaces
modeled by triangle meshes, we first need to extend the signal representation.
This is quite straightforward: any function defined on the mesh vertices can be
seen as a discrete mesh signal. Typically, we focus on the coordinate signal for
 (a) Original. (b) k = 300. (c) k = 200. (d) k = 100.

(e) k = 50. (f) k = 10. (g) k = 5. (h) k = 3.

Figure 8: Shape reconstruction based on spectral analysis using a typical mesh

Laplace operator, where k is the number of eigenvectors or spectral coefficients
used. The original model has 7,502 vertices and 15,000 faces.

the mesh, which, for a mesh with n vertices, is an n × 3 matrix whose columns
specify the x, y, and z coordinates of the mesh vertices.
The main task is to define an appropriate Laplace operator for the mesh.
Here a crucial difference to the classical DFTs is that while the DFT basis func-
tions are fixed as long as the length of the signal in question is determined, the
eigenvectors of a mesh Laplace operator would change with the mesh connec-
tivity and/or geometry. Formulations for the construction of appropriate mesh
Laplace operators will be the subjects of Sections 2 and 3.
Now with a mesh Laplace operator chosen, defining the spectral transform
of a mesh signal X with respect to the operator is exactly the same as the 1D
case for L in Section 1.2. Denote by e1 , e2 ,. . .,en the normalized eigenvectors of
the mesh Laplace operator, corresponding to eigenvalues λ1 ≤ λ2 ≤ . . . ≤ λn ,
and let E be the matrix whose columns are the eigenvectors. The vector of
spectral transform coefficients X̃ is obtained by X̃ = E T X. And for each i, we
obtain the spectral coefficient by x̃i = eTi · X via projection.
As a first visual example, we show in Figure 8 some results of spectral re-
construction, as define in (7), of a mesh model with progressively more spectral
coefficients added. As we can see, higher-frequency contents of the geometric
mesh signal manifest themselves as rough geometric features over the shape’s
surface; these features can be smoothed out by taking only the low-frequence
spectral coefficients in a reconstruction. A tolerance on such loss of geometric
features would lead to a JPEG-like compression of mesh geometry, as first pro-
posed by Karni and Gotsman [KG00a] in 2000. More applications using spectral
transforms of meshes will be given in Section 5.
2 Fourier analysis for meshes
The previous section introduced the idea of Fourier analysis applied to shapes,
with the example of a closed curve, for which the frequencies (sine waves) are
naturally obtained as the eigenvectors of the 1D discrete Laplacian. We now
study how to formalize the idea presented in the last subsection, i.e. porting
this setting to the case of arbitrary surfaces. Before diving into the heart of
the matter, we recall the definition of the Fourier transform, that is to say
the continuous version of the discrete signal processing framework presented in
Subsections 1.2 and 1.4.

2.1 Fourier analysis

As in Taubin’s article [Tau95b], we start by studying the case of a closed curve,
but staying in the continuous setting. Given a square-integrable periodic func-
tion f : x ∈ [0, 1] 7→ f (x), or a function f defined on a closed curve parameter-
ized by normalized arclength, it is well known that f can be expanded into an
infinite series of sines and cosines of increasing frequencies:
 0
X∞  H = 1
f (x) = f˜k H k (x) ; H 2k+1 = cos(2kπx) (10)
 2k+2
k=0 H = sin(2kπx)

where the coefficients f˜k of the decomposition are given by:

Z 1
˜
fk =< f, H >=k
f (x)H k (x)dx (11)
0

and where < ., . > denotes the inner product (i.e. the “dot product” for func-
tions defined on [0, 1]). See [Arv95] or [Lev06] for an introduction to func-
tional analysis. The ”Circle harmonics” basis H k is orthonormal with respect
to < ., . >: < H k , H k >= 1, < H k , H l >= 0 if k 6= l.
The set of coefficients f˜k (Equation 11) is called the Fourier Transform (FT)
of the function f . Given the coefficients f˜k , the function f can be reconstructed
by applying the inverse Fourier Transform FT−1 (Equation 10). Our goal is
now to explain the generalization of these notions to arbitrary manifolds. To do
so, we can consider the functions H k of the Fourier basis as the eigenfunctions
of −∂ 2 /∂x2 : the eigenfunctions H 2k+1 (resp. H 2k+2 ) are associated with the
eigenvalues (2kπ)2 :

∂ 2 H 2k+1 (x)
− = (2kπ)2 cos(2kπx) = (2kπ)2 H 2k+1 (x)
∂x2
This construction can be extended to arbitrary manifolds by considering the
generalization of the second derivative to arbitrary manifolds, i.e. the Laplace
operator and its variants, introduced below. Before studying the continuous
theory, we first do a step backward into the discrete setting, in which it is easier
Figure 9: The Fiedler vector gives a natural ordering of the nodes of a graph.
The displayed contours show that it naturally follows the shape of the dragon.

to grasp a geometric meaning of the eigenfunctions. In this setting, eigenfunc-

tions can be considered as orthogonal non-distorting 1D parameterizations of
the shape, as explained further.

2.2 The discrete setting: Graph Laplacians

Spectral graph theory was used for instance in [IL05] to compute an ordering
of vertices in a mesh that facilitates out-of-core processing. Such a natural
ordering can be derived from the Fiedler eigenvector of the Graph Laplacian.
The Graph Laplacian L = (ai,j ) is a matrix defined by:

ai,j = wi,j
P> 0 if (i, j) is an edge
ai,i = − j wi,j
ai,j = 0 otherwise

where the coefficients wi,j are weights associated to the edges of the graph. One
may use the uniform weighting wi,j = 1 or more elaborate weightings, computed
from the embedding of the graph. There are several variants of the Graph
Laplacian, the reader is referred to [ZvKDar] for an extensive survey. Among
these discrete Laplacians, the so-called Tutte Laplacian applies a normalization
in each row of L and is given by T = (ti,j ), where

ti,j = Pwi,j >0 if (i, j) is an edge

j wi,j
ti,i = −1
ti,j = 0 otherwise

The Tutte Laplacian was employed in the original work of Taubin [Tau95a],
among others [GGS03a, Kor03].
The first eigenvector of the Graph Laplacian is (1, 1 . . . 1) and its associ-
ated eigenvalue is 0. The second eigenvector is called the Fiedler vector and
has interesting properties, making it a good permutation vector for numerical
computations [Fie73, Fie75]. It has many possible applications, such as finding
natural vertices ordering for streaming meshes [IL05]. Figure 9 shows what it
looks like for a snake-like mesh (it naturally follows the shape of the mesh).
More insight on the Fiedler vector is given by the following alternative def-
inition. The Fiedler vector u = (u1 . . . un ) is the solution of the following con-
strained minimization problem:

Minimize: F (u) = ut Lu = i,j wi,j (ui − uj )2

P

P P 2 (12)
Subject to: i ui = 0 and i ui = 1

In other words, given a graph embedded in some space, and supposing that
the edge weight wi,j corresponds to the lengths of the edges in that space, the
Fielder vector (u1 . . . un ) defines a (1-dimensional) embedding of the graph on
a line that tries to respect the edge lengths of the graph.
This naturally leads to the question of whether embedding in higher-dimen-
sional spaces can be computed (for instance, computing a 2-dimensional embed-
ding of a surface corresponds the the classical surface parameterization prob-
lem). This general problem is well known by the automatic learning research
community as a Manifold learning problem, also called dimension reduction.
One of the problems in manifold learning is extracting from a set of input
(e.g. a set of images of the same object) some meaningful parameters (e.g.
camera orientation and lighting conditions), and sort these images with respect
to these parameters. From an abstract point of view, the images leave in a high-
dimensional space (the dimension corresponds to the number of pixels of the
images), and one tries to parameterize this image space. The first step constructs
a graph, by connecting each sample to its nearest neighbors, according to some
distance function. Then, different classes of methods have been defined, we
quickly review the most popular ones:
Local Linear Embedding [RS00a] tries to create an embedding that best
approximates the barycentric coordinates of each vertex relative to its neighbors.
In a certain sense, Floater’s Shape Preserving Parameterization (see [FH04]) is
a particular case of this approach.
Isomap [TdSL00] computes the geodesic distances between each pair of ver-
tex in the graph, and then uses MDS (multidimensional scaling) [You85] to com-
pute an embedding that best approximates these distances. Multidimensional
scaling simply minimizes an objective function that measures the deviation be-
tween the geodesic distances in the initial space and the Euclidean distances
in the embedding space (GDD for Geodesic Distance Deviation), by computing
the eigenvectors of the matrix D = (di,j ) where di,j denotes the geodesic dis-
tance between vertex i and vertex j. This is a multivariate version of Equation
12, that characterizes the Fiedler vector (in the univariate setting). Isomaps
and Multidimensional scaling were used to define parameterization algorithms
in [ZKK02], and more recently in the ISO-charts method [ZSGS04], used in Mi-
crosoft’s DirectX combined with the packing algorithm presented in [LPM02].
 At that point, we understand that the eigenvectors play an important role in
determining meaningful parameters. Just think about the simple linear case: in
PCA (principal component analysis), the eigenvectors of the covariance matrix
characterize the most appropriate hyperplane on which the data should be pro-
jected. In dimension reduction, we seek for eigenvectors that will fit non-linear
features. For instance, in MDS, these eigenvectors are computed in a way that
makes the embedding space mimic the global metric structure of the surface,
captured by the matrix D = (di,j ) of all geodesic distances between all pairs of
vertices in the graph.
Instead of using the dense matrix D, methods based on the Graph Laplacian
only use local neighborhoods (one-ring neighborhoods). As a consequence, the
used matrix is sparse, and extracting its eigenvectors requires lighter computa-
tions. Note that since the Graph Laplacian is a symmetric matrix, its eigenvec-
tors are orthogonal, and can be used as a vector basis to represent functions.
This was used in [KG00b] to define a compact encoding of mesh geometry. The
basic idea consists in encoding the topology of the mesh together with the coef-
ficients that define the geometry projected onto the basis of eigenvectors. The
decoder simply recomputes the basis of eigenvectors and multiplies them with
the coefficients stored in the file. A survey of spectral geometry compression and
its links with graph partitioning is given in [Got03]. Spectral graph theory also
enables to exhibit ways of defining valid graph embeddings. For instance, Colin-
de-verdière’s number [dV90] was used in [GGS03b] to construct valid spherical
embeddings of genus 0 meshes. Other methods that use spectral graph theory
to compute graph embeddings are reviewed in [Kor02]. Spectral graph theory
can also be used to compute topological invariants (e.g. Betti numbers), as
explained in [Fei96]. As can be seen from this short review of spectral graph
theory, the eigenvectors and eigenvalues of the graph Laplacian contain both
geometric and topological information.
However, as explained in [ZSGS04], only using the connectivity of the graph
may lead to highly distorted mappings. Methods based on MDS solve this issue
by considering the matrix D of the geodesic distances between all possible pairs
of vertices. However, we think that it is also possible to inject more geometry
in the Graph Laplacian approach, and understand how the global geometry and
topology of the shape may emerge from the interaction of local neighborhoods.
This typically refers to notions from the continuous setting, i.e. functional
analysis and operators. The next section shows its link with the Laplace-
Beltrami operator, that appears in the wave equation (Helmholtz’s equation).
We will also exhibit the link between the so-called stationary waves and spectral
graph theory.
2.3 The Continuous Setting: Laplace Beltrami
The Laplace operator (or Laplacian) plays a fundamental role in physics and
mathematics. In Rn , it is defined as the divergence of the gradient:
X ∂2
∆ = div grad = ∇.∇ =
i
∂x2i

Intuitively, the Laplacian generalizes the second order derivative to higher di-
mensions, and is a characteristic of the irregularity of a function as ∆f (P )
measures the difference between f (P ) and its average in a small neighborhood
of P .
Generalizing the Laplacian to curved surfaces require complex calculations.
These calculations can be simplified by a mathematical tool named exterior cal-
culus (EC) 1 . EC is a coordinate free geometric calculus where functions are
considered as abstract mathematical objects on which operators act. To use
these functions, we cannot avoid instantiating them in some coordinate frames.
However, most calculations are simplified thanks to higher-level considerations.
For instance, the divergence and gradient are known to be coordinate free oper-
ators, but are usually defined through coordinates. EC generalizes the gradient
by d and divergence by δ, which are built independently of any coordinate frame.
Using EC, the definition of the Laplacian can be generalized to functions
defined over a manifold S with metric g, and is then called the Laplace-Beltrami
operator:
X 1 ∂ p ∂
∆ = div grad = δd = p |g|
i
|g| ∂xi ∂x i
p
where |g| denotes the determinant of g. The additional term |g| can be inter-
pretedp as a local ”scale” factor since the local area element dA on S is given by
dA = |g|dx1 ∧ dx2 .
Finally, for the sake of completeness, we can mention that the Laplacian
can be extended to k-forms and is then called the Laplace-de Rham operator
defined by ∆ = δd + dδ. Note that for functions (i.e. 0-forms), the second term
dδ vanishes and the first term δd corresponds to the previous definition.
Since we aim at constructing a function basis, we need some notions from
functional analysis, quickly reviewed here. A similar review in the context of
light simulation is given in [Arv95].

2.3.1 Functional analysis

In a way similar to what is done for vector spaces, we need to introduce a dot
product (or inner product) to be able to define function bases and project func-
tions onto those bases. This corresponds to the notion of Hilbert space, outlined
1 To our knowledge, besides Hodge duality used to compute minimal surfaces [PP93],

one of the first uses of EC in geometry processing [GY02] applied some of the fundamental
notions involved in the proof of Poincaré’s conjecture to global conformal parameterization.
More recently, a Siggraph course was given by Schroeder et al. [SGD05], making these notions
usable by a wider community.
below.

Hilbert spaces

Given a surface S, let X denote the space of real-valued functions defined on

S. Given a norm k.k, the function space X is said to be complete with respect
to the norm if Cauchy sequences converge in X, where a Cauchy sequence is a
sequence of functions f1 , f2 , . . . such that limn,m−>∞ kfn −fm k = 0. A complete
vector space is called a Banach space.
The space√ X is called a Hilbert space in the specific case of the norm defined
by kf k = < f, f >, where < ., . > denotes an Rinner product. A possible
definition of an inner product is given by < f, g >= S f (x)g(x)dx, which yields
the L2 norm.
One of the interesting features provided by this additional level of structure
is the ability to define function bases and project onto these bases using the
innerP product. Using a function basis (Φi ), a function f will be defined by
f = αi Φi . Similarly to the definition in geometry, a function basis (Φi ) is
orthonormal if kΦi k = 1 for all i and < Φi , Φj >= 0 for all i 6= j. Still following
the analogy with geometry, given the function f , one can easily retrieve its
coordinates αi with respect to an orthonormal basis (Φi ) by projecting f onto
the basis, i.e. αi =< f, Φi >.
Operators
We now give the basic definitions related with operators. Simply put, an op-
erator is a function of functions (i.e. from X to X). An operator L applied
to a function f ∈ X is denoted by Lf , and results in another function in X.
An operator L is said to be linear if L(λf ) = λLf for all f ∈ X, λ ∈ R. An
eigenfunction of the operator L is a function f such that Lf = λf . The scalar
λ is an eigenvalue of L. In other words, the effect of applying the operator to
one of its eigenfunctions means simply scaling the function by the scalar λ.
A linear operator L is said to be Hermitian (or with Hermitian symmetry)2
if < Lf, g >=< f, Lg > for each f, g ∈ X. An important property of Hermitian
operators is that their eigenfunctions associated to different eigenvalues have
real eigenvalues and are orthogonal. This latter property can be easily proven
as follows, by considering two eigenfunctions f, g associated with the different
eigenvalues λ, µ respectively:

< Lf, g > = < f, Lg >

< λf, g > = < f, µg >
λ < f, g > = µ < f, g >

which gives the result (< f, g >= 0) since λ 6= µ.

As a consequence, considering the eigenfunctions of an Hermitian operator
is a possible way of defining an orthonormal function basis associated to a given
2 the general definition of Hermitian operators concerns complex-valued functions, we only

consider here real-valued functions.

Figure 10: Left: In the late 1700’s, the physicist Ernst Chladni was amazed
by the patterns formed by sand on vibrating metal plates. Right: numerical
simulations obtained with a discretized Laplacian.

function space X. The next section shows this method applied to the Laplace-
Beltrami operator. Before entering the heart of the matter, we will first consider
the historical perspective.

2.3.2 Chladni plates

In 1787, the physicist Ernst Chladni published the book entitled “Discoveries
Concerning the Theories of Music”. In this book, he reports his observations ob-
tained when putting a thin metal plate into vibration using a bow, and spreading
sand over it. Sand accumulates in certain zones, forming surprisingly complex
patterns (see Figure 10).
This behavior can be explained by the theory of stationary waves. When the
metal plate vibrates, some zones remain still, and sand naturally concentrates in
these zones. This behavior can be modeled as follows, by the spatial component
of Helmholtz’s wave propagation equation:

∆f = λf (13)
In this equation, ∆ denotes the Laplace-Beltrami operator on the considered
object. In Cartesian 2D space, ∆ = ∂ 2 /∂x2 + ∂ 2 /∂y 2 . We are seeking for the
eigenfunctions of this operator. To better understand the meaning of this equa-
tion, let us first consider a vibrating circle. This corresponds to the univariate
case on the interval [0, 2π] with cyclic boundary conditions (i.e. f (0) = f (2π)).
In this setting, the Laplace-Beltrami operator simply corresponds to the second
order derivative. Recalling that sin(ωx)00 = ω 2 sin(ωx), the eigenfunctions are
simply sin(N x), cos(N x) and the constant function, where N is an integer. Note
that the so-constructed function basis is the one used in Fourier analysis.
From the spectrum of the Laplace-Beltrami operator, it is well known that
one can extract the area of S, the length of its border and its genus. This
leads to the question asked by Kac in 1966: Can we hear the shape of a drum
Figure 11: Contours of the first eigenfunctions. Note how the protrusions
and symmetries are captured. The eigenfunctions combine the geometric and
topological information contained in the shape signal.

