Optimization Can Learn Johnson Lindenstrauss Embeddings

Embeddings play a pivotal role across various disciplines, offering compact representations of complex data structures. Randomized methods like Johnson-Lindenstrauss (JL) provide state-of-the-art and essentially unimprovable theoretical guarantees for achieving such representations. These guarantees are worst-case and in particular, neither the analysis, nor the algorithm, takes into account any potential structural information of the data. The natural question is: must we randomize? Could we instead use an optimization-based approach, working directly with the data? A first answer is no: as we show, the distance-preserving objective of JL has a non-convex landscape over the space of projection matrices, with many bad stationary points. But this is not the final answer. We present a novel method motivated by diffusion models, that circumvents this fundamental challenge: rather than performing optimization directly over the space of projection matrices, we use optimization over the larger space of random solution samplers, gradually reducing the variance of the sampler. We show that by moving through this larger space, our objective converges to a deterministic (zero variance) solution, avoiding bad stationary points. This method can also be seen as an optimization-based derandomization approach and is an idea and method that we believe can be applied to many other problems.

Paper

Similar papers

Peer review

Reviewer WAK98/10 · confidence 3/52024-06-17

Summary

This work shows that a deterministic optimization procedure can find a matrix $A$ that satisfies the Johnson Lindenstrauss guarantee. That is, a matrix $A$ maps a set of $n$ vectors to a lower dimensional space while preserving all pairwise distances up to some chosen multiplicative distortion. Typically, $A$ is constructed by sampling it from a random matrix distribution with i.i.d. entries. The authors prove that attempting to directly optimize the entries of $A$ through an optimization procedure by minimizing the maximum distortion is prone to being stuck at local minima. However, the authors show that by optimizing the mean of each entry and the entry-wise variance of the distribution $A$ is sampled from, one can maintain a fixed probability of $A$ being a JL-embedding while at the same time guaranteeing the entry-wise variance ‘sigma’ goes to zero. They then show that, when ‘sigma’ is sufficiently small, one may use the optimized expectation of $A$ as the embedding matrix while only slight increasing the maximum distortion, thereby deterministically finding the desired JL embedding matrix $A$. They show that $\rho$-SOSPs (second order stationary points) have sufficiently low variance when $\rho$ is small, and finally show that a method for finding $\rho$-SOSPs suffices to solve the designed optimization problem.

Strengths

Overall, the paper is clearly written and well-motivated. The intuition of the approach and analysis is easy to follow. The key idea of optimizing the parameters of a random matrix distribution to preserve the JL-embedding property while reducing the entry-wise variance seems like an innovative approach. The authors point out the original space of matrices is contained in this larger probabilistic space, since a deterministic matrix $A$ is equivalent to having mean $A$ and zero variance. Hence, this can be seen as a probabilistic relaxation of the original matrix optimization problem. I have not seen this type of relaxation used in the field of matrix sketching or more generally randomized numerical linear algebra before, and I believe it may be useful for other problems in the area. I am not very familiar with diffusion models, so I cannot speak on the novelty of the approach regarding that area. The empirical results are also strong in the sense that they show this procedure for constructing a JL embedding tends to achieve a much lower distortion factor than randomized constructions for a fixed dimension.

Weaknesses

The iterative method to find the matrix $A$ takes $\operatorname{poly}(n, k, d)$ steps, i.e., the complexity is proven to be polynomial but not explicitly determined. Since the paper is primarily theoretical with only limited experiments, it is unclear how efficient this method is in practice. While the results seem very interesting theoretically, the paper could be strengthened by pointing out some practical applications where this improved deterministic JL embedding would be useful. In the applications I am familiar with, oblivious JL embeddings are needed due to the large number of points in the high-dimensional space (e.g., preserving k-means loss). The authors point to embeddings in deep learning as motivation. It is unclear to me as to how the authors expect progress in understanding deterministic JL embeddings to relate to these embeddings in deep learning. Additional clarification of this point would be helpful.

Questions

In the conclusion, you mention the potential for this approach in applications beyond the Johnson Lindenstrauss setting. In your approach for the JL setting, you upper bound the failure probability of the distortion guarantee via the union bound in eqn. (3). This formulation of the objective function seems difficult to translate to other sketching guarantees on $A$ (e.g., projection cost preservation, L2-subspace embedding, affine embedding). Is there any intuitive reason why it may be possible to formulate a relaxed differentiable objective function when the embedding guarantee must hold over an uncountable number of points? How does learning a JL embedding relate to learning embeddings for application areas discussed in the introduction? In particular, how do you see the results of this paper affecting that line of work? As mentioned above, I think it would be helpful to expand on the link between your result and the motivation of deep learning embeddings given in the intro.

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

Yes

Reviewer XscG5/10 · confidence 3/52024-07-11

Summary

The paper proposes to calculate the embedding matrices used in the statement of the Johnson-Lindenstrauss lemma using optimization instead of randomization. The proposed algorithm is a Hessian descent. Authors prove that the algorithm finds the matrix of minimum distortion. Numerical results display the findings

Strengths

JL is a well celebrated result used for proving existence of optimal (low distortion) embeddings. It is stated in the formal result of JL that such embeddings can be found in polynomial time. But we often rely on randomization to exhibit them. It is useful to have an algorithm to calculate the embeddings. The paper tackles a well motivated problem and their presentation is clear and clean.

Weaknesses

The paper lacks complexity analysis of the algorithm. The algorithm proposed requires a full eigenvalue decomposition at every step. It is prohibitive to use this method in any practical scenario. A discussion on the complexity and how to scale the method up (using randomized methods??) would be nice.

Questions

The paper's main claim is that mean distortion embeddings are computationally well studied: spectral methods (SVD / eigenvalue) methods calculate those. Authors claim that the min (instead of mean) distortion method is what they want to find. Can authors explain why the relaxation of f* to f using the probability bound in Eq (2) and invoking union bound to obtain (3) does not reduce the max to a sum. The entire promise of the method is to work with the max directly to minimize distortion. It appears that relaxation of max to a sum using the union bound drops the nice property which was the primary motivation of the work. Can you explain what I am missing of misinterpreting?

Rating

5

Confidence

3

Soundness

2

Presentation

4

Contribution

3

Limitations

scaling the method up and a clear argument why relaxation of Eq (3) is not making the problem trivial would impove the paper.

Authorsrebuttal2024-08-13

Thanks for the response. We answer your follow-up questions below. **Q: is there a reason why a first-order optimizer would not work and we actually need a second-order (more expensive) optimizer? First order optimizers are not necessarily stochastic / randomized.** A simple answer to this is that second-order optimization is necessary because, without it, using first-order optimization **without any randomization** results in the optimization getting stuck at the initial point. Generally, to avoid getting stuck throughout the optimization process, you either need second-order information or first-order information combined with randomization. Stochasticity ensures that progress can always be made. **Q: If we remove the $\sigma^2$ term from Equation (4), what estimator do we get? is it not SVD?** If we remove the $\sigma^2$ term from Equation (4), we would not obtain SVD. Equation (4) provides a bound on the probability of generating a matrix that achieves **worst-case distortion** larger than a specified threshold $\epsilon$. In contrast, PCA, which relies on SVD, finds an embedding that minimizes **average distortion** of distances between the original and the embedded vectors. By fixing (or as you mentioned, removing) $\sigma^2$, we would essentially be looking to optimize the mean matrix $M$ for that specific $\sigma^2$. However, completely removing $\sigma^2$ from the optimization process would create an undesirable situation. While we might obtain a “better” mean matrix $M$, we would lose the intended derandomization at the end of the optimization, effectively resulting in $\sigma^2 = 0$.

Reviewer XscG2024-08-13

can you explain this? > second-order optimization is necessary because, without it, using first-order optimization without any randomization results in the optimization getting stuck at the initial point." is this an empirical finding on the numerical experiments you ran? did you try a larger learning rate? or is it theoretically proven and related to the objective function that we are optimizing? > Generally, to avoid getting stuck throughout the optimization process, you either need second-order information or first-order information combined with randomization. Stochasticity ensures that progress can always be made. Can you share a reference or a more formal statement on this?

Authorsrebuttal2024-08-13

**Q: [...] or is it theoretically proven and related to the objective function that we are optimizing?** A: It is related to the objective function and theoretically proven, so changing the learning rate won’t resolve it. More generally, to avoid getting stuck at any point during the optimization process, we require either a second-order method or a first-order method with some **randomness**. We expand further on this in the next question. **Q: Can you share a reference or a more formal statement on this?** A: Our main result is that the second-order stationary points of Equation (4) satisfy the JL guarantee (Theorem 2). To achieve this, we employ a deterministic second-order method, as detailed in our paper, and thus we achieve a **truly derandomized construction** of a JL matrix. Additionally, it has been shown that first-order methods with randomness also converge to second-order stationary points with high probability, see for example the work of Jin et al. (2017) on saddle points. Thus, whether using a deterministic second-order method or a first-order method with some degree of randomness, we can reach a second-order stationary point, which is precisely what is required to ensure the JL guarantee. We hope this clarifies your questions.

Reviewer y6Yv5/10 · confidence 3/52024-07-11

Summary

The paper considers using the optimization method to "learn" the Johnson Lindenstrauss transform. The paper first shows that the naive objective may not be good enough -- there are stationary points that are sub-optimal. Instead, they consider the way that optimize the random Gaussian space rather than the concrete matrix. Then the authors give an optimization method and show that using this way, every stationary point is a good point and claim that this gives way to deterministic learns the JL matrix. Finally, the paper gives an experiment that shows the advantages of the proposed method.

Strengths

The theoretical analysis of this paper is very interesting. To my knowledge, there are very few results about analyzing the landscape of the learned sketching matrix and this paper gives a strong analysis. The experiments also show the advantages of the proposed method. The presentation of the paper is also good.

Weaknesses

I am still confused about some parts of the paper. I can raise the score if the authors can adjust this. (see the below question)

Questions

1. I am still a little confused about the main conclusion of the results of this paper. That is -- it gives a deterministic construction of the JL lemma, or it gives a better optimization way and it works well empirically? (as the author mentioned, the bound of the JL lemma can not be improved) 2. The equation (4) is about probability, and B.1 says they use Monte Carlo sampling, however, would it means that the proposed method still contains some randomness part? 3. In the experiment section, the paper compares the proposed method with the JL lemma. It will make this stronger if the comparison with equation (1) is also given.

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

See above questions.

Reviewer pTrj7/10 · confidence 3/52024-07-19

Summary

This paper investigates the problem of using optimization-based approaches to learn Johnson Lindenstrauss(JL) embedding. The authors proposed a new framework to achieve the JL guarantee via optimization, instead of the traditional randomized methods. Similar with diffusion models, the authors proposed a novel method that uses optimization in an extended space of random Gaussian solution samplers, which circumvents direct optimization in non-convex landscape. Their approach uses second-order descent, gradually reduces the variance without increasing the expected distortion of the sampler, then can identify a specific projection matrix with the Gaussian distribution. Overall, theoretical guarantees and empirical results demonstrate that this method efficiently achieves a deterministic solution that satisfies the JL guarantee.

Strengths

The paper is well-written. The state of the art is well discussed by an extensive literature review. The proposed method combining optimized-based approaches and Johnson Lindenstrauss embeddings is an innovative contribution to the field. The paper is technically sound, provides rigorous theoretical analysis and proofs.

Weaknesses

It would be helpful to understand the main results if section 4 could be more organized, such as using subsections.

Questions

1. Could you explain the statement in lines 215-216? Are the values of 1/(3n) and 1/3 derived based on the chosen value of $\epsilon$? 2. Line 336, there is a typo “appplicability”. 3. Regarding the notation $\rho$-second-order stationary points($\rho$-SOSP), the paper uses $\rho$-second-order stationary points in some sections and uses $\rho$-SOSP in others.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

4

Limitations

No potential negative societal impact.

Reviewer WAK92024-08-10

Thanks to the authors for their reply. I will maintain my score. > Is there any intuitive reason why it may be possible to formulate a relaxed differentiable objective function when the embedding guarantee must hold over an uncountable number of points? I agree that this paper is valuable for raising the possibility of the general approach you've described being used to achieve deterministic optimization based algorithms for these other sketching guarantees. However, doing so indeed seems quite challenging at first glance. > How does learning a JL embedding relate to learning embeddings for application areas discussed in the introduction? In particular, how do you see the results of this paper affecting that line of work? This is a very interesting interpretation of the results in relation to deep learning. Thanks for clarifying.

Reviewer XscG2024-08-11

thanks for the rebuttal. Follow-up questions: * is there a reason why a first-order optimizer would not work and we actually need a second-order (more expensive) optimizer? First order optimizers are not necessarily stochastic / randomized. * If we remove the $\sigma^2$ term from Equation (4), what estimator do we get? is it not SVD?

Reviewer y6Yv2024-08-12

Thanks for the response. - Derandomization: Thanks for pointing this out. However, the calculation of the closed form of expression (4) seems to be non-trivial and some discussion about the computation should be included to support the claim (e.g., the time complexity). This is the main reason I keep my score. - Experiment: I agree that "optimizing Equation 1 directly in the matrix space can lead to bad solutions with high distortion" However, it was shown in the specific hard instance and the comparison over the real-world datasets will still be interesting.

Authorsrebuttal2024-08-13

Thank you for your response. Regarding the first point, it's important to note that the calculations of the gradient involve **only** cumulative distribution functions of known distributions, i.e. the chi-squared distribution, which simplifies the computations. We acknowledge that this can be explained more clearly and we will incorporate a discussion on the computational aspects in our work. We appreciate all your valuable feedback and will incorporate it to strengthen our paper. Thank you for your insightful comments and suggestions.

Reviewer pTrj2024-08-13

Thanks for the response. I will maintain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC