1 Introduction
Many seemingly complex systems described by highdimensional data sets are in fact governed by a surprisingly low number of parameters. Revealing the lowdimensional representation of such highdimensional data sets not only leads to a more compact description of the data, but also enhances our understanding of the system. Dimensionreducing algorithms attempt to simplify the system’s representation without losing significant structural information. Various dimensionreduction algorithms were developed recently to perform embeddings for manifoldbased data sets. These include the following algorithms: Locally Linear Embedding
(LLE, Roweis and Saul, 2000), Isomap (Tenenbaum et al., 2000), Laplacian Eigenmaps (LEM, Belkin and Niyogi, 2003), Local Tangent Space Alignment (LTSA, Zhang and Zha, 2004), Hessian Eigenmap (HLLE, Donoho and Grimes, 2004), Semidefinite Embedding (SDE, Weinberger and Saul, 2006) and Diffusion Maps (DFM, Coifman and Lafon, 2006).These manifoldlearning algorithms compute an embedding for some given input. It is assumed that this input lies on a lowdimensional manifold, embedded in some highdimensional space. Here a manifold is defined as a topological space that is locally equivalent to a Euclidean space. It is further assumed that the manifold is the image of a lowdimensional domain. In particular, the input points are the image of a sample taken from the domain. The goal of the manifoldlearning algorithms is to recover the original domain structure, up to some scaling and rotation. The nonlinearity of these algorithms allows them to reveal the domain structure even when the manifold is not linearly embedded.
The central question that arises when considering the output of a manifoldlearning algorithm is, whether the algorithm reveals the underlying lowdimensional structure of the manifold. The answer to this question is not simple. First, one should define what “revealing the underlying lowerdimensional description of the manifold” actually means. Ideally, one could measure the degree of similarity between the output and the original sample. However, the original lowdimensional data representation is usually unknown. Nevertheless, if the lowdimensional structure of the data is known in advance, one would expect it to be approximated by the dimensionreducing algorithm, at least up to some rotation, translation, and global scaling factor. Furthermore, it would be reasonable to expect the algorithm to succeed in recovering the original sample’s structure asymptotically, namely, when the number of input points tends to infinity. Finally, one would hope that the algorithm would be robust in the presence of noise.
Previous papers have addressed the central question posed earlier. Zhang and Zha (2004)
presented some bounds on the localneighborhoods’ errorestimation for LTSA. However, their analysis says nothing about the global embedding.
Huo and Smith (2006) proved that, asymptotically, LTSA recovers the original sample up to an affine transformation. They assume in their analysis that the level of noise tends to zero when the number of input points tends to infinity. Bernstein et al. (2000.) proved that, asymptotically, the embedding given by the Isomap algorithm (Tenenbaum et al., 2000) recovers the geodesic distances between points on the manifold.In this paper we develop theoretical results regarding the performance of a class of manifoldlearning algorithms, which includes the following five algorithms: Locally Linear Embedding (LLE), Laplacian Eigenmap (LEM), Local Tangent Space Alignment (LTSA), Hessian Eigenmaps (HLLE), and Diffusion maps (DFM).
We refer to this class of algorithms as the normalizedoutput algorithms. The normalizedoutput algorithms share a common scheme for recovering the domain structure of the input data set. This scheme is constructed in three steps. In the first step, the local neighborhood of each point is found. In the second step, a description of these neighborhoods is computed. In the third step, a lowdimensional output is computed by solving some convex optimization problem under some normalization constraints. A detailed description of the algorithms is given in Section 2.
In Section 3 we discuss informally the criteria for determining the success of manifoldlearning algorithms. We show that one should not expect the normalizedoutput algorithms to recover geodesic distances or local structures. A more reasonable criterion for success is a high degree of similarity between the output of the algorithms and the original sample, up to some affine transformation; the definition of similarity will be discussed later. We demonstrate that under certain circumstances, this high degree of similarity does not occur. In Section 4 we find necessary conditions for the successful performance of LEM and DFM on the twodimensional grid. This section serves as an explanatory introduction to the more general analysis that appears in Section 5. Some of the ideas that form the basis of the analysis in Section 4 were discussed independently by both Gerber et al. (2007) and ourselves (Goldberg et al., 2007). Section 5 finds necessary conditions for the successful performance of all the normalizedoutput algorithms on general twodimensional manifolds. In Section 6 we discuss the performance of the algorithms in the asymptotic case. Concluding remarks appear in Section 7. The detailed proofs appear in the Appendix.
