MERL – A MITSUBISHI ELECTRIC RESEARCH LABORATORY

http://www.merl.com

Charting a manifold

Matthew Brand

TR-2003-13 March 2003

Abstract
We construct a nonlinear mapping from a high-dimensional sample space to a low-dimensional
vector space, effectively recovering a Cartesian coordinate system for the manifold from which
the data is sampled. The mapping preserves local geometric relations in the manifold and is
pseudo-invertible. We show how to estimate the intrinsic dimensionality of the manifold from
samples, decompose the sample data into locally linear low-dimensional patches, merge these
patches into a single lowdimensional coordinate system, and compute forward and reverse map-
pings between the sample and coordinate spaces. The objective functions are convex and their
solutions are given in closed form.

This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in
part without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies
include the following: a notice that such copying is by permission of Mitsubishi Electric Information Technology Center America; an
acknowledgment of the authors and individual contributions to the work; and all applicable portions of the copyright notice. Copying,
reproduction, or republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Information
Technology Center America. All rights reserved.

Copyright
c Mitsubishi Electric Information Technology Center America, 2003
201 Broadway, Cambridge, Massachusetts 02139
Presented at NIPS-15, December 2002. Appears in Proceedings, Neural Information Processing Systems,
volume 15. MIT Press, March 2003.
 Charting a manifold

Matthew Brand
Mitsubishi Electric Research Labs
201 Broadway, Cambridge MA 02139 USA
www.merl.com/people/brand/

Abstract
We construct a nonlinear mapping from a high-dimensional sample space
to a low-dimensional vector space, effectively recovering a Cartesian
coordinate system for the manifold from which the data is sampled.
The mapping preserves local geometric relations in the manifold and is
pseudo-invertible. We show how to estimate the intrinsic dimensionality
of the manifold from samples, decompose the sample data into locally
linear low-dimensional patches, merge these patches into a single low-
dimensional coordinate system, and compute forward and reverse map-
pings between the sample and coordinate spaces. The objective functions
are convex and their solutions are given in closed form.

1 Nonlinear dimensionality reduction (NLDR) by charting

Charting is the problem of assigning a low-dimensional coordinate system to data points
in a high-dimensional sample space. It is presumed that the data lies on or near a low-
dimensional manifold embedded in the sample space, and that there exists a 1-to-1 smooth
nonlinear transform between the manifold and a low-dimensional vector space. The data-
modeler’s goal is to estimate smooth continuous mappings between the sample and co-
ordinate spaces. Often this analysis will shed light on the intrinsic variables of the data-
generating phenomenon, for example, revealing perceptual or configuration spaces.
Our goal is to find a mapping—expressed as a kernel-based mixture of linear projections—
that minimizes information loss about the density and relative locations of sample points.
This constraint is expressed in a posterior that combines a standard gaussian mixture model
(GMM) likelihood function with a prior that penalizes uncertainty due to inconsistent pro-
jections in the mixture. Section 3 develops a special case where this posterior is unimodal
and maximizable in closed form, yielding a GMM whose covariances reveal a patchwork of
overlapping locally linear subspaces that cover the manifold. Section 4 shows that for this
(or any) GMM and a choice of reduced dimension d, there is a unique, closed-form solution
for a minimally distorting merger of the subspaces into a d-dimensional coordinate space,
as well as an reverse mapping defining the surface of the manifold in the sample space.
The intrinsic dimensionality d of the data manifold can be estimated from the growth pro-
cess of point-to-point distances. In analogy to differential geometry, we call the subspaces
“charts” and their merger the “connection.” Section 5 considers example problems where
these methods are used to untie knots, unroll and untwist sheets, and visualize video data.
1.1 Background
Topology-neutral NLDR algorithms can be divided into those that compute mappings, and
those that directly compute low-dimensional embeddings. The field has its roots in map-
ping algorithms: DeMers and Cottrell [3] proposed using auto-encoding neural networks
with a hidden layer “bottleneck,” effectively casting dimensionality reduction as a com-
pression problem. Hastie defined principal curves [5] as nonparametric 1 D curves that pass
through the center of “nearby” data points. A rich literature has grown up around properly
regularizing this approach and extending it to surfaces. Smola and colleagues [10] analyzed
the NLDR problem in the broader framework of regularized quantization methods.
More recent advances aim for embeddings: Gomes and Mojsilovic [4] treat manifold com-
pletion as an anisotropic diffusion problem, iteratively expanding points until they connect
to their neighbors. The I SO M AP algorithm [12] represents remote distances as sums of a
trusted set of distances between immediate neighbors, then uses multidimensional scaling
to compute a low-dimensional embedding that minimally distorts all distances. The locally
linear embedding algorithm (LLE) [9] represents each point as a weighted combination of
a trusted set of nearest neighbors, then computes a minimally distorting low-dimensional
barycentric embedding. They have complementary strengths: I SO M AP handles holes well
but can fail if the data hull is nonconvex [12]; and vice versa for LLE [9]. Both offer em-
beddings without mappings. It has been noted that trusted-set methods are vulnerable to
noise because they consider the subset of point-to-point relationships that has the lowest
signal-to-noise ratio; small changes to the trusted set can induce large changes in the set of
constraints on the embedding, making solutions unstable [1].
In a return to mapping, Roweis and colleagues [8] proposed global coordination—learning
a mixture of locally linear projections from sample to coordinate space. They constructed
a posterior that penalizes distortions in the mapping, and gave a expectation-maximization
(EM) training rule. Innovative use of variational methods highlighted the difficulty of even
hill-climbing their multimodal posterior. Like [2, 7, 6, 8], the method we develop below is
a decomposition of the manifold into locally linear neighborhoods. It bears closest relation
to global coordination [8], although by a different construction of the problem, we avoid
hill-climbing a spiky posterior and instead develop a closed-form solution.

2 Estimating locally linear scale and intrinsic dimensionality

.
We begin with matrix of sample points Y = [y1 , · · · , yN ], yn ∈ RD populating a D-
dimensional sample space, and a conjecture that these points are samples from a man-
ifold M of intrinsic dimensionality d < D. We seek a mapping onto a vector space
.
G(Y) → X = [x1 , · · · , xN ], xn ∈ Rd and 1-to-1 reverse mapping G−1 (X) → Y such that
local relations between nearby points are preserved (this will be formalized below). The
map G should be non-catastrophic, that is, without folds: Parallel lines on the manifold in
RD should map to continuous smooth non-intersecting curves in Rd . This guarantees that
linear operations on X such as interpolation will have reasonable analogues on Y.
Smoothness means that at some scale r the mapping from a neighborhood on M to Rd is
effectively linear. Consider a ball of radius r centered on a data point and containing n(r)
data points. The count n(r) grows as rd , but only at the locally linear scale; the grow rate
is inflated by isotropic noise at smaller scales and by embedding curvature at larger scales.
.
To estimate r, we look at how the r-ball grows as points are added to it, tracking c(r) =
d
d log n(r) log r. At noise scales, c(r) ≈ 1/D < 1/d, because noise has distributed points in
all directions with equal probability. At the scale at which curvature becomes significant,
c(r) < 1/d, because the manifold is no longer perpendicular to the surface of the ball, so
the ball does not have to grow as fast to accommodate new points. At the locally linear
scale, the process peaks at c(r) = 1/d, because points are distributed only in the directions
of the manifold’s local tangent space. The maximum of c(r) therefore gives an estimate
of both the scale and the local dimensionality of the manifold (see figure 1), provided that
the ball hasn’t expanded to a manifold boundary—boundaries have lower dimension than
 Scale behavior of a 1D manifold in 2-space Point−count growth process on a 2D manifold in 3−space
1
10
samples radial growth process
noise scale 1D hypothesis
locally linear scale 2D hypothesis
3D hypothesis
curvature scale

radius (log scale)

0
10

1 2 3
2 10 10 10
#points (log scale)

Figure 1: Point growth processes. L EFT: At the locally linear scale, the number of points
in an r-ball grows as rd ; at noise and curvature scales it grows faster. R IGHT: Using the
point-count growth process to find the intrinsic dimensionality of a 2D manifold nonlinearly
embedded in 3-space (see figure 2). Lines of slope 1/3 , 1/2 , and 1 are fitted to sections of the
log r/ log nr curve. For neighborhoods of radius r ≈ 1 with roughly n ≈ 10 points, the slope
peaks at 1/2 indicating a dimensionality of d = 2. Below that, the data appears 3 D because
it is dominated by noise (except for n ≤ D points); above, the data appears >2 D because of
manifold curvature. As the r-ball expands to cover the entire data-set the dimensionality
appears to drop to 1 as the process begins to track the 1D edges of the 2D sheet.

the manifold. For low-dimensional manifolds such as sheets, the boundary submanifolds
(edges and corners) are very small relative to the full manifold, so the boundary effect is
typically limited to a small rise in c(r) as r approaches the scale of the entire data set. In
practice, our code simply expands an r-ball at every point and looks for the first peak in
c(r), averaged over many nearby r-balls. One can estimate d and r globally or per-point.

3 Charting the data

In the charting step we find a soft partitioning of the data into locally linear low-dimensional
neighborhoods, as a prelude to computing the connection that gives the global low-
dimensional embedding. To minimize information loss in the connection, we require that
the data points project into a subspace associated with each neighborhood with (1) minimal
loss of local variance and (2) maximal agreement of the projections of nearby points into
nearby neighborhoods. Criterion (1) is served by maximizing the likelihood function of a
Gaussian mixture model (GMM) density fitted to the data:
.
p(yi |µ, Σ) = ∑ j p(yi |µ j , Σ j ) p j = ∑ j N (yi ; µ j , Σ j ) p j . (1)

Each gaussian component defines a local neighborhood centered around µ j with axes de-
fined by the eigenvectors of Σ j . The amount of data variance along each axis is indicated
by the eigenvalues of Σ j ; if the data manifold is locally linear in the vicinity of the µ j , all
but the d dominant eigenvalues will be near-zero, implying that the associated eigenvec-
tors constitute the optimal variance-preserving local coordinate system. To some degree
likelihood maximization will naturally realize this property: It requires that the GMM com-
ponents shrink in volume to fit the data as tightly as possible, which is best achieved by
positioning the components so that they “pancake” onto locally flat collections of data-
points. However, this state of affairs is easily violated by degenerate (zero-variance) GMM
components or components fitted to overly small enough locales where the data density off
the manifold is comparable to density on the manifold (e.g., at the noise scale). Conse-
quently a prior is needed.
Criterion (2) implies that neighboring partitions should have dominant axes that span sim-
ilar subspaces, since disagreement (large subspace angles) would lead to inconsistent pro-
jections of a point and therefore uncertainty about its location in a low-dimensional co-
ordinate space. The principal insight is that criterion (2) is exactly the cost of coding the
location of a point in one neighborhood when it is generated by another neighborhood—the
cross-entropy between the gaussian models defining the two neighborhoods:
N (y; µ1 ,Σ1 )
Z
D(N1 kN2 ) = dy N (y; µ1 ,Σ1 ) log
N (y; µ2 ,Σ2 )
> −1
1 Σ2 | + trace(Σ2 Σ1 ) + (µ2 −µ1 ) Σ2 (µ2 −µ1 ) − D)/2. (2)
= (log |Σ−1 −1