? [Kac66]. The answer to this question is “no”: one can find drums that have
the same spectrum although they do not have the same shape [Cip93] (they are
then referred to as isospectral shapes). However, the spectrum contains much
information, which led to the idea of using it as a signature for shape matching
and classification, as explained in the “shape DNA” approach [RWP05a].
We are now going now to take a look at the eigenfunctions. Mathematicians
mostly studied bounds and convergence of the spectrum. However, some results
are known about the geometry of the eigenfunctions [JNT]. More precisely,
we are interested in the so-called nodal sets, defined to be the zero-set of an
eigenfunction. Intuitively, they correspond to the locations that do not move on
a Chladni plate, where sand accumulates (see Figure 10). A nodal set partitions
the surface into a set of nodal domains of constant sign. The nodal sets and
nodal domains are characterized by the following theorems:
1. the n-th eigenfunction has at most n nodal domains
2. the nodal sets are curves intersecting at constant angles
Besides their orthogonality, these properties make eigenfunction an interesting
choice for function bases. Theorem 1 exhibits their multi-resolution nature,
and from theorem 2, one can suppose that they are strongly linked with the
geometry of the shape. Note also that these theorems explain the appearance
of Chladni plates. This may also explain the very interesting re-meshing results
obtained by Dong et. al [DBG+ 05], that use a Morse-smale decomposition of
one of the eigenfunctions.
In the case of simple objects, a closed form of the eigenfunctions can be
derived. This made it possible to retrieve the patterns observed by Chladni in
the case of a square and a circle. For curved geometry, Chladni could not make
the experiment, since sand would not remain in the nodal set. However, one
can still study the eigenfunctions. For instance, on a sphere, the eigenfunction
correspond to spherical harmonics (see e.g. [JNT]), often used in computer
graphics to represent functions defined on the sphere (such as radiance fields or
Figure 12: Contours of the 4th eigenfunction, computed from the Graph Lapla-
cian (left) and cotangent weights (right) on an irregular mesh.

Bidirectional Reflectance Distribution Functions). In other words, on a sphere,

the eigenfunctions of the Laplace-Beltrami operator define an interesting hierar-
chical function basis. One can now wonder whether the same approach could be
used to create function bases on more complex geometries. In the general case,
a closed form cannot be derived, and one needs to use a numerical approach, as
explained in the next section.

3 Discretizing the Laplace operator

The previous section mentioned two different settings for spectral analysis :
• the discrete setting is concerned with graphs, matrices and vectors;

• the continuous setting is concerned with manifolds, operators and func-

tions.
The continuous setting belongs to the realm of differential geometry, a pow-
erful formalism for studying relations between manifolds and functions defined
over them. However, computer science can only manipulate discrete quantities.
Therefore, a practical implementation of the continuous setting requires a dis-
cretization. This section shows how to convert the continuous setting into a
discrete counterpart using the standard Finite Element Modeling technique. In
Geometry Processing, another approach consists in trying to define a discrete
setting that mimics the property of the continous theory. As such, we will also
show how spectral geometry processing can be derived using Discrete Exterior
Calculus (DEC).
The eigenfunctions and eigenvalues of the Laplacian on a (manifold) surface
S, are all the pairs (H k , λk ) that satisfy:

−∆H k = λk H k (14)
Figure 13: Some functions of the Manifold Harmonic Basis (MHB) on the Gar-
goyle dataset

The “−” sign is here required for the eigenvalues to be positive. On a closed
curve, the eigenfunctions of the Laplace operator define the function basis (sines
and cosines) of Fourier analysis, as recalled in the previous section. On a square,
they correspond to the function basis of the DCT (Discrete Cosine Transform),
used for instance by the JPEG image format. Finally, the eigenfunctions of the
Laplace-Beltrami operator on a sphere define the Spherical Harmonics basis.
In these three simple cases, two reasons make the eigenfunctions a function
basis suitable for spectral analysis of manifolds:
1. Because the Laplacian is symmetric (< ∆f, g >=< f, ∆g >), its eigen-
functions are orthogonal, so it is extremely simple to project a function
onto this basis, i.e. to apply a Fourier-like transform to the function.
2. For physicists, the eigenproblem (Equation 14) is called the Helmoltz equa-
tion, and its solutions H k are stationary waves. This means that √ the H k
are functions of constant wavelength (or spatial frequency) ωk = λk .
Hence, using the eigenfunctions of the Laplacian to construct a function ba-
sis on a manifold is a natural way to extend the usual spectral analysis to this
manifold. In our case, the manifold is a mesh, so we need to port this construc-
tion to the discrete setting. The first idea that may come to the mind is to apply
spectral analysis to a discrete Laplacian matrix (e.g. the cotan weights). How-
ever, the discrete Laplacian is not a symmetric matrix (the denominator of the
ai,j coefficient is the area of vertex i0 s neighborhood, that does not necessarily
correspond to the area of vertex j’s neighborhood). Therefore, we lose the sym-
metry of the Laplacian and the orthogonality of its eigenvectors. This makes it
difficult to project functions onto the basis. For this reason, we will clarify the
relations between the continuous setting (with functions and operators), and
the discrete one (with vectors and matrices) in the next section.
In this section, we present two ways of discretizing the Laplace operator on
a mesh. The first approach is based on Finite Element Modeling (FEM) such
as done in [WBH+ 07], and converge to the continuous Laplacian under certain
conditions as explained in [HPW06] and [AFW06]. Reuter et al. [RWP05b]
also use FEM to compute the spectrum (i.e. the eigenvalues) of a mesh, which
provides a signature for shape classification. The cotan weights were also used
in [DBG+ 06b] to compute an eigenfunction to steer a quad-remeshing process.
The cotan weights alone are still dependent on the sampling as shown in Fig-
ure 18-B, so they are usually weighted by the one ring or dual cell area of each
vertex, which makes them loose their symmetry. As a consequence, they are
improper for spectral analysis (18-C). An empirical symmetrization was pro-
posed in [Lev06] (see Figure 18-D). The FEM formulation enables to preserve
the symmetry property of the continuous counterpart of the Laplace operator.
It is also possible to derive a symmetric Laplacian by using the DEC formu-
lation (Discrete Exterior Calculus). Then symmetry is recovered by expressing
the operator in a proper basis. This ensures that its eigenfunctions are both ge-
ometry aware and orthogonal (Figure 18-E). Note that a recent important proof
[WMKG07] states that a ”perfect” discrete Laplacian that satisfies all the prop-
erties of the continuous one cannot exist on general meshes. This explains the
large number of definitions for a discrete Laplacian, depending on the desired
properties.

3.1 The FEM Laplacian

In this subsection, we explain the Finite Element Modeling (FEM) formulation
of the discrete Laplacian. The Discrete Exterior Calculus (DEC)formululation,
presented during the course, is described in the next subsection. We have chosen
to include in these course nodes the full derivations for the FEM Laplacian for
the sake of completeness.
To setup the finite element formulation, we first need to define a set of basis
functions used to express the solutions, and a set of test functions onto which
the eigenproblem (Equation 14) will be projected. As it is often done in FEM,
we choose for both basis and test functions the same set Φi (i = 1 . . . n). We use
the “hat” functions (also called P1 ), that are piecewise-linear on the triangles,
and that are such that Φi (i) = 1 and Φi (j) = 0 if i 6= j. Geometrically,
Φi corresponds to the barycentric coordinate associated with vertex i on each
triangle containing i. Solving the finite element formulation of Equation
Pn 14
relative to the Φi ’s means looking for functions of the form: H k = i=1 Hik Φi
which satisfy Equation 14 in projection on the Φj ’s:

∀j, < −∆H k , Φj >= λk < H k , Φj >

or in matrix form:
−Qhk = λBhk (15)
where Qi,j =< ∆Φi , Φj >, Bi,j =< Φi , Φj > and where hk denotes the vector
[H1k , . . . Hnk ]. The matrix Q is called the stiffness matrix, and B the mass matrix.
This appendix derives the expressions for the coefficients of the stiffness
matrix Q and the mass matrix B. To do so, we start by parameterizing a triangle
t = (i, j, k) by the barycentric coordinates (or hat functions) Φi and Φj of a point
P ∈ t relative to vertices i and j. This allows to write P = k + Φi ej − Φj ei
(Figure 14). This yields an area element dA(P ) = ei ∧ ej dΦi dΦj = 2|t|dΦi dΦj ,
Figure 14: Notations for matrix coefficients computation. Vectors are typed in
bold letters (ei )

where |t| is the area of t, so we get the integral:

Z Z 1 Z 1−Φi
Φi Φj dA = 2|t| Φi Φj dΦi dΦj =
P ∈t Φi =0 Φj =0

1  
|t|
Z
2 1 2 1
|t| Φi (1 − Φi ) dΦi = |t| − + =
Φi =0 2 3 4 12
which we sum up on the 2 triangles sharing (i, j) to get Bi,j = (|t| + |t0 |/12. We
get the diagonal terms by:
Z Z 1 Z 1−Φi
2
Φi dA = 2|t| Φ2i dΦj =
P ∈t Φi =0 Φj =0

1  
|t|
Z
1 1
2|t| Φ2i (1 − Φi )dΦi = 2|t| − =
Φi =0 3 4 6
which
P are summed up over the set St(i) of triangles containing i to get Bi,i =
( t∈St(i) |t|)/6.
To compute the coefficients of the stiffness matrix Q, we use the fact that d
and δ are adjoint to get the more symmetric expression:
Z
Qi,j =< ∆Φi , Φj >=< δdΦi , Φj >=< dΦi , dΦj >= ∇Φi .∇Φj
S

In t, the gradients of barycentric coordinates are the constants :

−e⊥i ei .ej
∇Φi = ∇Φi .∇Φj =
2|t| 4|t|2
Where e⊥
i denotes ei rotated by π/2 around t’s normal. By integrating on t we
get:
||ei ||.||ej || cos(βij )
Z
ei .ej cot(βij )
∇Φi .∇Φj dA = = =
t 4|t| 2||ei ||.||ej || sin(β ij ) 2
Summing these expressions, the coefficients of the stiffness matrix Q are
given by:
X X e2i
Qi,i = ∇Φi .∇Φi =
4|t|
t∈St(i) t∈St(i)
Z
1 0
∇Φi .∇Φj =


Qi,j = cot(βij ) + cot(βij
t∪t0 2
Note that this expression is equivalent to the numerator of the classical cotan
weights.
0


Qi,j = cotan(βP i,j ) + cotan(β i,j ) /2
Qi,i = − j Qi,j
(16)
0

Bi,j = (|t| P + |t |) /12
Bi,i = ( t∈St(i) |t|)/6
where t, t0 are the two triangles that share the edge (i, j), |t| and |t0 | denote
0
their areas, βi,j , βi,j denote the two angles opposite to the edge (i, j) in t and
0
t , and St(i) denotes the set of triangles incident to i.
To simplify the computations, a common practice of FEM consists in replac-
ing this equation with an approximation:

−Qhk = λDhk or − D−1 Qhk = λhk

(17)

where the mass matrix B is replaced with a diagonal matrix D called the lumped
mass matrix, and defined by:
X X
Di,i = Bi,j = ( |t|)/3. (18)
j t∈St(i)

Note that D is non-degenerate (as long as mesh triangles have non-zero areas).
FEM researchers [Pra99] explain that besides simplifying the computations this
approximation fortuitously improves the accuracy of the result, due to a can-
cellation of errors, as pointed out in [Dye06]. The practical solution mechanism
to solve Equation 17 will be explained further in the section about numerics.

Remark: The matrix D−1 Q in (Equation 17) exactly corresponds to the usual
discrete Laplacian (cotan weights). Hence, in addition to direct derivation of
triangle energies [PP93] or averaged differential values [MDSB03], the discrete
Laplacian can be derived from a lumped-mass FEM formulation.
Figure 15: Reconstructions obtained with an increasing number of eigenfunc-
tions.

As will be seen further, the FEM formulation and associated inner product
will help us understand why the orthogonality of the eigenvectors seems to be
lost (since D−1 Q is not symmetric), and how to retrieve it.
Without entering the details, we mention some interesting features and de-
grees of freedom obtained with the FEM formulation. The function basis Φ onto
the eigenfunction problem is projected can be chosen. One can use piecewise
polynomial functions. Figure 16 shows how the eigenfunctions look like with
the usual piecewise linear basis (also called P1 in FEM parlance) and degree 3
polynomial basis (P3). Degree 3 triangles are defined by 10 values (1 value per
vertex, two values per edge and a central value). As can be seen, more complex
function shapes can be obtained (here displayed using a fragment shader). It is
also worth mentioning that the boundaries can be constrained in two different
ways (see Figure 17). With Dirichlet boundary conditions, the value of the func-
tion is constant on the boundary (contour lines are parallel to the boundary).
With Neumann boundary condistions, the gradient of the eigenfunction is par-
allel to the boundary (therefore, contour lines are orthogonal to the boundary).

3.2 The DEC Laplacian

For a complete introduction to DEC we refer the reader to [DKT05], [Hir03] and
to [AFW06] for proofs of convergence. We quickly introduce the few notions and
notations that we are using to define the inner product < ., . > and generalized
second-order derivative (i.e. Laplacian operator).
A k-simplex sk is the geometric span of k + 1 points. For instance, 0, 1 and
2-simplices are points, edges and triangles respectively. In our context, a mesh
can be defined as a 2-dimensional simplicial complex S, i.e. a collection of nk
k-simplices (k = 0, 1, 2), with some conditions to make it manifold. A discrete
k-form ω k on S is given by a real value ω k (sk ) associated with each oriented k-
simplex (that corresponds to the integral of a smooth k-form over the simplex).
The set Ωk (S) of k-forms on S is a vector-space of dimension nk . With a proper
numbering of the k-simplices, ω k can be assimilated to a vector of size nk , and
linear operators from Ωk (S) to Ωl (S) can be assimilated to (nk , nl ) matrices.
The exterior derivative dk : Ωk (S) → Ωk+1 (S) is defined by the signed
adjacency matrix: (dk )sk ,sk+1 = ±1 if sk belongs to the boundary of sk+1 , with
the sign depending on their respective orientations.
Figure 16: Contour lines of eigenfunctions obtained with a linear basis (left)
and a cubic basis (right)

DEC provides Ωk (S) with a L2 inner product:

< ω1k , ω2k >= (ω1k )T ?k ω2k (19)
where ?k is the so-called Hodge star. As a matrix, the Hodge star is diagonal
with elements |s∗k |/|sk | where s∗k denotes the circumcentric dual of simplex sk ,
and |.| is the simplex volume. In particular, for vertices, edges and triangles:
|e∗ |
(?0 )vv = |v ∗ | ; (?1 )ee = = cot βe + cot βe0 ; (?2 )tt = |t|−1
|e|
where βe and βe0 denote the two angles opposite to e.
The codifferential δk : Ωk (S) → Ωk−1 (S) is the adjoint of the exterior deriva-
tive for the inner product:
< δk ω1k , ω2k−1 >=< ω1k , dk−1 ω2k−1 >
which yields δk = − ?−1 T
k−1 dk−1 ?k .
Finally the Laplace de Rham operator on k-forms is given by: ∆k = dk−1 δk +
δk+1 dk . In this paper, we are only interested in the Laplacian ∆ = ∆0 , i.e. the
Laplace de Rham operator for 0-forms . Since δ0 and d2 are undefined (zero by
convention), ∆ = δ1 d0 = − ?−1 T
0 d1 ?1 d0 0 and its coefficients are given by:
cot(βij ) + cot(βij ) X
∆ij = − ∗ ; ∆ii = − ∆ij
|vi | j
0
For surfaces with borders, if the edge ij is on the border, the term cot(βij )
∗
vanishes and the dual cell vi is cropped by the border. This matches the FEM
formulation with Neumann boundary conditions (not detailed here).
Figure 17: Contour lines of eigenfunctions obtained with Dirichlet boundary
conditions (left) and Neumann boundary conditions (right)
Figure 18: Filtering an irregularly sampled surface (twice denser on the
right half) with different discrete laplacians. Combinatorial Laplacian (A),
0
unweighted cotan cot(βij ) + cot(βij ) (B) and symmetrized weighted cotan
0
(cot(βij )+cot(βij ))/(Ai +Aj ) (D) produce deformed results (right leg is bigger).
0
Weighted cotan (cot(βij )+cot(βij ))/Ai is not symmetric which does not allow for
0
p
correct reconstruction (C). Only symmetric weights (cot(βij )+cot(βij ))/ Ai Aj
is fully mesh-independent (E).

Remark: The matrix ∆ corresponds to the standard discrete Laplacian,

except for the sign. The sign difference comes from the distinction between
Laplace-Beltrami and Laplace de Rham operators.
The so-defined Laplacian ∆ apparently looses the symmetry of its contin-
uous counterpart (∆ij 6= ∆ji ). This makes the eigenfunction basis no longer
orthonormal, which is problematic for our spectral processing purposes (Fig-
ure 18-C). To recover symmetry, consider the canonical basis (φi ) of 0-forms:
φi = 1 on vertex i and φi = 0 on other vertices. This basis is orthogo-
nal but not normal with respect to the inner product defined in Equation 19
(< φi , φi >= (φi )T ?0 φi 6= 1). However, since the Hodge star ?0 is a diagonal
matrix, one can easily normalize (φi ) as follows:
−1/2
φ̄i = ?0 φi
¯ is symmetric, and its coef-
In this orthonormalized basis (φ̄i ), the Laplacian ∆
ficients are given by:
0
cot βij + cot βij
¯ = ?−1/2 ∆?1/2
∆ ; ¯ ij = −
∆ (20)
0 0 q
|vi∗ ||vj∗ |

4 Computing eigenfunctions
Now that we have seen two equivalent discretizations of the Laplacien, namely
FEM (Finite Elements Modeling) and DEC (Discrete Exterior Calculus), we
will now focus on the problem of computing the eigenfunctions in practice. An
implementation of the algorithm is available from the WIKI of the course (see
web reference at the beginning of this document).
Computing the eigenfunctions means solving for the eigenvalues λk and
¯
eigenvectors H̄ k for the symmetric positive semi-definite matrix ∆:
¯ H̄ k = λk H̄ k
∆

However, eigenvalues and eigenvectors computations are known to be ex-

tremely computationally intensive. To reduce the computation time, Karni et
al. [KG00c] partition the mesh into smaller charts, and [DBG+ 06b] use mul-
tiresolution techniques. In our case, we need to compute multiple eigenvectors
(typically a few thousands). This is known to be currently impossible for meshes
with more than a few thousand vertices [WK05]. In this section, we show how
this limit can be overcome by several orders of magnitude.
To compute the solutions of a large sparse eigenproblems, several iterative
algorithms exist. The publicly available library ARPACK (used in [DBG+ 06b])
provides an efficient implementation of the Arnoldi method. Yet, two charac-
teristics of eigenproblem solvers hinder us from using them directly to compute
the MHB for surfaces with more than a few thousand vertices:
• first of all, we are interested in the lower frequencies, i.e. eigenvectors with
2
associated eigenvalues lower than ωm . Unfortunately, iterative solvers
performs much better for the other end of the spectrum. This can be
explained in terms of filtering as lower frequencies correspond to higher
powers of the smoothing kernel, which may have a poor condition number;
• secondly, we need to compute a large number of eigenvectors (typically
a thousand), and it is well known that computation time is superlinear
in the number of requested eigenpairs. In addition, if the surface is large
(millions of vertices), the MHB does not fit in system RAM.

4.1 Band-by-band computation of the MHB

We address both issues by applying spectral transforms to the eigenproblem.
To get the eigenvectors of a spectral band centered around a value λS , we start
by shifting the spectrum by λS , by replacing ∆ ¯ with ∆
¯ − λS Id. Then, we can
swap the spectrum by inverting this matrix in ∆ = (∆SI ¯ − λS Id)−1 . This is
called the Shift-Invert spectral transform, and the new eigenproblem to solve is
given by:
∆SI H̄ k = µk H̄ k
¯ and ∆SI have the same eigenvectors, and that their
It is easy to check that ∆
eigenvalues are related by λk = λS + 1/µk . Applying an iterative solver to ∆SI
will return the high end of the spectrum (largest µ’s), corresponding to a band
of eigenvalues of ∆¯ centered around λS . It is then possible to split the MHB
computation into multiple bands, and obtain a computation time that is linear
in the number of computed eigenpairs. In addition, if the MHB does not fit in
RAM, each band can be streamed into a file. This gives the following algorithm:

(1) λS ← 0 ; λlast ← 0
2
(2) while(λlast < ωm )
(3) compute an inverse ∆SI of (∆ ¯ − λS Id)
(4) find the 50 first eigenpairs (H̄ k , µk ) of ∆SI
(5) for k = 1 to 50
(6) λk ← λS + 1/µk
(7) if (λk > λlast ) write(H̄ k , λk )
(8) end //f or
(9) λS ← max(λk ) + 0.4(max(λk ) − min(λk ))
(10) λlast ← max(λk )
(11) end //while

Before calling the eigen solver, we pre-compute ∆SI with a sparse direct
solver (Line 3). Note that ∆ ¯ − λS Id may be singular. This is not a problem
since the spectral transform still works with an indefinite factorization. The
factorized ∆¯ − λS Id is used in the inner loop of the eigen solver (Line 4). To
¯
factorize ∆−λ S Id, we used sparse direct solvers (TAUCS, SUPERLU). For large
models (millions of vertices), we used the sparse OOC (out-of-core) symmetric
indefinite factorization [MIT06] implemented in the future release of TAUCS,
kindly provided by S. Toledo. We then recover the λ’s from the µ’s (Line 6)
and stream-write the new eigenpairs into a file (Line 7). Since the eigenvalues
are centered around the shift λS , the shift for the next band is given by the last
computed eigenvalue plus slightly less than half the bandwidth to ensure that
the bands overlap and that we are not missing any eigenvalue (Line 9). If the
bands do not overlap, we recompute a larger band until they do.
Note that this is different from the shift-invert spectral transform imple-
mented by ARPACK, dedicated to iterative solvers. Ours makes use of the
factorization of the matrix, resulting in much better performances.

5 Applications
The eigendecomposition of a discrete mesh Laplace operator provides a set of
eigenvalues and eigenvectors, which can be directly used by an application to
accomplish different tasks. Moreover, the eigenvectors can also be used as a
basis onto which a signal defined on a triangle mesh is projected. The resulting
spectral transform coefficients can be further analyzed or manipulated. Here we
discuss a subset of the applications which utilize the spectral transform or the
eigenvectors of mesh Laplace or more general linear mesh operators.

5.1 Use of eigenvectors

Eigenvectors are typically used to obtain an embedding of the input shape
in the spectral domain, which we call a spectral embedding. After obtain-
ing the eigendecomposition of a specific operator, the coordinates of vertex
Figure 19: Compared parameterization results obtained with LSCM (left), spec-
tral parameterization (center) and ABF++ (right).

i in a k-dimensional spectral embedding are given by the i-th row of matrix

Vk = [v1 , . . . , vk ], where v1 , . . . , vk are the first k eigenvectors from the spec-
trum (possibly after scaling). Whether the eigenvectors should be in ascending
or descending order of eigenvalues depends on the operator that is being used.
The choice of k also varies between applications. For planar mesh parameteriza-
tion, k is typically 2, while in other applications, a single eigenvector possessing
desirable properties may be selected for a task.
In computer vision and machine learning, spectral methods usually employ a
different operator, the so-called affinity matrix [SM00, Wei99]. Each entry Wij
of an affinity matrix W represents a numerical relation, the affinity, between
two data points i and j, e.g., pixels in an image, vertices in a mesh, or two
face models in the context of face recognition. Note that the affinity matrix
differs from the Laplacian in that affinities between all data pairs are defined.
Therefore this matrix is not sparse in general. In practice, this non-sparse
structure implies more memory requirements and more expensive computations.
The use of affinity matrices for spectral mesh processing has received wide appeal
in computer graphics as well, e.g., for mesh segmentation.

5.1.1 Parameterization and remeshing

In the context of mesh parameterization, spectral methods have the interest-
ing property of connecting local entities in a way that lets a global behavior
emerge. This property can be used to compute a good base complex [DBG+ 06a],
[HZM+ 08], or to directly define a parameterization [ZSGS04], [MTAD08], as ex-
plained below. Figure 19 shows that Mullen et.al’s spectral parameterization
achieves a result comparable to ABF++. The interesting point is that it uses a
simple and elegant formulation of the problem (just compute the Fiedler vector
Figure 20: Spectral Surface Quadrangulation first computes a Laplace eigen-
function (A), then extracts its Morse complex (B), smoothes it (C) and uses it
to partition the mesh into quads, that can be parameterized (D).

of the LSCM matrix).

Besides corresponding to a “Fourier basis on meshes”, eigenfunctions can be
used to define parameterizations. For instance, in the context of data analysis,
LLE (Local Linear Embedding) [RS00b] computes an embedding of a graph
that preserves metric relations in local neighborhoods. Similarly, the MDS
method (multidimensional scaling) [You85] computes an embedding that best
approximates all the geodesic distances between the vertices of a graph. Mul-
tidimensional scaling simply minimizes an objective function that measures the
deviation between the geodesic distances in the initial space and the Euclidean
distances in the embedding space (GDD for Geodesic Distance Deviation), by
computing the eigenvectors of the matrix D = (di,j ) where di,j denotes the
geodesic distance between vertex i and vertex j. Isomaps and Multidimensional
scaling were used to define parameterization algorithms in [ZKK02], and more
recently in the ISO-charts method [ZSGS04]. Spectral analysis also has the
interesting property of defining an orthogonal basis. This property can be used
to avoid the degenerate configurations encountered by linear conformal param-
eterization methods [LPRM02], [DMA02]. Those linear methods use “vertex
pinning” to avoid the trivial constant solution, whereas the spectral method
introduced in [MTAD08] computes the first solution orthogonal to the trivial
one, that is to say, the eigenvector associated with the first non-zero eigenvalue.
However, the spectral parameterization methods listed above still need to
partition the mesh into charts. More recently, Dong et al. used the Laplacian to
decompose a mesh into quadrilaterals [DBG+ 05, DBG+ 06a], in a way that facil-
itates constructing a globally smooth parameterization. As shown in Figure 20,
their method first computes one eigenfunction of the Laplacian (the 38th in this
example), then extract the Morse complex of this function, filters and smoothes
the Morse complex, and uses it to partition the mesh into quads. These quads
are parameterized, and inter-chart smoothness can be further optimized us-
ing global relaxation [KLS03, THCM04]. More recently, a generalization was
proposed [HZM+ 08], still based on Laplacian eigenfunctions, and steered by a
guidance vector field.

5.1.2 Clustering and segmentation

One of the most well-known techniques in this regard is spectral clustering [BN03,
KVV00, NJW02]. Interested readers should refer to the recent survey by von
Luxburg [vL06] and the comparative study by Verma and Meilă [VM03]. Ng et
al. [NJW02] presented a clustering method where the entries of the first k eigen-
vectors corresponding to the largest eigenvalues of an affinity matrix are used
to obtain the transformed coordinates of the input data points. Additionally,
the embedded points are projected onto the unit k-sphere. Points that possess
high affinities tend to be grouped together in the spectral domain, where a sim-
ple clustering algorithm, such as k-means, can reveal the final clusters. Other
approaches differ only slightly from the core solution paradigm, e.g., in terms
of the operator used and the dimensionality of the embedding.
The ubiquity of the clustering problem makes spectral clustering an ex-
tremely useful technique. Kolluri et al. [KSO04] used spectral clustering to
determine the inside and outside tetrahedra in the context of a Delaunay-based
surface reconstruction algorithm. Liu and Zhang [LZ04] performed mesh seg-
mentation via spectral clustering. Basically, an affinity matrix is constructed
where the affinity measure depends on both geodesic distances and curvature
information. Next, the eigenvectors given by the eigendecomposition of this
matrix guide a clustering method, which provides patches of faces that define
the different segments of the mesh returned by the segmentation algorithm. It
is shown by example that it can be advantageous to perform segmentation in
the spectral domain, e.g., in terms of higher-quality cut boundaries.
In a follow-up work, Zhang and Liu [ZL05] presented a mesh segmentation
approach based on a recursive 2-way spectral clustering method. Only the first
two largest eigenvectors of the affinity matrix are computed. This provides a
one-dimensional embedding of mesh faces given by the quotient of the entries of
the two eigenvectors. The most salient cut in this embedding is located, given
by a part salience measure [HS97]. The cut provides a segmentation of the
faces into two parts. This process is recursively repeated in order to obtain a
hierarchical binary partitioning of the input mesh.
In yet another follow-up work, Liu and Zhang [LZ07] proposed an algorithm
for mesh segmentation via recursive bisection where at each step, a sub-mesh
embedded in 3D is spectrally projected to 2D and then a contour is extracted
from the planar embedding. Two operators are used in combination to com-
pute the projection: the well-known graph Laplacian and a geometric operator
designed to emphasize concavity. The two embeddings reveal distinctive shape
semantics of the 3D model and complement each other in capturing the struc-
tural or geometrical aspects of a segmentation. Transforming the shape analysis
problem to the 2D domain also facilitates segmentability analysis and sampling,
where the latter is needed to identify two samples residing on different parts of
the sub-mesh. These two samples are used by the Nyström method in the con-
struction of a 1D face sequence for finding an optimal cut, as in [ZL05]. Similar
to the technique presented in Section 4, the Nyström method is designed to ef-
ficiently compute eigenvectors of large matrices. However, it relies on sampling
and extrapolation and only approximates the eigenvectors.
Recently, de Goes et al. [dGGV08] presented a hierarchical segmentation
method for articulated bodies. Their approach relies on the diffusion distance,
which is a multi-scale metric based on the heat kernel and computed from the
eigenvectors of the Laplace-Beltrami operator. The diffusion distance is used
to compute a bijection between medial structures and segments of the model.
The medial structures yield a means to further refine the segmentation in an
iterative manner and provide a full hierarchy of segments for the shape.
Huang et al. [HWAG09] also performed hierarchical shape decomposition via
spectral analysis. However, the operator they use encapsulates shape geometry
beyond the static setting. The idea is to define a certain deformation energy and
use the eigenvectors of the Hessian of the deformation energy to characterize the
space of possible deformations of a given shape. The eigenmodes corresponding
to the low-end of the spectrum of the Hessian capture low-energy or in their for-
mulation, more rigid deformations, called “typical” deformations. The optimal
shape decomposition they compute is one whose optimal articulated (piecewise
rigid) deformation defined on the parts of the decomposition best conforms to
the basis vectors of the space of typical deformations. As a result, their method
tends to identify parts of a shape which would likely remain rigid during the
“typical” deformations.

5.1.3 Shape correspondence and retrieval

Depending on the requirement of the problem at hand, the mesh operator we
use to compute a spectral embedding can be made to incorporate any intrinsic
measure on a shape in order to obtain useful invariance properties, e.g., with
respect to part articulation or bending. In Figure 21, we show 3D spectral
embeddings of a few human and hand models obtained from an operator derived
from geodesic distances over the mesh surfaces. As geodesic distance is bending-
tolerant, the resulting embeddings are normalized with respect to bending and
can facilitate shape retrieval under part articulation [EK03, JZ07].
Elad and Kimmel [EK03] used MDS to compute bending-invariant signa-
tures for meshes. Geodesic distances between mesh vertices are computed by
fast marching. The resulting spectral embedding effectively normalizes the mesh
shapes with respect to translation, rotation, as well as bending transformations.
The similarity between two shapes is then given by the Euclidean distance be-
tween the moments of the first few eigenvectors, usually less than 15, and these
similarity distances can be used for shape classification.
Inspired by works in computer vision on spectral point correspondence [SB92],
Jain and Zhang [JZvK07] relied on higher-dimensional embeddings based on
the eigenvectors of an affinity matrix to obtain point correspondence between
two mesh shapes. The first k eigenvectors of the affinity matrix encoding the
geodesic distances between pairs of vertices are used to embed the model in a
k-dimensional space; typically k = 5 or 6. After this process is performed for
Figure 21: Spectral embeddings (bottom row) of some articulated 3D shapes
(top row). Since the mesh operator is constructed from geodesic distances, the
embeddings are normalized with respect to shape bending.

Figure 22: Eigenvector plots for two similar shapes, both with 252 vertices.
The entries in an eigenvector are color-mapped. As we can see, there is an
“eigenvector switching” occurring between the fifth and sixth eigenvectors. Such
“switching” is difficult to detect from the magnitude of the eigenvalues. The
first 8 eigenvalues for the two shapes are [205.6, 11.4, 4.7, 3.8, 1.8, 0.4, 0.26, 0.1]
and [201.9, 10.9, 6.3, 3.4, 1.8, 1.2, 0.31, 0.25], respectively.

two models, the two embeddings are non-rigidly aligned via thin-plate splines
and the correspondence between the two shapes is given by the proximity of the
vertices after such alignment. Any measure for the cost of a correspondence,
e.g., sum of distances between corresponding vertices, can be used as a similarity
distance for shape retrieval.
One of the key observations made in [JZvK07] is the presence of “eigen-
vector switching” due to non-uniform scaling of the shapes. Specifically, the
eigenmodes of similar shapes do not line up with respect to the magnitude of
their corresponding eigenvalues, as illustrated in Figure 22. As a result, it is
unreliable to sort the eigenvectors according to the magnitude of their respec-
tive eigenvalues, as has been done in all works on spectral correspondence so far.
Jain and Zhang [JZvK07] relied on a heuristic to “unswitch” the eigenmodes and
thin-plate splines to further align the shapes in the spectral domain [JZvK07].
Recent work of Mateus et al. [MCBH07] addressed the issue using an alignment
by the Expectation-Maximization (EM) algorithm instead.
The method by Leordeanu and Hebert [LH05] focuses on the global char-
acteristics of correspondence computation and aims at finding consistent cor-
respondences between two sets of shape or image features, where the spectral
approach has also found its utility. They build a graph whose nodes represent
possible feature pairings and edge weights measure how agreeable the pairings
are. The principal eigenvector of an affinity matrix W , one corresponding to the
largest eigenvalue, is inspected to detect how strongly the graph nodes belong
to the main cluster of W . The idea is that correct correspondences are likely to
establish links among each other and thus form a strongly connected cluster.
To define informative shape descriptors, another possibility consists in using
the heat diffusion equation and the heat kernel. Heat diffusion is governed by
the following differential equation :
∂f
= k∆f
∂t
where k is a constant. This equation admits a solution that can be written as :
Z
f (t, x) = K(t, y, x)dy

where K(t, x, y) denotes the heat kernel. Intuitively, considering that a Dirac of
heat was emitted from point x at time t = 0, the heat kernel K(t, x, y) indicates
the quantity of heat that can be measured at point y after time t. The heat
kernel admits an expression as an (infinite) sum of eigenfunctions :
∞
X
K(t, x, y) = e−λi t φi (x)φi (y)
i=0

Using this expression, the idea of using the auto-diffusion function was simul-
taneously proposed in [SOG] and [GBAL]. The auto-diffusion function corre-
sponds to the amount of heat that remains at point x after time t for a given
t, in other words K(t, x, x). This defines a scalar field on the surface that prov-
ably has good properties to be used as a shape descriptor [SOG]. In particular,
protrusions correspond to extrema of the function. The value of t acts as a
smoothing parameter (smaller values of t preserve most details, and higher val-
ues of t tend to a constant function). Another application of the auto-diffusion
function is to compute a Morse-Smale complex for shape segmentation [GBAL].

5.1.4 Global intrinsic symmetry detection

Ovsjanikov et al. [OSG08] propose an approach to detect the intrinsic sym-
metries of a shape which are invariant to isometry preserving transformations.
They show that if the shape is embedded into a signature space defined by the
eigenfunctions of the Laplace-Beltrami operator, then the intrinsic symmetries
are transformed into extrinsic Euclidean symmetries (rotations or reflections).
However, it is possible to restrict the search of symmetries only to reflections,
avoiding the search for rotational symmetries, a task that can be hard in high-
dimensional space. This result allows to obtain the intrinsic symmetries by
first computing the eigenvectors of the operator, then embedding the shape into
the signature space, and finally finding point-to-point correspondences between
symmetric points. The signature adopted is derived from the GPS embedding
of Rustamov [Rus07].