Our paper has two main results. First, we give welldefined necessary conditions for the successful performance of the normalizedoutput algorithms. Second, we show that there exist simple manifolds that do not fulfill the necessary conditions for the success of the algorithms. For these manifolds, the normalizedoutput algorithms fail to generate output that recovers the structure of the original sample. We show that these results hold asymptotically for LEM and DFM. Moreover, when noise, even of small variance, is introduced, LLE, LTSA, and HLLE will fail asymptotically on some manifolds. Throughout the paper, we present numerical results that demonstrate our claims.
2 Description of Outputnormalized Algorithms
In this section we describe in short the normalizedoutput algorithms. The presentation of these algorithms is not in the form presented by the respective authors. The form used in this paper emphasizes the similarities between the algorithms and is bettersuited for further derivations. In Appendix A.1 we show the equivalence of our representation of the algorithms and the representations that appear in the original papers.
Let be the input data where is the dimension of the ambient space and is the size of the sample. The normalizedoutput algorithms attempt to recover the underlying structure of the input data in three steps.
In the first step, the normalizedoutput algorithms assign neighbors to each input point based on the Euclidean distances in the highdimensional space^{1}^{1}1The neighborhoods are not mentioned explicitly by Coifman and Lafon (2006). However, since a sparse optimization problem is considered, it is assumed implicitly that neighborhoods are defined (see Sec. 2.7 therein).. This can be done, for example, by choosing all the input points in an ball around or alternatively by choosing ’s nearestneighbors. The neighborhood of is given by the matrix where are the neighbors of . Note that can be a function of , the index of the neighborhood, yet we omit this index to simplify the notation. For each neighborhood, we define the radius of the neighborhood as
(1) 
where we define . Finally, we assume throughout this paper that the neighborhood graph is connected.
In the second step, the normalizedoutput algorithms compute a description of the local neighborhoods that were found in the previous step. The description of the th neighborhood is given by some weight matrix . The matrices for the different algorithms are presented.

LEM and DFM: is a matrix,
For LEM is a natural choice, yet it is also possible to define the weights as , where is the width parameter of the kernel. For the case of DFM,
(2) where is some rotationinvariant kernel, and is again a width parameter. We will use in the normalization of the diffusion kernel, yet other values of can be considered (see details in Coifman and Lafon, 2006). For both LEM and DFM, we define the matrix to be a diagonal matrix where .

LLE: is a matrix,
The weights are chosen so that can be best linearly reconstructed from its neighbors. The weights minimize the reconstruction error function
(3) under the constraint . In the case where there is more than one solution that minimizes , regularization is applied to force a unique solution (for details, see Saul and Roweis, 2003).

LTSA: is a matrix,
Let be the SVD of where is the sample mean of and 1
is a vector of ones
(for details about SVD, see, for example, Golub and Loan, 1983). Let be the matrix that holds the first columns of where is the output dimension. The matrix is the centering matrix. See also Huo and Smith (2006) regarding this representation of the algorithm. 
HLLE: is a matrix,
where 0 is a vector of zeros and is the Hessian estimator.
The estimator can be calculated as follows. Let be the SVD of . Letwhere the operator diag returns a column vector formed from the diagonal elements of the matrix. Let be the result of the GramSchmidt orthonormalization on . Then is defined as the transpose of the last columns of .
The third step of the normalizedoutput algorithms is to find a set of points where is the dimension of the manifold. is found by minimizing a convex function under some normalization constraints, as follows. Let be any matrix. We define the th neighborhood matrix using the same pairs of indices as in . The cost function for all of the normalizedoutput algorithms is given by
(4) 
under the normalization constraints
(5) 
where stands for the Frobenius norm, and is algorithmdependent.