Roughly speaking, the terms in (2) measure differences in size, orientation, and position,
respectively, of two coordinate frames located at the means µ1 , µ2 with axes specified by
the eigenvectors of Σ1 , Σ2 . All three terms decline to zero as the overlap between the two
frames is maximized. To maximize consistency between adjacent neighborhoods, we form
.
the prior p(µ, Σ) = exp[− ∑i6= j mi (µ j )D(Ni kN j )], where mi (µ j ) is a measure of co-locality.
Unlike global coordination [8], we are not asking that the dominant axes in neighboring
charts are aligned—only that they span nearly the same subspace. This is a much easier
objective to satisfy, and it contains a useful special case where the posterior p(µ, Σ|Y) ∝
∑i p(yi |µ, Σ)p(µ, Σ) is unimodal and can be maximized in closed form: Let us associate a
gaussian neighborhood with each data-point, setting µi = yi ; take all neighborhoods to be
a priori equally probable, setting pi = 1/N; and let the co-locality measure be determined
from some local kernel. For example, in this paper we use mi (µ j ) ∝ N (µ j ; µi , σ2 ), with
the scale parameter σ specifying the expected size of a neighborhood on the manifold in
sample space. A reasonable choice is σ = r/2, so that 2erf(2) > 99.5% of the density of
mi (µ j ) is contained in the area around yi where the manifold is expected to be locally linear.
With uniform pi and µi , mi (µ j ) and fixed, the MAP estimates of the GMM covariances are
!,
∑ mi (µ j ) (y j − µi )(y j − µi )> + (µ j − µi )(µ j − µi )> + Σ j ∑ mi (µ j ) (3).
 
Σi =
j j

Note that each covariance Σi is dependent on all other Σ j . The MAP estimators for all
covariances can be arranged into a set of fully constrained linear equations and solved ex-
actly for their mutually optimal values. This key step brings nonlocal information about
the manifold’s shape into the local description of each neighborhood, ensuring that ad-
joining neighborhoods have similar covariances and small angles between their respective
subspaces. Even if a local subset of data points are dense in a direction perpendicular to
the manifold, the prior encourages the local chart to orient parallel to the manifold as part
of a globally optimal solution, protecting against a pathology noted in [8]. Equation (3) is
easily adapted to give a reduced number of charts and/or charts centered on local centroids.

4 Connecting the charts

We now build a connection for set of charts specified as an arbitrary nondegenerate GMM. A
GMM gives a soft partitioning of the dataset into neighborhoods of mean µk and covariance
Σk . The optimal variance-preserving low-dimensional coordinate system for each neigh-
borhood derives from its weighted principal component analysis, which is exactly specified
by the eigenvectors of its covariance matrix: Eigendecompose Vk Λk V> k ← Σk with eigen-
.
values in descending order on the diagonal of Λk and let Wk = [Id , 0]V> k be the operator
.
projecting points into the kth local chart, such that local chart coordinate uki = Wk (yi − µk )
.
and Uk = [uk1 , · · · , ukN ] holds the local coordinates of all points.
Our goal is to sew together all charts into a globally consistent low-dimensional coordinate
system. For each chart there will be a low-dimensional affine transform Gk ∈ R(d+1)×d
that projects Uk into the global coordinate space. Summing over all charts, the weighted
average of the projections of point yi into the low-dimensional vector space is
   
. W j (y − µ j ) . u ji
x|y = ∑ G j
c p j|y (y) ⇒ xi |yi = ∑ G j
d p j|y (yi ), (4)
j
1 j
1

where pk|y (y) ∝ pk N (y; µk , Σk ), ∑k pk|y (y) = 1 is the probability that chart k generates
point y. As pointed out in [8], if a point has nonzero probabilities in two charts, then there
should be affine transforms of those two charts that map the point to the same place in a
global coordinate space. We set this up as a weighted least-squares problem:
    2
. uki u ji
G = [G1 , · · · , GK ] = arg min ∑ pk|y (yi )p j|y (yi )

Gk − G j
. (5)
Gk ,G j i 1 1 F

Equation (5) generates a homogeneous set of equations that determines a solution up to an

affine transform of G. There are two solution methods. First, let us temporarily anchor one
neighborhood at the origin to fix this indeterminacy. This adds the constraint G1 = [I, 0]> .
.
To solve, define indicator matrix Fk = [0, · · · , 0, I, 0, · · · , 0]> with the identity ma-
.
trix occupying the kth block, such that Gk = GFk . Let the diagonal of Pk =
diag([pk|y (y1 ), · · · , pk|y (yN )]) record the per-point posteriors of chart k. The squared error
of the connection is then a sum of of all patch-to-anchor and patch-to-patch inconsistencies:
"   2 #  
. U . U
E = ∑ (GUk − 0 )Pk P1 + ∑ (GU j − GUk )P j Pk F ; Uk = Fk 1k .
1
2
k F j6=k
(6)
Setting dE /dG = 0 and solving to minimize convex E gives
! !−1  > !
U1
G = >
∑ Uk P2k ∑ P2j U>
k − ∑ Uk P2k P2j U>j ∑ Uk P2k P21
0
. (7)
k j6=k j6=k k

We now remove the dependence on a reference neighborhood G1 by rewriting equation 5,

 
G = arg min ∑ j6=k k(GU j − GUk )P j Pk k2F = kGQk2F = trace(GQQ> G> ) , (8)
G
.
where Q = ∑ j6=k U j − Uk P j Pk . If we require that GG> = I to prevent degenerate

solutions, then equation (8) is solved (up to rotation in coordinate space) by setting G> to
the eigenvectors associated with the smallest eigenvalues of QQ> . The eigenvectors can be
computed efficiently without explicitly forming QQ> ; other numerical efficiencies obtain
by zeroing any vanishingly small probabilities in each Pk , yielding a sparse eigenproblem.
A more interesting strategy is to numerically condition the problem by calculating the
trailing eigenvectors of QQ> + 1. It can be shown that this maximizes the posterior
2
p(G|Q) ∝ p(Q|G)p(G) ∝ e−kGQkF e−kG1k , where the prior p(G) favors a mapping G
whose unit-norm rows are also zero-mean. This maximizes variance in each row of G
and thereby spreads the projected points broadly and evenly over coordinate space.
The solutions for MAP charts (equation (5)) and connection (equation (8)) can be applied
to any well-fitted mixture of gaussians/factors1 /PCAs density model; thus large eigen-
problems can be avoided by connecting just a small number of charts that cover the data.
1 We thank reviewers for calling our attention to Teh & Roweis ([11]—in this volume), which
shows how to connect a set of given local dimensionality reducers in a generalized eigenvalue prob-
lem that is related to equation (8).
 LLE, n=5 LLE, n=6 LLE, n=7
original data embedding, XY view XYZ view

data (linked)

XZ view
random subset
LLE, n=8 LLE, n=9 LLE, n=10

of local charts
XY view
charting reconstruction
(projection onto coordinate space) (back−projected coordinate grid)
charting best Isomap best LLE
(regularized)

Figure 2: The twisted curl problem. L EFT: Comparison of charting, I SO M AP, & LLE.
400 points are randomly sampled from the manifold with noise. Charting is the only
method that recovers the original space without catastrophes (folding), albeit with some
shear. R IGHT: The manifold is regularly sampled (with noise) to illustrate the forward
and backward projections. Samples are shown linked into lines to help visualize the man-
ifold structure. Coordinate axes of a random selection of charts are shown as bold lines.
Connecting subsets of charts such as this will also give good mappings. The upper right
quadrant shows various LLE results. At bottom we show the charting solution and the
reconstructed (back-projected) manifold, which smooths out the noise.

Once the connection is solved, equation (4) gives the forward projection of any point y
down into coordinate space. There are several numerically distinct candidates for the back-
projection: posterior mean, mode, or exact inverse. In general, there may not be a unique
posterior mode and the exact inverse is not solvable in closed form (this is also true of [8]).
Note that chart-wise projection defines a complementary density in coordinate space
   
0 [Id , 0]Λk [Id , 0]> 0
px|k (x) = N (x; Gk , Gk G>
k ). (9)
1 0 0
Let p(y|x, k), used to map x into subspace k on the surface of the manifold, be a Dirac delta
function whose mean is a linear function of x. Then the posterior mean back-projection is
obtained by integrating out uncertainty over which chart generates x:
  +   !
I 0
c=
y|x ∑ pk|x (x) µk + Wk Gk 0
>
x − Gk
1
, (10)
k

where (·)+ denotes pseudo-inverse. In general, a back-projecting map should not recon-
struct the original points. Instead, equation (10) generates a surface that passes through the
weighted average of the µi of all the neighborhoods in which yi has nonzero probability,
much like a principal curve passes through the center of each local group of points.

5 Experiments
Synthetic examples: 400 2 D points were randomly sampled from a 2 D square and embed-
ded in 3 D via a curl and twist, then contaminated with gaussian noise. Even if noiselessly
sampled, this manifold cannot be “unrolled” without distortion. In addition, the outer curl
is sampled much less densely than the inner curl. With an order of magnitude fewer points,
higher noise levels, no possibility of an isometric mapping, and uneven sampling, this is
arguably a much more challenging problem than the “swiss roll” and “s-curve” problems
featured in [12, 9, 8, 1]. Figure 2LEFT contrasts the (unique) output of charting and the
best outputs obtained from I SO M AP and LLE (considering all neighborhood sizes between
2 and 20 points). I SO M AP and LLE show catastrophic folding; we had to change LLE’s
regularization in order to coax out nondegenerate (>1 D) solutions. Although charting is
 a. data, xy view b. data, yz view c. local charts d. 2D embedding e. 1D embedding

1D ordinate
true manifold arc length
Figure 3: Untying a trefoil knot ( ) by charting. 900 noisy samples from a 3 D-embedded
1 D manifold are shown as connected dots in front (a) and side (b) views. A subset of charts
is shown in (c). Solving for the 2 D connection gives the “unknot” in (d). After removing
some points to cut the knot, charting gives a 1 D embedding which we plot against true
manifold arc length in (e); monotonicity (modulo noise) indicates correctness.

Three principal degrees of freedom recovered from raw jittered images

pose

scale

expression
images synthesized via backprojection of straight lines in coordinate space
Figure 4: Modeling the manifold of facial images from raw video. Each row contains
images synthesized by back-projecting an axis-parallel straight line in coordinate space
onto the manifold in image space. Blurry images correspond to points on the manifold
whose neighborhoods contain few if any nearby data points.

not designed for isometry, after affine transform the forward-projected points disagree with
the original points with an RMS error of only 1.0429, lower than the best LLE (3.1423) or
best I SO M AP (1.1424, not shown). Figure 2RIGHT shows the same problem where points
are sampled regularly from a grid, with noise added before and after embedding. Figure 3
shows a similar treatment of a 1 D line that was threaded into a 3 D trefoil knot, contaminated
with gaussian noise, and then “untied” via charting.
Video: We obtained a 1965-frame video sequence (courtesy S. Roweis and B. Frey) of
20 × 28-pixel images in which B.F. strikes a variety of poses and expressions. The video
is heavily contaminated with synthetic camera jitters. We used raw images, though image
processing could have removed this and other uninteresting sources of variation. We took a
500-frame subsequence and left-right mirrored it to obtain 1000 points in 20 × 28 = 560D
image space. The point-growth process peaked just above d = 3 dimensions. We solved for
25 charts, each centered on a random point, and a 3D connection. The recovered degrees
of freedom—recognizable as pose, scale, and expression—are visualized in figure 4.

original data stereographic map to 3D fishbowl charting

Figure 5: Flattening a fishbowl. From the left: Original 2000×2D points; their stereo-
graphic mapping to a 3D fishbowl; its 2D embedding recovered using 500 charts; and the
stereographic map. Fewer charts lead to isometric mappings that fold the bowl (not shown).
Conformality: Some manifolds can be flattened conformally (preserving local angles) but
not isometrically. Figure 5 shows that if the data is finely charted, the connection behaves
more conformally than isometrically. This problem was suggested by J. Tenenbaum.

6 Discussion
Charting breaks kernel-based NLDR into two subproblems: (1) Finding a set of data-
covering locally linear neighborhoods (“charts”) such that adjoining neighborhoods span
maximally similar subspaces, and (2) computing a minimal-distortion merger (“connec-
tion”) of all charts. The solution to (1) is optimal w.r.t. the estimated scale of local linearity
r; the solution to (2) is optimal w.r.t. the solution to (1) and the desired dimensionality d.
Both problems have Bayesian settings. By offloading the nonlinearity onto the kernels,
we obtain least-squares problems and closed form solutions. This scheme is also attractive
because large eigenproblems can be avoided by using a reduced set of charts.
The dependence on r, like trusted-set methods, is a potential source of solution instabil-
ity. In practice the point-growth estimate seems fairly robust to data perturbations (to be
expected if the data density changes slowly over a manifold of integral Hausdorff dimen-
sion), while the use of a soft neighborhood partitioning appears to make charting solutions
reasonably stable to variations in r. Eigenvalue stability analyses may prove useful here.
Ultimately, we would prefer to integrate r out. In contrast, use of d appears to be a virtue:
Unlike other eigenvector-based methods, the best d-dimensional embedding is not merely
a linear projection of the best d + 1-dimensional embedding; a unique distortion is found
for each value of d that maximizes the information content of its embedding.
Why does charting performs well on datasets where the signal-to-noise ratio confounds
recent state-of-the-art methods? Two reasons may be adduced: (1) Nonlocal information
is used to construct both the system of local charts and their global connection. (2) The
mapping only preserves the component of local point-to-point distances that project onto
the manifold; relationships perpendicular to the manifold are discarded. Thus charting uses
global shape information to suppress noise in the constraints that determine the mapping.

Acknowledgments
Thanks to J. Buhmann, S. Makar, S. Roweis, J. Tenenbaum, and anonymous reviewers for
insightful comments and suggested “challenge” problems.

References
[1] M. Balasubramanian and E. L. Schwartz. The IsoMap algorithm and topological stability. Sci-
ence, 295(5552):7, January 2002.
[2] C. Bregler and S. Omohundro. Nonlinear image interpolation using manifold learning. In
NIPS–7, 1995.
[3] D. DeMers and G. Cottrell. Nonlinear dimensionality reduction. In NIPS–5, 1993.
[4] J. Gomes and A. Mojsilovic. A variational approach to recovering a manifold from sample
points. In ECCV, 2002.
[5] T. Hastie and W. Stuetzle. Principal curves. J. Am. Statistical Assoc, 84(406):502–516, 1989.
[6] G. Hinton, P. Dayan, and M. Revow. Modeling the manifolds of handwritten digits. IEEE
Trans. Neural Networks, 8, 1997.
[7] N. Kambhatla and T. Leen. Dimensionality reduction by local principal component analysis.
Neural Computation, 9, 1997.
[8] S. Roweis, L. Saul, and G. Hinton. Global coordination of linear models. In NIPS–13, 2002.
[9] S. T. Roweis and L. K. Saul. Nonlinear dimensionality reduction by locally linear embedding.
Science, 290:2323–2326, December 22 2000.
[10] A. Smola, S. Mika, B. Schölkopf, and R. Williamson. Regularized principal manifolds. Ma-
chine Learning, 1999.
[11] Y. W. Teh and S. T. Roweis. Automatic alignment of hidden representations. In NIPS–15, 2003.
[12] J. B. Tenenbaum, V. de Silva, and J. C. Langford. A global geometric framework for nonlinear
dimensionality reduction. Science, 290:2319–2323, December 22 2000.