5.2 Use of spectral transforms

The spectral mesh transforms are closely related to the Fourier transform that
is the foundation of signal processing theory. Conceivably, any application of
the classical Fourier transform in signal or image processing has the potential
to be realized in the mesh setting. In geometry processing, the mesh signal
considered is often the embedding function that specifies the 3D coordinates of
each vertex. This signal serves to represent mesh geometry.

5.2.1 Geometry compression

Karni and Gotsman [KG00c] proposed an approach to compress the geometry
of triangle meshes. Firstly, the set of eigenvectors of the Tutte Laplacian is
computed. Next, the mesh vertex coordinates are projected into the spectral
space spanned by the computed eigenvectors. Part of the coefficients obtained by
this transformation is eliminated in order to reduce the storage space required
for mesh geometry. The coefficients related to the eigenvectors associated to
larger eigenvalues are firstly removed, which would correspond to high frequency
detail, when following an analogy with Fourier analysis.
The main drawback of this method is that many eigenvectors need to be
computed. Karni and Gotsman propose to partition the mesh into smaller sets of
vertices. Although that alleviates the problem of computing the eigenvectors for
large matrices, it still requires a good partitioning of the mesh for the efficiency
of the compression, and artifacts along the partition boundaries are evident
when higher compression rates are employed.

5.2.2 Watermarking
Ohbuchi et al. [OTMM01, OMT02] also employed the spectral transform ap-
proach, but to insert watermarks into triangle meshes. In their method, the
eigenvectors of the graph Laplacian are used as the basis for the projection.
After transforming the geometry of the mesh and obtaining the spectral coeffi-
cients, a watermark is inserted into the model by modifying coefficients at the
low-frequency end of the spectrum. In this way, modifications on the geometry
of the mesh are well-spread over the model and less noticeable than if they were
directly added to the vertex coordinates. This method also requires the compu-
tation of eigenvectors of the Laplace operator, which is prohibitive in the case
of large meshes. Mesh partitioning is again used to address this problem.

5.2.3 Fourier descriptors

2D Fourier descriptors have been quite successful as a means to characterize
2D shapes. Using eigendecomposition with respect to the mesh Laplacians,
one can compute analogous Fourier-like descriptors to describe mesh geometry.
However, there have not seen such mesh Fourier descriptors being proposed for
shape analysis so far. There have been methods, e.g., [VS01], which rely on
3D Fourier descriptors for 3D shape retrieval. In this case, the mesh shapes
are first voxelized and 3D Fourier descriptors are extracted from the resulting
volumetric data. We suspect that the main difficulties with the use of mesh
Fourier descriptors for shape matching include computational costs and the fact
that when the eigenmodes vary between the two mesh shapes to be matched,
it becomes doubtful whether their associated eigenspace projections can serve
as reliable shape descriptors. Also, even when the shapes are very similar,
eigenvector switching, as depicted in Figure 22, can occur when the eigenvectors
are ordered by the magnitude of their eigenvalues.

References
[AFW06] D. N. Arnold, R. S. Falk, and R. Winther. Finite element exterior
calculus, homological techniques, and applications. Acta Numerica
15, 2006.
[Arv95] James Arvo. The Role of Functional Analysis in Global Illumina-
tion. In P. M. Hanrahan and W. Purgathofer, editors, Rendering
Techniques ’95 (Proceedings of the Sixth Eurographics Workshop on
Rendering), pages 115–126, New York, NY, 1995. Springer-Verlag.

[BN03] M. Belkin and P. Niyogi. Laplacian eigenmaps and spectral

techniques for embedding and clustering. Neural Computations,
15(6):1373–1396, 2003.
[Bra99] Ronald N. Bracewell. The Fourier Transform And Its Applications.
McGraw-Hill, 1999.
[Cip93] Barri Cipra. You can’t always hear the shape of a drum. What’s
Happening in the Mathematical Sciences, 1, 1993.
[DBG+ 05] S. Dong, P.-T. Bremer, M. Garland, V. Pascucci, and J. C. Hart.
Quadrangulating a mesh using laplacian eigenvectors. Technical
report, June 2005.
[DBG+ 06a] S. Dong, P.-T. Bremer, M. Garland, V. Pascucci, and J. C. Hart.
Spectral mesh quadrangulation. ACM Transactions on Graphics
(SIGGRAPH 2006 special issue), 2006.

[DBG+ 06b] Shen Dong, Peer-Timo Bremer, Michael Garland, Valerio Pascucci,
and John C. Hart. Spectral surface quadrangulation. In SIG-
GRAPH ’06: ACM SIGGRAPH 2006 Papers, pages 1057–1066,
New York, NY, USA, 2006. ACM Press.
[dGGV08] Fernando de Goes, Siome Goldenstein, and Luiz Velho. A hierarchi-
cal segmentation of articulated bodies. Computer Graphics Forum
(Symposium on Geometry Processing), 27(5):1349–1356, 2008.
[DKT05] Mathieu Desbrun, Eva Kanzo, and Yiying Tong. Discrete differen-
tial forms for computational modeling. Siggraph ’05 course notes
on Discrete Differential Geometry, Chapter 7, 2005.
[DMA02] Mathieu Desbrun, Mark Meyer, and Pierre Alliez. Intrinsic pa-
rameterizations of surface meshes. In Proceedings of Eurographics,
pages 209–218, 2002.

[dV90] Y. Colin de Verdiere. Sur un nouvel invariant des graphes et un

critere de planarite. J. of Combinatorial Theory, 50, 1990.
[Dye06] Ramsey Dyer. Mass weights and the cot operator (personal com-
munication). Technical report, Simon Fraser University, CA, 2006.
[EK03] A. Elad and R. Kimmel. On bending invariant signatures for sur-
faces. IEEE Trans. Pattern Anal. Mach. Intell., 25(10):1285–1295,
2003.
[Fei96] J. Feidman. Computing betti numbers via combinatorial laplacians.
In Proc. 28th Sympos. Theory Comput., pages 386–391. ACM, 1996.

[FH04] M. S. Floater and K. Hormann. Surface parameterization: a tuto-

rial and survey. Springer, 2004.
[Fie73] Miroslav Fiedler. Algebraic connectivity of graphs. Czech. Math.
Journal, 23:298–305, 1973.
[Fie75] Miroslav Fiedler. A property of eigenvectors of nonnegative sym-
metric matrices and its application to graph theory. Czech. Math.
Journal, 25:619–633, 1975.
[GBAL] Katarzyna Gebal, Andreas Baerentzen, Henrik Aanaes, and Ras-
mus Larsen. Shape analysis using the auto diffusion function. Com-
puter Graphics Forum (Proc. of Symp. on Geom. Proc.).

[GGS03a] C. Gotsman, X. Gu, and A. Sheffer. Fundamentals of spherical

parameterization for 3d meshes. ACM Trans. Graph., 22(3):358–
363, 2003.
[GGS03b] C. Gotsman, X. Gu, and A. Sheffer. Fundamentals of spherical
parameterization for 3d meshes, 2003.

[Got03] Craig Gotsman. On graph partitioning, spectral analysis, and dig-

ital mesh processing. In Shape Modeling International, pages 165–
174, 2003.
[GY02] X. Gu and S.-T. Yau. Computing conformal structures of surfaces.
Communications in Information and Systems, 2(2):121–146, 2002.
[Hir03] Anil Hirani. Discrete exterior calculus. PhD thesis, 2003.
[HPW06] Klaus Hildebrandt, Konrad Polthier, and Max Wardetzky. On the
convergence of metric and geometric properties of polyhedral sur-
faces. Geom Dedicata, 2006.
[HS97] D. D. Hoffman and M. Singh. Salience of visual parts. Cognition,
63:29–78, 1997.
[HWAG09] Qixing Huang, Martin Wicke, Bart Adams, and Leonidas J.
Guibas. Shape decomposition using modal analysis. 28(2):to ap-
pear, 2009.
[HZM+ 08] Jin Huang, Muyang Zhang, Jin Ma, Xinguo Liu, Leif Kobbelt,
and Hujun Bao. Spectral quadrangulation with orientation and
alignment control. ACM Transactions on Graphics (SIGGRAPH
Asia conf. proc., 2008.
[IL05] Martin Isenburg and Peter Lindstrom. Streaming meshes. In IEEE
Visualization, page 30, 2005.
[Jai89] A. K. Jain. Fundamentals of Digital Image Processing. Prentice
Hall, 1989.
[JNT] Dmitry Jakobson, Nikolai Nadirashvili, and John Toth. Geometric
properties of eigenfunctions.
[JZ07] Varun Jain and Hao Zhang. A spectral approach to shape-based
retrieval of articulated 3D models. Computer Aided Design, 39:398–
407, 2007.
[JZvK07] Varun Jain, Hao Zhang, and Oliver van Kaick. Non-rigid spectral
correspondence of triangle meshes. International Journal on Shape
Modeling, 2007. to appear.
[Kac66] Mark Kac. Can you hear the shape of a drum? Amer. Math.
Monthly, 73, 1966.
[KG00a] Z. Karni and C. Gotsman. Spectral compression of mesh geometry.
In Proc. of ACM SIGGRAPH, pages 279–286, 2000.
[KG00b] Zachi Karni and Craig Gotsman. Spectral compression of mesh
geometry. In SIGGRAPH, pages 279–286, 2000.
[KG00c] Zachi Karni and Craig Gotsman. Spectral compression of mesh
geometry. In SIGGRAPH ’00: Proceedings of the 27th annual
conference on Computer graphics and interactive techniques, pages
279–286, New York, NY, USA, 2000. ACM Press/Addison-Wesley
Publishing Co.
[KLS03] A. Khodakovsky, N. Litke, and P. Schröder. Globally smooth pa-
rameterizations with low distortion. ACM TOG (SIGGRAPH),
2003.
[Kor02] Y. Koren. On spectral graph drawing, 2002.
[Kor03] Y. Koren. On spectral graph drawing. In Proc. of the International
Computing and Combinatorics Conference, pages 496–508, 2003.
[KSO04] Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F.
O’Brien. Spectral surface reconstruction from noisy point clouds. In
Proc. of Eurographics Symposium on Geometry Processing, pages
11–21, 2004.
[KVV00] R. Kannan, S. Vempala, and A. Vetta. On clustering - good, bad,
and spectral. In FOCS, pages 367–377, 2000.
[Lev06] Bruno Levy. Laplace-beltrami eigenfunctions: Towards an algo-
rithm that understands geometry. In IEEE International Confer-
ence on Shape Modeling and Applications, 2006.
[LH05] Marius Leordeanu and Martial Hebert. A spectral technique for
correspondence problems using pairwise constraints. In Interna-
tional Conference of Computer Vision (ICCV), volume 2, pages
1482–1489, October 2005.
[LPM02] Bruno Levy, Sylvain Petitjean, and Nicolas Ray Nicolas Jerome
Maillot. Least squares conformal maps for automatic texture atlas
generation. In ACM, editor, SIGGRAPH conf. proc, 2002.
[LPRM02] B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least squares confor-
mal maps for automatic texture atlas generation. In Proc. of ACM
SIGGRAPH 02, pages 362–371, 2002.
[LZ04] R. Liu and H. Zhang. Segmentation of 3D meshes through spectral
clustering. In Pacific Graphics, pages 298–305, 2004.
[LZ07] Rong Liu and Hao Zhang. Mesh segmentation via spectral em-
bedding and contour analysis. Computer Graphics Forum (Special
Issue of Eurographics 2007), 26:385–394, 2007.
[MCBH07] Diana Mateus, Fabio Cuzzolin, Edmond Boyer, and Radu Horaud.
Articulated shape matching by robust alignment of embedded rep-
resentations. In ICCV ’07 Workshop on 3D Representation for
Recognition (3dRR-07), 2007.
[MDSB03] Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H.
Barr. Discrete differential-geometry operators for triangulated 2-
manifolds. In Hans-Christian Hege and Konrad Polthier, editors,
Visualization and Mathematics III, pages 35–57. Springer-Verlag,
Heidelberg, 2003.
[MIT06] Omer Meshar, Dror Irony, and Sivan Toledo. An out-of-core sparse
symmetric indefinite factorization method. ACM Transactions on
Mathematical Software, 32:445–471, 2006.
[MTAD08] Patrick Mullen, Yiying Tong, Pierre Alliez, and Mathieu Desbrun.
Spectral conformal parameterization. In ACM/EG Symposium of
Geometry Processing, 2008.
[NJW02] A. Y. Ng, M. I. Jordan, and Y. Weiss. On spectral clustering: anal-
ysis and an algorithm. In Neural Information Processing Systems,
volume 14, pages 849–856, 2002.

[OMT02] R. Ohbuchi, A. Mukaiyama, and S. Takahashi. A frequency-domain

approach to watermarking 3D shapes. Computer Graphics Forum,
21(3):373–382, 2002.
[OSG08] Maks Ovsjanikov, Jian Sun, and Leonidas Guibas. Global intrinsic
symmetries of shapes. Computer Graphics Forum (Symposium on
Geometry Processing), 27(5):1341–1348, 2008.
[OTMM01] R. Ohbuchi, S. Takahashi, T. Miyazawa, and A. Mukaiyama. Wa-
termarking 3D polygonal meshes in the mesh spectral domain. In
Proc. of Graphics Interface, pages 9–18, 2001.
[PP93] Ulrich Pinkall and Konrad Polthier. Computing discrete minimal
surfaces and their conjugates. Experimental Mathematics, 2(1),
1993.
[Pra99] G. Prathap. Towards a science of fea: Patterns, predictability and
proof through some case studies. Current Science, 77:1311–1318,
1999.

[RS00a] S. T. Roweis and L. K. Saul. Nonlinear dimensionality reduction

by locally linear embedding. Science, 290:2323–2326, 2000.
[RS00b] Sam Roweis and Lawrence Saul. Nonlinear dimensionality reduc-
tion by locally linear embedding. Science, 290(5500):2323–2326,
Dec 2000.

[Rus07] R. M. Rustamov. Laplace-beltrami eigenfunctions for deformation

invariant shape representation. In Proc. of Eurographics Sympo-
sium on Geometry Processing, pages 225–233, 2007.
[RWP05a] M. Reuter, F.-E. Wolter, and N. Peinecke. Laplace-beltrami spectra
as ”shape-dna” of surfaces and solids. CAD Journal, 2005.
[RWP05b] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-
spectra as fingerprints for shape matching. In SPM ’05: Proceedings
of the 2005 ACM symposium on Solid and physical modeling, pages
101–106, New York, NY, USA, 2005. ACM Press.
[SB92] L. S. Shapiro and J. M. Brady. Feature-based correspondence: an
eigenvector approach. Image and Vision Computing, 10(5):283–
288, 1992.
[SGD05] P. Schröder, E. Grinspun, and M. Desbrun. Discrete differential
geometry: an applied introduction. In SIGGRAPH Course Notes,
2005.
[SM00] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmen-
tation. IEEE Trans. on Pattern Analysis and Machine Intelligence,
22(8):888–905, 2000.

[SOG] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A concise and
provably informative multi-scale signature based on heat diffusion.
Computer Graphics Forum (Proc. of Symp. on Geom. Proc.).
[Tau95a] G. Taubin. A signal processing approach to fair surface design. In
Proc. of ACM SIGGRAPH, pages 351–358, 1995.

[Tau95b] Gabriel Taubin. A signal processing approach to fair surface design.

In SIGGRAPH ’95: Proceedings of the 22nd annual conference on
Computer graphics and interactive techniques, pages 351–358, New
York, NY, USA, 1995. ACM Press.
[TB97] Lloyd N. Trefethen and David Bau. Numerical Linear Algebra.
SIAM, 1997.
[TdSL00] J. B. Tenenbaum, V. de Silva, and J. C. Langford. A global geo-
metric framework for nonlinear dimensionality reduction. Science,
290:2319–2323, 2000.

[THCM04] M. Tarini, K. Hormann, P. Cignoni, and C. Montani. Polycube-

maps. ACM TOG (SIGGRAPH), 2004.
[vL06] Ulrike von Luxburg. A tutorial on spectral clustering. Technical
Report TR-149, Max Plank Institute for Biological Cybernetics,
August 2006.

[VM03] D. Verma and M. Meila. A comparison of spectral clustering algo-

rithms. Technical Report UW-CSE-03-05-01, University of Wash-
ington, 2003.
[VS01] D. V. Vranić and D. Saupe. 3D shape descriptor based on 3D
Fourier transform. In Proc. EURASIP Conf. on Digital Signal
Processing for Multimedia Communications and Services, 2001.
[WBH+ 07] Max Wardetzky, Miklos Bergou, David Harmon, Denis Zorin, and
Eitan Grinspun. Discrete quadratic curvature energies. Computer
Aided Geometric Design (CAGD), 2007.
[Wei99] Y. Weiss. Segmentation using eigenvectors: A unifying view. In
Proc. of the International Conference on Computer Vision, pages
975–983, 1999.
[WK05] Jianhua Wu and Leif Kobbelt. Efficient spectral watermarking
of large meshes with orthogonal basis functions. In The Visual
Computer, 2005.
[WMKG07] Max Wardetzky, Saurabh Mathur, Felix Kalberer, and Eitan Grin-
spun. Discrete laplace operators: No free lunch. Eurographics
Symposium on Geometry Processing, 2007.

[You85] F. W. Young. Multidimensional scaling. Encyclopedia of Statistical

Sciences, 5:649–658, 1985.
[ZKK02] Gil Zigelman, Ron Kimmel, and Nahum Kiryati. Texture map-
ping using surface flattening via multidimensional scaling. IEEE
Transactions on Visualization and Computer Graphics, 8(2), 2002.

[ZL05] H. Zhang and R. Liu. Mesh segmentation via recursive and visually
salient spectral cuts. In Proc. of Vision, Modeling, and Visualiza-
tion, 2005.
[ZSGS04] Kun Zhou, John Snyder, Baining Guo, and Heung-Yeung Shum.
Iso-charts: Stretch-driven mesh parameterization using spectral
analysis. In Symposium on Geometry Processing, pages 47–56,
2004.
[ZvKDar] Hao Zhang, Oliver van Kaick, and Ramsay Dyer. Spectral mesh
processing. Computer Graphics Forum, 2009, to appear. http:
//www.cs.sfu.ca/∼haoz/pubs/zhang cgf09 spect survey.pdf.