Define the output matrix to be the matrix that achieves the minimum of under the normalization constraints of Eq. 5 ( is defined up to rotation). Then we have the following: the embeddings of LEM and LLE are given by the according output matrices ; the embeddings of LTSA and HLLE are given by the according output matrices
; and the embedding of DFM is given by a linear transformation of
as discussed in Appendix A.1. The discussion of the algorithms’ output in this paper holds for any affine transformation of the output (see Section 3). Thus, without loss of generality, we prefer to discuss the output matrix directly, rather than the different embeddings. This allows a unified framework for all five normalizedoutput algorithms.3 Embedding quality
In this section we discuss possible definitions of “successful performance” of manifoldlearning algorithms. To open our discussion, we present a numerical example. We chose to work with LTSA rather arbitrarily. Similar results can be obtained using the other algorithms.
The example we consider is a uniform sample from a twodimensional strip, shown in Fig. 1A. Note that in this example, ; i.e., the input data is identical to the original data. Fig. 1B presents the output of LTSA on the input in Fig. 1A. The most obvious difference between input and output is that while the input is a strip, the output is roughly square. While this may seem to be of no importance, note that it means that the algorithm, like all the normalizedoutput algorithms, does not preserve geodesic distances even up to a scaling factor. By definition, the geodesic distance between two points on a manifold is the length of the shortest path on the manifold between the two points. Preservation of geodesic distances is particularly relevant when the manifold is isometrically embedded. In this case, assuming the domain is convex, the geodesic distance between any two points on the manifold is equal to the Euclidean distance between the corresponding domain points. Geodesic distances are conserved, for example, by the Isomap algorithm (Tenenbaum et al., 2000).
Figs. 1E and 1F present closeups of Figs. 1A and 1B, respectively. Here, a less obvious phenomenon is revealed: the structure of the local neighborhood is not preserved by LTSA. By local structure we refer to the angles and distances (at least up to a scale) between all points within each local neighborhood. Mappings that preserve local structures up to a scale are called conformal mappings (see for example de Silva and Tenenbaum, 2003; Sha and Saul, 2005). In addition to the distortion of angles and distances, the nearestneighbors of a given point on the manifold do not necessarily correspond to the nearestneighbors of the respective output point, as shown in Figs. 1E and 1F. Accordingly, we conclude that the original structure of the local neighborhoods is not necessarily preserved by the normalizedoutput algorithms.
The above discussion highlights the fact that one cannot expect the normalizedoutput algorithms to preserve geodesic distances or local neighborhood structure. However, it seems reasonable to demand that the output of the normalizedoutput algorithms resemble an affine transformation of the original sample. In fact, the output presented in Fig. 1B is an affine transformation of the input, which is the original sample, presented in Fig. 1A. A formal similarity criterion based on affine transformations is given by Huo and Smith (2006). In the following, we will claim that a normalizedoutput algorithm succeeds (or fails) based on the existence (or lack thereof) of resemblance between the output and the original sample, up to an affine transformation.
Fig. 1D presents the output of LTSA on a noisy version of the input, shown in Fig. 1C. In this case, the algorithm prefers an output that is roughly a onedimensional curve embedded in . While this result may seem incidental, the results of all the other normalizedoutput algorithms for this example are essentially the same.
Using the affine transformation criterion, we can state that LTSA succeeds in recovering the underlying structure of the strip shown in Fig. 1A. However, in the case of the noisy strip shown in Fig. 1C, LTSA fails to recover the structure of the input. We note that all the other normalizedoutput algorithms perform similarly.
For practical purposes, we will now generalize the definition of failure of the normalizedoutput algorithms. This definition is more useful when it is necessary to decide whether an algorithm has failed, without actually computing the output. This is useful, for example, when considering the outputs of an algorithm for a class of manifolds.
We now present the generalized definition of failure of the algorithms. Let be the original sample. Assume that the input is given by , where is some smooth function, and is the dimension of the input. Let be an affine transformation of the original sample , such that the normalization constraints of Eq. 5 hold. Note that is algorithmdependent, and that for each algorithm, is unique up to rotation and translation. When the algorithm succeeds it is expected that the output will be similar to a normalized version of , namely to . Let be any matrix that satisfies the same normalization constraints. We say that the algorithm has failed if , and is substantially different from , and hence also from . In other words, we say that the algorithm has failed when a substantially different embedding has a lower cost than the most appropriate embedding . A precise definition of “substantially different” is not necessary for the purposes of this paper. It is enough to consider substantially different from when is of lower dimension than , as in Fig. 1D.
We emphasize that the matrix is not necessarily similar to the output of the algorithm in question. It is a mathematical construction that shows when the output of the algorithm is not likely to be similar to , the normalized version of the true manifold structure. The following lemma shows that if , the inequality is also true for a small perturbation of . Hence, it is not likely that an output that resembles will occur when and is substantially different from .
Lemma 3.1
Let be an matrix. Let be a perturbation of , where is an matrix such that and where . Let be the maximum number of neighborhoods to which a single input point belongs. Then for LLE with positive weights , LEM, DFM, LTSA, and HLLE, we have
where is a constant that depends on the algorithm.
4 Analysis of the twodimensional grid
In this section we analyze the performance of LEM on the twodimensional grid. In particular, we argue that LEM cannot recover the structure of a twodimensional grid in the case where the aspect ratio of the grid is greater than . Instead, LEM prefers a onedimensional curve in . Implications also follow for DFM, as explained in Section 4.3, followed by a discussion of the other normalizedoutput algorithms. Finally, we present empirical results that demonstrate our claims.
In Section 5 we prove a more general statement regarding any twodimensional manifold. Necessary conditions for successful performance of the normalizedoutput algorithms on such manifolds are presented. However, the analysis in this section is important in itself for two reasons. First, the conditions for the success of LEM on the twodimensional grid are more limiting. Second, the analysis is simpler and points out the reasons for the failure of all the normalizedoutput algorithms when the necessary conditions do not hold.
4.1 Possible embeddings of a twodimensional grid
We consider the input data set to be the twodimensional grid , where . We denote . For convenience, we regard as an matrix, where is the number of points in the grid. Note that in this specific case, the original sample and the input are the same.
In the following we present two different embeddings, and . Embedding is the grid itself, normalized so that . Embedding collapses each column to a point and positions the resulting points in the twodimensional plane in a way that satisfies the constraint (see Fig. 2 for both). The embedding is a curve in and clearly does not preserve the original structure of the grid.
We first define the embeddings more formally. We start by defining . Note that this is the only linear transformation of (up to rotation) that satisfies the conditions and , which are the normalization constraints for LEM (see Eq. 5). However, the embedding depends on the matrix , which in turn depends on the choice of neighborhoods. Recall that the matrix is a diagonal matrix, where equals the number of neighbors of the th point. Choose to be the radius of the neighborhoods. Then, for all inner points , the number of neighbors is a constant, which we denote as . We shall call all points with less than neighbors boundary points. Note that the definition of boundary points depends on the choice of . For inner points of the grid we have . Thus, when we have .
We define . Note that , and for , . In this section we analyze the embedding instead of , thereby avoiding the dependence on the matrix and hence simplifying the notation. This simplification does not significantly change the problem and does not affect the results we present. Similar results are obtained in the next section for general twodimensional manifolds, using the exact normalization constraints (see Section 5.2).
Note that can be described as the set of points , where . The constants and ensure that the normalization constraint holds. Straightforward computation (see Appendix A.3) shows that
(6) 
The definition of the embedding is as follows:
(7) 
where ensures that , and (the same as before; see below) and are chosen so that sample variance of and is equal to one. The symmetry of about the origin implies that , hence the normalization constraint holds. is as defined in Eq. 6, since (with both defined similarly to ). Finally, note that the definition of does not depend on .
4.2 Main result for LEM on the twodimensional grid
We estimate by (see Eq. 4), where is an inner point of the grid and is the neighborhood of ; likewise, we estimate by for an inner point . For all inner points, the value of is equal to some value . For boundary points, is bounded by multiplied by some constant that depends only on the number of neighbors. Hence, for large and , the difference between and is negligible.
The main result of this section states:
Theorem 4.1
Let be an inner point and let the ratio be greater than . Then
for neighborhoodradius that satisfies , or similarly, for nearest neighborhoods where .
This indicates that for aspect ratios that are greater than and above, mapping , which is essentially onedimensional, is preferred to , which is a linear transformation of the grid. The case of general ball neighborhoods is discussed in Appendix A.4 and indicates that similar results should be expected.
The proof of the theorem is as follows. It can be shown analytically (see Fig. 3) that
(8) 
where
(9) 
For higher , can be approximated for any ball neighborhood of (see Appendix A.4).
4.3 Implications to other algorithms
We start with implications regarding DFM. There are two main differences between LEM and DFM. The first difference is the choice of the kernel. LEM chooses , which can be referred to as the “window” kernel (a Gaussian weight function was also considered by Belkin and Niyogi, 2003). DFM allows a more general rotationinvariant kernel, which includes the “window” kernel of LEM. The second difference is that DFM renormalizes the weights (see Eq. 2). However, for all the inner points of the grid with neighbors that are also inner points, the renormalization factor is a constant. Therefore, if DFM chooses the “window” kernel, it is expected to fail, like LEM. In other words, when DFM using the “window” kernel is applied to a grid with aspect ratio slightly greater than or above, DFM will prefer the embedding over the embedding (see Fig 2). For a more general choice of kernel, the discussion in Appendix A.4 indicates that a similar failure should occur. This is because the relation between the estimations of and presented in Eqs. 8 and 10 holds for any rotationinvariant kernel (see Appendix A.4). This observation is also evident in numerical examples, as shown in Figs. 4 and 5.
In the cases of LLE with no regularization, LTSA, and HLLE, it can be shown that . Indeed, for LTSA and HLLE, the weight matrix projects on a space that is perpendicular to the SVD of the neighborhood , thus . Since , we have , and, therefore, . For the case of LLE with no regularization, when , each point can be reconstructed perfectly from its neighbors, and the result follows. Hence, a linear transformation of the original data should be the preferred output. However, the fact that relies heavily on the assumption that both the input and the output are of the same dimension (see Theorem 5.1 for manifolds embedded in higher dimensions), which is typically not the case in dimensionreducing applications.
4.4 Numerical results
For the following numerical results, we used the Matlab implementation written by the respective algorithms’ authors as provided by Wittman (retrieved Jan. 2007) (a minor correction was applied to the code of HLLE).
We ran the LEM algorithm on data sets with aspect ratios above and below . We present results for both a grid and a uniformly sampled strip. The neighborhoods were chosen using nearest neighbors with , and . We present the results for ; the results for , and are similar. The results for the grid and the random sample are presented in Figs. 4 and 5, respectively.
We ran the DFM algorithm on the same data sets. We used the normalization constant and the kernel width ; the results for , and are similar. The results for the grid and the random sample are presented in Figures 4 and 5, respectively.
Both examples clearly demonstrate that for aspect ratios sufficiently greater than , both LEM and DFM prefer a solution that collapses the input data to a nearly onedimensional output, normalized in . This is exactly as expected, based on our theoretical arguments.
Finally, we ran LLE, HLLE, and LTSA on the same data sets. In the case of the grid, both LLE and LTSA (roughly) recovered the grid shape for , and , while HLLE failed to produce any output due to large memory requirements. In the case of the random sample, both LLE and HLLE succeeded for but failed for . LTSA succeeded for , and but failed for . The reasons for the failure for lower values of are not clear, but may be due to roundoff errors. In the case of LLE, the failure may also be related to the use of regularization in LLE’s second step.
5 Analysis for general twodimensional manifolds
The aim of this section is to present necessary conditions for the success of the normalizedoutput algorithms on general twodimensional manifolds embedded in highdimensional space. We show how this result can be further generalized to manifolds of higher dimension. We demonstrate the theoretical results using numerical examples.
5.1 Two different embeddings for a twodimensional manifold
We start with some definitions. Let be the original sample. Without loss of generality, we assume that
As in Section 4, we assume that . Assume that the input for the normalizedoutput algorithms is given by where is a smooth function and is the dimension of the input. When the mapping is an isometry, we expect to be small. We now take a close look at .
where is the th column of the neighborhood . Define , and note that depends on the different algorithms through the definition of the matrices . The quantity is the portion of error obtained by using the th column of the th neighborhood when using the original sample as output. Denote , the average error originating from the th column.
We define two different embeddings for , following the logic of Sec. 4.1. Let
(13) 
be the first embedding. Note that is just the original sample up to a linear transformation that ensures that the normalization constraints and hold. Moreover, is the only transformation of that satisfies these conditions, which are the normalization constraints for LLE, HLLE, and LTSA. In Section 5.2 we discuss the modified embeddings for LEM and DFM.
The second embedding, , is given by
(14) 
Here
(15) 
ensures that , and and are chosen so that the sample mean and variance of are equal to zero and one, respectively. We assume without loss of generality that .
Note that depends only on the first column of . Moreover, each point is just a linear transformation of . In the case of neighborhoods , the situation can be different. If the first column of is either nonnegative or nonpositive, then is indeed a linear transformation of . However, if is located on both sides of zero, is not a linear transformation of . Denote by the set of indices of neighborhoods that are not linear transformations of . The number depends on the number of nearest neighbors . Recall that for each neighborhood, we defined the radius . Define to be the maximum radius of neighborhoods , such that .
5.2 The embeddings for LEM and DFM
So far we have claimed that given the original sample , we expect the output to resemble (see Eq. 13). However, does not satisfy the normalization constraints of Eq. 5 for the cases of LEM and DFM. Define to be the only affine transformation of (up to rotation) that satisfies the normalization constraint of LEM and DFM. When the original sample is given by , we expect the output of LEM and DFM to resemble . We note that unlike the matrix that was defined in terms of the matrix only, depends also on the choice of neighborhoods through the matrix that appears in the normalization constraints.
We define more formally. Denote . Note that is just a translation of that ensures that . The matrix is positive definite and therefore can be presented by where is a orthogonal matrix and
where . Define ; then is the only affine transformation of that satisfies the normalization constraints of LEM and DFM; namely, we have and .
A similar analysis to that of and can be performed for and . The same necessary conditions for success are obtained, with , , and replaced by , , and , respectively. In the case where the distribution of the original points is uniform, the ratio is close to the ratio and thus the necessary conditions for the success of LEM and DFM are similar to the conditions in Corollary 5.2.
5.3 Characterization of the embeddings
The main result of this section provides necessary conditions for the success of the normalizedoutput algorithms. Following Section 3, we say that the algorithms fail if , where and are defined in Eqs. 13 and 14, respectively. Thus, a necessary condition for the success of the normalizedoutput algorithms is that .
Theorem 5.1
Let be a sample from a twodimensional domain and let be its embedding in highdimensional space. Let and be defined as above. Then
where is a constant that depends on the specific algorithm. For the algorithms LEM and DFM a more restrictive condition can be defined:
For the proof, see Appendix A.6.
Adding some assumptions, we can obtain a simpler criterion. First note that, in general, and
should be of the same order, since it can be assumed that, locally, the neighborhoods are uniformly distributed. Second, following Lemma
A.2 (see Appendix A.8), when is a sample from a symmetric unimodal distribution it can be assumed that and . Then we have the following corollary:Corollary 5.2
Let be as in Theorem 5.1. Let be the ratio between the variance of the first and second columns of . Assume that , , and . Then
For LEM and DFM, we can write
We emphasize that both Theorem 5.1 and Corollary 5.2 do not state that is the output of the normalizedoutput algorithms. However, when the difference between the right side and the left side of the inequalities is large, one cannot expect the output to resemble the original sample (see Lemma 3.1). In these cases we say that the algorithms fail to recover the structure of the original domain.
5.4 Generalization of the results to manifolds of higher dimensions
The discussion above introduced necessary conditions for the normalizedoutput algorithms’ success on twodimensional manifolds embedded in . Necessary conditions for success on general dimensional manifolds, , can also be obtained. We present here a simple criterion to demonstrate the fact that there are dimensional manifolds that the normalizedoutput algorithms cannot recover.
Let be a sample from a dimensional domain. Assume that the input for the normalizedoutput algorithms is given by where is a smooth function and is the dimension of the input. We assume without loss of generality that and that is a diagonal matrix. Let . We define the matrix as follows. The first column of , , equals the first column of , namely, . We define the second column similarly to the definition in Eq. 14:
where is defined as in Eq. 15, and and are chosen so that the sample mean and variance of are equal to zero and one, respectively. We define the next columns of by
where . Note that and . Denote .
We bound from above:
Since we may write , we have
When the sample is taken from a symmetric distribution with respect to the axes, one can expect to be small. In the specific case of the dimensional grid, . Indeed, is symmetric around zero, and all values of appear for a given value of . Hence, both LEM and DFM are expected to fail whenever the ratio between the length of the grid in the first and second coordinates is slightly greater than or more, regardless of the length of grid in the other coordinates, similar to the result presented in Theorem 4.1. Corresponding results for the other normalizedoutput algorithms can also be obtained, similar to the derivation of Corollary 5.2.
5.5 Numerical results
We ran all five normalizedoutput algorithms, along with Isomap, on three data sets. We used the Matlab implementations written by the algorithms’ authors as provided by Wittman (retrieved Jan. 2007).
The first data set is a point sample from the swissroll as obtained from Wittman (retrieved Jan. 2007). The results for the swissroll are given in Fig. 7, A1F1. The results for the same swissroll, after its first dimension was stretched by a factor , are given in Fig. 7, A2F2. The original and stretched swissrolls are presented in Fig. 6A. The results for are given in Fig. 7. We also checked for ; but “shortcircuits” occur (see Balasubramanian et al., 2002, for a definition and discussion of “shortcircuits”).
The second data set consists of points, uniformly sampled from a “fishbowl”, where a “fishbowl” is a twodimensional sphere minus a neighborhood of the northern pole (see Fig. 6B for both the “fishbowl” and its stretched version). The results for are given in Fig. 8. We also checked for ; the results are roughly similar. Note that the “fishbowl” is a twodimensional manifold embedded in , which is not an isometry. While our theoretical results were proved under the assumption of isometry, this example shows that the normalizedoutput algorithms prefer to collapse their output even in more general settings.
The third data set consists of images of the globe, each of pixels (see Fig. 6C). The images, provided by Hamm et al. (2005), were obtained by changing the globe’s azimuthal and elevation angles. The parameters of the variations are given by a array that contains to degrees of azimuth and to degrees of elevation. We checked the algorithms both on the entire set of images and on a strip of angular variations. The results for are given in Fig. 9. We also checked for ; the results are roughly similar.
6 Asymptotics
In the previous sections we analyzed the phenomenon of global distortion obtained by the normalizedoutput algorithms on a finite input sample. However, it is of great importance to explore the limit behavior of the algorithms, i.e., the behavior when the number of input points tends to infinity. We consider the question of convergence in the case of input that consists of a dimensional manifold embedded in , where the manifold is isometric to a convex subset of Euclidean space. By convergence we mean recovering the original subset of up to a nonsingular affine transformation.
Some previous theoretical works presented results related to the convergence issue. Huo and Smith (2006) proved convergence of LTSA under some conditions. However, to the best of our knowledge, no proof or contradiction of convergence has been given for any other of the normalizedoutput algorithms. In this section we discuss the various algorithms separately. We also discuss the influence of noise on the convergence. Using the results from previous sections, we show that there are classes of manifolds on which the normalizedoutput algorithms cannot be expected to recover the original sample, not even asymptotically.
6.1 LEM and DFM
Let be a uniform sample from the twodimensional strip . Let be the sample of size . Let be the number of nearest neighbors. Then when
there exists with probability one an
, such that for all the assumptions of Corollary 5.2 hold. Thus, if we do not expect either LEM or DFM to recover the structure of the strip as the number of points in the sample tends to infinity. Note that this result does not depend on the number of neighbors or the width of the kernel, which can be changed as a function of the number of points , as long as . Hence, we conclude that LEM and DFM generally do not converge, regardless of the choice of parameters.In the rest of this subsection we present further explanations regarding the failure of LEM and DFM based on the asymptotic behavior of the graph Laplacian (see Belkin and Niyogi, 2003, for details)
. Although it was not mentioned explicitly in this paper, it is well known that the outputs of LEM and DFM are highly related to the lower nonnegative eigenvectors of the normalized graph Laplacian matrix (see Appendix
A.1). It was shown by Belkin and Niyogi (2005), Hein et al. (2005), and Singer (2006)that the graph Laplacian operator converges to the continuous Laplacian operator. Thus, taking a close look at the eigenfunctions of the continuous Laplacian operator may reveal some additional insight into the behavior of both LEM and DFM.
Comments
There are no comments yet.