Breaking the curse of dimensionality in structured density estimation

We consider the problem of estimating a structured multivariate density, subject to Markov conditions implied by an undirected graph. In the worst case, without Markovian assumptions, this problem suffers from the curse of dimensionality. Our main result shows how the curse of dimensionality can be avoided or greatly alleviated under the Markov property, and applies to arbitrary graphs. While existing results along these lines focus on sparsity or manifold assumptions, we introduce a new graphical quantity called"graph resilience"and show how it controls the sample complexity. Surprisingly, although one might expect the sample complexity of this problem to scale with local graph parameters such as the degree, this turns out not to be the case. Through explicit examples, we compute uniform deviation bounds and illustrate how the curse of dimensionality in density estimation can thus be circumvented. Notable examples where the rate improves substantially include sequential, hierarchical, and spatial data.

Paper

References (70)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer yapz6/10 · confidence 5/52024-07-06

Summary

This paper proposed a new summary of graph, called graph resilience, which measures the complexity of density estimation. This concept is different from more its well-known counterparts, including rank, sparsity, manifold dimension etc. Quite a few concrete examples of graph along with their resilience are given. Both the concept and the theory are sound, the contribution to the literature of density estimation is solid.

Strengths

The concept of resilience is novel. The definition is clear, esp Figure 3. Section 4 is very well-written, which helps the audience a lot to understand the concept of resilience.

Weaknesses

More intuitions are expected, to define $r$. The concept is beautiful, but I am curious how the authors come up with the definition. Is there is any clear intuition to define it, or the authors start from the proof of the density estimation rate and then identify this key quantity?

Questions

Given an arbitrary graph, is there any concrete computational algorithm to calculate the exact resilience? If so, what’s the computational cost? I suppose the exact calculation is very expensive, that’s why the authors focused more on how to bound it.

Rating

6

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

Can the authors discuss some possible extensions? For example, what does the resilience look like for directed graphs, or for non-Markovian graphs?

Reviewer HdwX6/10 · confidence 3/52024-07-10

Summary

The paper studies the problem of estimating a multivariate density, assuming that the density is Markov w.r.t an underlying undirected graph $G$. It is shown that the sample complexity for estimating such a density scales with the resilience of the graph, as opposed to the dimension d. Several examples of G are provided for which the resilience can be exactly computed or bounded.

Strengths

1. The paper is written very well with a clear exposition which makes it easy to follow. The problem is well defined and motivated, along with clear notation throughout. The related work is also thorough and provides a good overview of the literature. 2. The results are to my knowledge novel, I am not aware of similar results for this problem setting. I believe the paper would be of interest for researchers in high dimensional statistics, and related problems.

Weaknesses

1. It would have been helpful to provide a proof-sketch of the main theorems in the main text, this is particularly relevant here given that this is a theoretical paper. Moreover, I think it is important to outline the algorithm in the main text and not relegate it to the appendix. To handle space constraints, I think parts of Section 4 could be moved to the appendix as they are relatively less important as opposed to the aforementioned details. 2. It would have also been useful to provide some simulation results on synthetic data as this would have helped empirically validate the theoretical results.

Questions

1. What is the running time of the algorithm w.r.t n, d and r? This is important to discuss in the main text and is related to the point I had raised earlier. 2. The density is assumed to be Lipschitz for the analysis, but can the results be extended to handle higher order smoothness (s times continuous differentiability)? 3. Can something be said about the optimality of the bounds? In particular, is the graph resilience the right quantity for this problem? 4. In Corollary 4.12, why not show the dependence on t explicitly in the bound? Minor remarks: - I think there is a typo towards the end of line 182. Another typos in line 252 ("... tend to have ...") - In the statement of Theorem 3.1, it might be good to specify that $G$ is known.

Rating

6

Confidence

3

Soundness

2

Presentation

4

Contribution

3

Limitations

I do not see a dicussion on limitations in the main text.

Reviewer P9wL5/10 · confidence 4/52024-07-13

Summary

The authors, through the formalism of graph models for probability density functions, identify the “graph resilience” as a key quantity to estimate the sample complexity of computing the density.

Strengths

The work is original and technically sound, claims are generally well supported (with one exception that I will mention later on). The ideas of graphs resilience and disintegration are interesting and are nicely defined.

Weaknesses

The only claim that I consider not well supported is related to lines 160-161, more concretely to the lower dimensional manifold. In order to support this claim, the authors should prove this claim by finding an example where the resilience r is lower than the dimension d and that cannot be mapped into a lower dimensional manifold. The paper is not clearly written and in a way that only specialists in the field of graph modelling can understand. The relationship with classical methods like kernel density estimation is not clear. No practical applications are shown. From certain point of view, the novelty is not high: The authors shown that, for structured data, the Markov property is enough to simplify the complexity of density estimation. However, the structure in the data is, by itself, a correlation measure, and therefore, it is not surprising that it reduces the effect of the curse of dimensionality.

Questions

Why is the manifold assumption violated in structured data if the effective dimension is lower than the embedding dimension? Are you willing to provide a code that allows to use your findings in practical data?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

No limitations are discussed. Aspects like the limited application field and the difficulties when trying to use the findings in real data should be addressed.

Reviewer yapz2024-08-11

Thanks for the response. The intuition is helpful to understand the concept. For the calculation, I understand it's not necessary to calculate it exactly for certain purpose. I asked this question because I'd like to understand the concept deeper, not for any specific (practical) purpose. Can I understand your answer as "we don't have a principled way to exactly calculate it, and we don't know the cost"?

Authorsrebuttal2024-08-13

Response to Reviewer yapz comment: On calculating graph resilience

__”Can I understand your answer as "we don't have a principled way to exactly calculate it, and we don't know the cost"?”__ Correct, we don’t have a general way to calculate it exactly, and we stress that part of our contribution is to emphasize the relevance of this new quantity, so as to inspire future investigations that might provide principled algorithms. We suspect, for example, that greedy methods may work well and give useful bounds. An example of this could be greedily removing the most central vertex in each graph component according to some vertex centrality measure (e.g. vertex degree).

Reviewer yapz2024-08-14

Thanks for the clarification. I suggest the authors to clarify this computational aspect somewhere in the paper, even in the appendix. I'll keep my (positive) score.

Reviewer P9wL2024-08-11

Follow up

I thank the authors for the reply, now things are more clear. Regarding practical applications, I'm still not convinced that a paper with no practical applications is worth publication in Neurips, but before deciding to maintain or raise my score I would like to have a clear reply from the authors. Is there a practical way to take advantage of resilience to compute densities? Can you show that in a real case?

Authorsrebuttal2024-08-13

Response to Reviewer P9wL comment: On practical applications

__”Is there a practical way to take advantage of resilience to compute densities? Can you show that in a real case?”__ Yes, there is a practical way to take advantage of this. This is related to our explanation of graph resilience in our rebuttal to yapz. Graph resilience essentially gives the most efficient disintegration of a graph, and can be thought of as a sort of “meta-algorithm.” The disintegration gives an ordering as to how to estimate conditional densities, however it would be up to a practitioner to decide how to take care of the one-dimensional and conditional density estimation, which are themselves structured in an inductive way. More precisely, a disintegration outlines a way to inductively condition out the entries of a random vector. For instance, consider a random vector $[X_1,X_2]$. The disintegration $(X_1)-(X_2) \Rightarrow (X_1) \Rightarrow \emptyset$ corresponds to first estimating $p_{X_2}$, and then, for each $x$, estimating the density $p_{X_1|X_2=x}$. The Lipschitz condition tells us that $p_{X_1|X_2=x}$ doesn’t change drastically as $x$ changes, thus making the problem tractable. Removing one vertex from __each__ component of a graph captures the idea that, after sufficient conditioning, these components become independent. This allows us to avoid estimating the high-dimensional joint density of all vertices in the graph. Instead, we can estimate the low-dimensional components individually and take their product which is justified due to the following inequality: $\Vert\prod_{i=1}^d p_i - \prod_{j=1}^d q_j \Vert_1 \le \sum_{i=1}^d \Vert p_i - q_i \Vert_1$ For a concrete example, consider the graph $(X)-(Y)-(Z)$, with the disintegration $(X)-(Y)-(Z) \Rightarrow (X)\quad (Z) \Rightarrow \emptyset$. This corresponds to estimating $p_Y$, then estimating $p_{X,Z| Y}$ utilizing the fact that $X$ and $Z$ are independent given $Y$ so $p_{X,Z| Y}(x,z) = p_{X| Y}(x)p_{Z| Y}(z)$ thus instead of having to estimate a two dimensional conditional density, one can simply estimate two one dimensional conditional densities and take the product, thereby avoiding the curse of dimensionality. This example only contains one simple instance of taking advantage of the product structure, but there may be layers of product structure, which is exemplified in our result that the graph resilience of a linear Markov graph grows only logarithmically in dimension. One also need not know the disintegration corresponding to the graph resilience in order to apply this. The resilience simply characterizes the best possible disintegration. While disintegrations are based on estimating many one-dimensional conditional distributions, one could instead estimate low-dimensional distributions. We would be happy to add an appendix (based on this discussion) outlining how a practitioner can take advantage of the disintegration structure in a more practical and readable way.

Reviewer HdwX2024-08-12

Read authors response

Thank you for the clarifications, I am satisfied with them. I am not insistent on the simulations, although I do think it would have strengthened the paper. But adding a proof sketch in the main text is more important given that this is a theory paper. I am increasing my score to 6.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC