Structured Semidefinite Programming for Recovering Structured Preconditioners

We develop a general framework for finding approximately-optimal preconditioners for solving linear systems. Leveraging this framework we obtain improved runtimes for fundamental preconditioning and linear system solving problems including the following. We give an algorithm which, given positive definite $\mathbf{K} \in \mathbb{R}^{d \times d}$ with $\mathrm{nnz}(\mathbf{K})$ nonzero entries, computes an $\epsilon$-optimal diagonal preconditioner in time $\widetilde{O}(\mathrm{nnz}(\mathbf{K}) \cdot \mathrm{poly}(\kappa^\star,\epsilon^{-1}))$, where $\kappa^\star$ is the optimal condition number of the rescaled matrix. We give an algorithm which, given $\mathbf{M} \in \mathbb{R}^{d \times d}$ that is either the pseudoinverse of a graph Laplacian matrix or a constant spectral approximation of one, solves linear systems in $\mathbf{M}$ in $\widetilde{O}(d^2)$ time. Our diagonal preconditioning results improve state-of-the-art runtimes of $\Omega(d^{3.5})$ attained by general-purpose semidefinite programming, and our solvers improve state-of-the-art runtimes of $\Omega(d^{\omega})$ where $\omega>2.3$ is the current matrix multiplication constant. We attain our results via new algorithms for a class of semidefinite programs (SDPs) we call matrix-dictionary approximation SDPs, which we leverage to solve an associated problem we call matrix-dictionary recovery.

Paper

Similar papers

Peer review

Reviewer 2VGS6/10 · confidence 3/52023-06-26

Summary

This paper develops a general preconditioning framework called matrix-dictionary recovery. This framework follows the matrix multiplicative weights update method and is applied to solve two classes of problems: (1) Two diagonal preconditioning problems: outer scaling and inner scaling This paper gives the first nontrivial approximation algorithms in nearly linear time, which is beyond the generic SDP solvers. (2) Graph-structured matrix families This paper proposes algorithms for perturbed Laplacian solver and the recovery of two structured matrices: M-matrix and Laplacian matrix.

Strengths

(1) The proposed algorithm for computing diagonal preconditioners (Theorem 1 and Theorem 2) improves the method using general SDP solvers by $\text{poly}(d)$ factors when $\kappa_o^\star (\mathbf{K})$ and $\kappa_i^\star (\mathbf{A})$ are small. (2) The proposed framework for solving graph-structured matrices (Theorem 3, 4, and 5) obtains $\widetilde{O}(n^2)$ running time and improves the state-of-the-art methods which have running time $O(n^\omega)$ by virtue of general linear systems solvers. (2) This paper is well-written and easy to understand.

Weaknesses

(1) The algorithms for diagonal preconditioning improve the existing method by a factor of $\text{poly}(d)$ at the cost of $(\kappa^\star)^{1.5}$. The applicable scenario is when $\kappa^\star$ is small but $\kappa$ is large. (2) The algorithms for solving graph-structured matrices have time complexity $\widetilde{O}(n^2)$ that improves $O(n^\omega)$. They work for the case when the input is dense but in practice, the input is often sparse. (3) Section 5 is placed at the end of the manuscript. It would be better to show the comparison with existing work in the Introduction part so that we can clearly know your contributions and advantages.

Questions

The proposed method for graph-structured matrices has running time $\widetilde{O}(n^2)$ which is nearly linear when the input is dense. (1) Can it be further improved when the input is sparse? (2) The existing work has complexity $O(n^\omega)$. Is it still $O(n^\omega)$ when the input is sparse?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See Weaknesses.

Reviewer pK6D8/10 · confidence 3/52023-06-29

Summary

This paper studies preconditioning, which is one of the most important techniques in numerical linear algebra with numerous applications in optimization and machine learning. It proposes a general framework based on the matrix-dictionary recovery problem, where are given a matrix $M$ and $M_1,\dots, M_n$, and the goal is to find $w$ such that $\sum_{i\in [n]}w_i M_i$ approximates the optimal preconditioner for $M$. This paper develops a general-purpose solver for this problem. Then, they apply this framework to propose nearly linear-time algorithms for diagonal preconditioning (including outer scaling and inner scaling), semi-random regression, and structured liner systems (including dense matrix approximated by graph Laplacians, dense inverse Laplacians, and dense inverse M-matrices). Technically, their meta-solver for the matrix-dictionary recovery problem is based on the matrix multiplicative weights update method and employs packing SDP solvers to form the gain matrices. They also use other techniques, such as JL sketching, the homotopy method, etc., to implement their solvers.

Strengths

This research question considered in this paper is crucial for many applications in both theory and practice. Their results show that a bunch of numerical linear algebra problems can be solved in nearly linear time, which was not known before. More specifically, for the diagonal preconditioning problem, prior to this work, we did not know how to $O(1)$-approximate the optimal condition number of the rescaled matrix without using a generic SDP solver, which takes $\Omega(d^{3.5})$-time. They improve it to a linear dependence on the number of non-zero entries of the matrix ($\leq d^2$). In addition, they also provide a tighter analysis of the Jacobi preconditioning method. For numerically solving linear systems, when the matrix is a perturbed Laplacians, prior to this work, the only approach is to apply generic matrix multiplication, which runs in $d^{\omega}$-time. Their new algorithm runs in about $d^2$-time when the matrix is well-approximated by some unknown Laplacian. Moreover, their algorithm also generalizes to larger families of dense matrices, which are very useful in many different applications. The framework of matrix dictionary recovery is also very interesting and useful. There are some previous results concerning the isotropic case with some restrictions on the condition number. In this paper, their solver works not only for the non-isotropic case but also for general condition numbers. This framework and the general-purpose solver will be very helpful in future studies of numerical linear algebra algorithms.

Weaknesses

The running times of their preconditioning algorithms depend on $\kappa^{1.5}$, which may not be optimal. Moreover, the accuracy dependence is $poly(1/\epsilon)$. Also, the paper does not discuss the practical implications of their algorithms, e.g., is it possible to implement them in practice?

Questions

In Algorithm 1, it is required that $\lambda_{\max}(M_i)\in [1,2]$. Is it just for convenience? What is the runtime bottleneck of Algorithm 1?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

The limitations are addressed.

Reviewer 1BXc3/10 · confidence 1/52023-06-30

Summary

In this paper, a framework is presented to compute approximately-optimal preconditioners in order to solve linear systems. In the case of diagonal preconditioning, an algorithm is provided whose runtime is (up to log factors) polynomial in the desired accuracy, optimal condition number of the re-scaled matrix (thus in the output conditioning) and the number of nonzero entries of the input matrix. This is explicitly stated in Theorem 1 and Theorem 2 for outer and inner scaling, respectively. In the case of structured linear systems, an algorithm is provided whose runtime is (up to log factors) quadratic in the size of the input matrix. The underlying algorithms are based on the so-called "matrix-dictionary" approximation semidefinite programs. They rely on these algorithms to solve a related problem called "matrix-dictionary recovery".

Strengths

If they would be correct, the resulting algorithms improve current runtime of the best solvers given by SDP (by a factor of d sqrt(d)). The comparison with existing literature seems well-explained.

Weaknesses

Not being an expert in this field and with a very limited amount of time (too short to read the 63 pages of supplementary material), it is completely impossible to judge whether the framework is correct or not. The authors just state the theorems and do not provide neither proof sketches nor hints that could possibly convince the reader. I believe that this work, possibly sound and surely interesting, is not a good fit at all, considering the currently requested NeurIPS format.

Questions

l43: What is an M-matrix? The definition should be recalled.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

There is no limitation section provided by the authors in the main submission.

Reviewer HjkM7/10 · confidence 3/52023-07-07

Summary

This is a theoretical paper that studies the problem of diagonal matrix preconditioning, where given a PSD matrix $A$, the goal is to find a (positive) diagonal scaling $W$, such that $WAW$ has a small condition number, given the promise that such scaling exists. This problem can be solved using SDP, but that is too expensive. Following previous work that uses spectral sparsification techniques, the authors give improved guarantees for this problem. In particular, previously it was only known how to compute an approximately optimal scaling when the resulting condition number is $\leq 1.1$, while the authors' result works for any condition number (and has a polynomial condition number dependency in the runtime). The authors use the matrix multiplicative weights algorithm, together with sketching techniques to efficiently implement its iterations, leading to a near-linear time algorithm for bounded condition number. The results can be generalized to other problem classes like graph-structured preconditioning.

Strengths

- Preconditioning is a fundamental problem and this is a significant and original contribution to it. It is significantly stronger than previously known results. - The analysis looks correct and technically strong. - The presentation is thorough and the intro is well written.

Weaknesses

- There is no empirical evaluation of the proposed algorithm, e.g. compared to Jacobi iteration, and so the practical impact of this approach is not clear.

Questions

- Do the results extend to the case when we have both outer and inner preconditioning? - Since the algorithm is based on multiplicative weights, it seems natural to try and generalize this to the online setting. Are there any significant difficulties there? - Are there potential implications for preconditioned optimizers such as Adagrad, Shampoo, etc?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

No

Reviewer HjkM2023-08-18

Thank you very much for the detailed response.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC