REVIEW 5 major objections 6 minor 39 references
On Squared-Variable Formulations for Nonlinear Semidefinite programming
T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read This paper proves that a second-order necessary point of the squared-variable reformulation is exactly a second-order necessary point of the original nonlinear semidefinite program when the factor is square.
desk verdict A genuinely new and clean equivalence: 2NP of the nonsymmetric squared-variable reformulation maps exactly to 2NP of the original nonlinear SDP, without transversality or strict complementarity, and the paper deserves a serious referee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The argument is carried by two trace lemmas (Lemmas 2.1 and 2.2) that connect curvature in the factor space to curvature in the matrix space. For a direction $\Delta$ in $F$-space, the corresponding direction in $X$-space is $W = F\Delta^\top + \Delta F^\top$; the lemmas relate the term $\operatorname{tr}(\nabla h(X)\Delta\Delta^\top)$ to $\operatorname{tr}(W X^\dagger W \nabla h(X))$, using the pseudoinverse $X^\dagger$ and a null-space basis $V$ of $X$. Lemma 2.2 constructs, for every $W$ in the relevant subspace, a $\Delta$ that realizes $W$ while making the trace correction vanish, and it is this construction that removes the strict-complementarity assumption needed in earlier treatments.
What would settle it
Theorems 3.1 and 2.1 assert that no example exists of a square-factor second-order point whose projected matrix point fails the weak second-order condition. A single smooth instance with $F\in\mathbb{R}^{d\times d}$ satisfying (SSV-2NC) whose $(x,\Lambda)$ violates (NSDP-2NC) would falsify the main claim; the paper's own Example 3.1 already supplies exactly this failure for the symmetric variant, so the decisive experiment is to check whether the same construction can succeed with a nonsymmetric factor.
Extended reading notes
Core claim
Formally, Theorem 3.1 states that if $(x,\Lambda)$ satisfies the paper's weak second-order necessary condition (2NC) for (NSDP), then any $F$ with $C(x)=F F^\top$ makes $(x,F,\Lambda)$ satisfy 2NC for (SSV); conversely, every 2NC point $(x,F,\Lambda)$ of (SSV) gives a 2NC point $(x,\Lambda)$ of (NSDP). Theorem 2.1 proves the analogous equivalence between (BC) and (DSS): a matrix $X$ is a 2NP of the semidefinite-constrained problem exactly when its square factors are 2NPs of the unconstrained factored problem. The paper also shows local minimizers correspond in both directions for the nonsymmetric factorization, while strict local minimizers generically do not exist in the factored problem because $F$ can be rotated without changing $F F^\top$. For the symmetric reformulation, the converse direction needs the eigenvalue condition that no two nonzero eigenvalues of $F$ sum to zero; without it, a second-order point of the factored problem can fail even the first-order conditions of the original.
Load-bearing premise
The load-bearing premise is that the weak subspace version of the second-order necessary condition is the right target; the clean correspondence does not survive if one uses the stronger cone version, except in the nondegenerate case where the multiplier fills the null space of the constraint.
Editorial extensions
If this is right
- Any algorithm that provably converges to a second-order necessary point of an equality-constrained problem can be applied to (SSV) and will automatically certify a second-order necessary point of the original (NSDP).
- For convex objectives in the nuclear-norm application, a second-order point of the factored formulation is globally optimal, so the factored form needs no rank assumptions and no strong measurement assumptions.
- The exact equivalence depends on square factors; with rectangular $d\times k$ factors for $k<d$, there are convex examples where a second-order point of the factored problem is not even first-order for the original.
- The symmetric-factor variant satisfies the correspondence only under the eigenvalue condition, so a user of symmetric squared variables must check that no two nonzero eigenvalues of $F$ sum to zero.
- Strict local minimizers are generically absent from the factored problem because of rotational symmetry, while local minimizers themselves do correspond.
Reading between the lines
- The paper stops at the theoretical equivalence; an immediate empirical program is to run equality-constrained NLP solvers on (SSV) for benchmark nonlinear SDPs and check whether the second-order points they reach are competitive, which the theorem does not by itself guarantee.
- Because the weak 2NC is checkable while the stronger cone-based version is generally not, the paper reframes the practical meaning of second-order optimality for this problem class.
- The same overparametrization logic could be exported to other PSD-constrained problems with convex objectives: square factoring plus a second-order point of the factored form would replace specialized SDP solvers.
- The rotation symmetry that kills strict local minima in (SSV) is an implicit warning that optimization methods on the factored form must either quotient by that symmetry or tolerate flat directions, a point the paper notes but does not develop into an algorithmic prescription.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper studies second-order necessary conditions (2NC) for nonlinear semidefinite programming problems (NSDP) and their squared-variable reformulations (SSV), in which the positive semidefinite matrix C(x) is replaced by FF^T, as well as the direct-substitution cases (BC)/(DSS) and symmetric variants (SSV-Sym)/(DSS-Sym). The main results are Theorem 3.1, stating that a point satisfies the (weak) second-order necessary conditions for (NSDP) if and only if any corresponding square factor F satisfies the second-order necessary conditions for (SSV), and Theorem 2.1, the analog for (BC) and (DSS). The paper also proves one-directional results for the symmetric factor variant, characterizes when the eigenvalue condition (EC) is needed, analyzes local-minimizer correspondences in Appendix B, locates the paper's weak 2NC notion relative to the stronger s2NC of Shapiro and Lourenço et al. in Appendix C, and applies the (BC)/(DSS) equivalence to nuclear-norm minimization via an overparametrized factorization. The proofs are algebraic and rely on two technical lemmas (Lemmas 2.1 and 2.2); Lemma 2.2 is proved in Appendix A. The paper is careful to state limitations, including the gap between weak 2NC and stronger cone-based conditions, and it supplies counterexamples (Examples 2.1, 2.2, 3.1, B.1) showing the necessity of stated conditions.
Significance. If the results hold, the paper gives a clean, parameter-free bridge between second-order points of PSD-constrained problems and their squared-factor reformulations, thereby justifying the use of equality-constrained and unconstrained solvers for a broader class of nonconvex matrix optimization problems. The claimed equivalence for the nonsymmetric factor formulation requires no constraint qualification, strict complementarity, or eigenvalue condition, which is a genuine improvement over earlier results in [LFF18] that needed strict complementarity and transversality. The paper is also honest about the scope of its central notion: the equivalence is for the weak 2NC over a subspace, not the stronger cone-based s2NC, and Appendix C explains precisely when the two notions coincide. The proofs are self-contained, with the key algebra in Lemmas 2.1 and 2.2 and their application in the theorem proofs; the counterexamples are concrete and demonstrate tightness of the stated hypotheses.
major comments (5)
- [Theorem 3.1 and Definition 4] The central equivalence is stated for the weak second-order necessary condition (NSDP-2NC), in which the curvature inequality is required only on the subspace V^T DC_x[z] V = 0. This is a strictly weaker condition than the cone-based s2NC of Shapiro [Sha97] and Lourenço et al. [LFF18], and the paper acknowledges this in footnote 2 and Appendix C. The stated theorem is internally consistent; however, the algorithmic promise in the introduction—that algorithms converging to second-order points of (SSV) can be used to solve (NSDP)—should be read as applying to this weak notion. Since s2NC reduces to 2NC only under strict complementarity (Appendix C, equation (54)), the practical implications for the strong notion remain open. This is a scope limitation, not a mathematical error, and it is handled honestly. I do not regard it as a blocker, but the abstract and introduction could state more prominently that 'second-order point' throughout means the weak 2NC.
- [Lemma 2.1] The statement of Lemma 2.1 says 'a rank r matrix X ... has a factorization X = F F^T for some F in R^{d x k} with k >= r', and the proof uses F in R^{d x k}. However, the lemma is applied in Theorem 2.1 and Theorem 3.1 with F square (k = d), and in the proof of Lemma 2.1 the matrix Delta is taken in R^{d x d}, while the statement should presumably allow Delta in R^{d x k} for consistency with the lemma's general F. This is a minor notational mismatch that does not affect the applications, but the lemma's dimension parameters should be stated uniformly.
- [Section 3.1, proof of Theorem 3.1] The converse direction of Theorem 3.1 uses Lemma 2.2 with S = Lambda after establishing that Lambda is PSD. This ordering is correct: Lambda >= 0 is proved first via the z = 0, Delta = w v^T argument, and only then is Lemma 2.2 invoked. The reader's concern about circularity is therefore not realized. The only point to note is that the proof of Lambda >= 0 relies on choosing v with F v = 0, which is valid when F is singular; the invertible case is handled separately. This is sound.
- [Example 2.1] The example showing that a square factor F is necessary for the 2NP equivalence involves a convex quadratic h and a rank-1 unique minimizer, and the computation of the second-order condition at F_k is algebraic and verifiable. The example is persuasive for the claim that rectangular Burer-Monteiro factors with k < d do not inherit the equivalence. One minor clarification: the statement 'for any k < d there is a 2NP F of (DSS-BM) such that F F^T is not a 1P of (BC)' is established by constructing F_k with a specific eta_k; the displayed calculation shows 2NC holds for all Delta, and the conclusion that F_k F_k^T is not a 1P uses the uniqueness of the 1P of (BC), which is argued. This is complete.
- [Appendix C] The appendix correctly explains the relationship between 2NC and s2NC and shows that under strict complementarity the two conditions coincide. However, the claim that the cone of z in Definition 7 contains the subspace of Definition 4 is derived after a nontrivial decomposition (equations (50)-(53)); this is correct. The appendix could be more explicit that the weak 2NC is what is certified by the squared-variable reformulation, while s2NC is not generally certified without strict complementarity. This is already stated in footnote 2, and the appendix is a useful clarification.
minor comments (6)
- [Throughout] The label 'Theorem 3.1' in the main text appears as 'Theorem 3.1' in Section 3.1 and is cited correctly, but in the sentence following Definition 5 the text says 'if (x, F, Lambda) satisfies 2NC for (DSS)' where it should say 'for (SSV)'. This is a typo.
- [Section 1.2] The notation section defines e_i, I_k, 0_k, and V_X but uses 0_{d-1}, 0_{(n-k) x k}, and I_{d-1} in examples without comment. The usage is clear, but a brief note that subscripts denote sizes would help.
- [Equation (3)] The bracehtip markers in equation (3) appear as LaTeX artifacts in the manuscript rendering and obscure the displayed formula. The underlying algebra is correct, but the authors should ensure the final published version renders these annotations properly.
- [Section 2.3] The application to nuclear norm minimization uses the notation Y and Z in (NNM-DSS) and then switches to Y1, Y2, Y3 in the symmetric variant. The relationship between these notations is stated, but a diagram or explicit block-matrix display would improve readability.
- [Appendix B, Example B.1] The local-minimizer example is long and the displayed expression for g(F+Delta)-g(F) in (47) would benefit from a check of the quadratic term; the subsequent inequality involving 10(z+y1^2)^2 and (z+y2^2)^2 is plausible but the algebra is dense. Adding a sentence explaining the choice of constants would improve readability without changing the result.
- [References] The related work discussion cites [LKB24] but does not compare the present theorem's 2NP-to-2NP correspondence with the general framework's Proposition 2.2 in a formal remark; the current comparison in Section 1.1 is helpful but could be sharpened by stating the relationship between the assumptions of [LKB24, Proposition 2.2] and the failure of 1P equivalence for (SSV)/(NSDP).
Circularity Check
No circularity: Theorem 3.1 is proved from explicit definitions and independently stated matrix lemmas; self-citations are motivational only.
full rationale
The paper's central claim (Theorem 3.1) is an equivalence between weak second-order necessary conditions for (NSDP) and (SSV). The proof does not assume the target result. Direction (NSDP-2NC) => (SSV-2NC): after deriving Lambda F = 0 from Lambda C(x) = 0, the paper writes D2Lssv = D2L + T1 + T2, with T1 >= 0 from (NSDP-2NC) and T2 >= 0 from Lemma 2.1, whose proof is a direct pseudoinverse trace computation. Direction (SSV-2NC) => (NSDP-2NC): Lambda >= 0 is obtained by choosing z = 0 and Delta = w v^T with F v = 0, and the remaining curvature inequality is supplied by Lemma 2.2, which constructs Delta explicitly via SVD and verifies tr(S(W X^dagger W - Delta Delta^T)) = 0. No fitted parameters are introduced and no 'prediction' is a renamed input. The paper's earlier work [DW23] is cited only as a scalar analog ('Our result can be regarded as a matrix analog ... [DW23, Theorem 2.3 and Theorem 3.3]'), not as a premise of any theorem. The eigenvalue-condition caveats in Theorems 2.2 and 3.2, Example 3.1, and Appendix C are explicit scope limitations rather than hidden assumptions. Hence there is no circular step; the honest finding is no significant circularity, score 0.
Assumptions & free parameters
assumptions (4)
- domain assumption The functions f, C, and h are twice continuously differentiable.
- domain assumption The weak 2NC in Definitions 1 and 4 are necessary conditions for local optimality of (BC) and (NSDP).
- standard math Pseudoinverse facts: (F F^T)^dagger = (F^dagger)^T F^dagger and I - F^T (F F^T)^dagger F is the orthogonal projector onto the null space of F^T, hence PSD.
- standard math Davis-Kahan and Weyl inequalities for eigenvalue perturbation.
Cite this review
Pith. "Pith review of On Squared-Variable Formulations for Nonlinear Semidefinite programming." pith.science (2026). https://pith.science/paper/NJUWQQUY
@misc{pith2026250202099,
author = {Pith},
title = {Pith review of: On Squared-Variable Formulations for Nonlinear Semidefinite programming},
year = {2026},
howpublished = {\url{https://pith.science/paper/NJUWQQUY}},
note = {Machine review of arXiv:2502.02099}
}
abstract
In optimization problems involving smooth functions and real and matrix variables, that contain matrix semidefiniteness constraints, consider the following change of variables: Replace the positive semidefinite matrix $X \in \mathbb{S}^d$, where $\mathbb{S}^d$ is the set of symmetric matrices in $\mathbb{R}^{d\times d}$, by a matrix product $FF^\top$, where $F \in \mathbb{R}^{d \times d}$ or $F \in \mathbb{S}^d$. The formulation obtained in this way is termed ``squared variable," by analogy with a similar idea that has been proposed for real (scalar) variables. It is well known that points satisfying first-order conditions for the squared-variable reformulation do not necessarily yield first-order points for the original problem. There are closer correspondences between second-order points for the squared-variable reformulation and the original formulation. These are explored in this paper, along with correspondences between local minimizers of the two formulations.
Reference graph
Works this paper leans on
- [1]
-
[2]
Smoothed analysis for low-rank solutions to semidefinite programs in quadratic penalty form
Srinadh Bhojanapalli, Nicolas Boumal, Prateek Jain, and Praneeth Netrapalli. Smoothed analysis for low-rank solutions to semidefinite programs in quadratic penalty form. In Conference On Learning Theory , pages 3243--3270. PMLR, 2018
work page 2018
-
[3]
Afonso S Bandeira, Nicolas Boumal, and Vladislav Voroninski. On the low-rank approach for semidefinite programs arising in synchronization and community detection. In Conference on Learning Theory , pages 361--382. PMLR, 2016
work page 2016
-
[4]
D. P. Bertsekas. Nonlinear Programming . Athena Scientific, second edition, 1999
work page 1999
-
[5]
A nonlinear programming algorithm for solving semidefinite programs via low-rank factorization
Samuel Burer and Renato DC Monteiro. A nonlinear programming algorithm for solving semidefinite programs via low-rank factorization. Mathematical programming , 95(2):329--357, 2003
work page 2003
-
[6]
Local minima and convergence in low-rank semidefinite programming
Samuel Burer and Renato DC Monteiro. Local minima and convergence in low-rank semidefinite programming. Mathematical programming , 103(3):427--444, 2005
work page 2005
-
[7]
The non-convex Burer-Monteiro approach works on smooth semidefinite programs
Nicolas Boumal, Vlad Voroninski, and Afonso Bandeira. The non-convex Burer-Monteiro approach works on smooth semidefinite programs. Advances in Neural Information Processing Systems , 29, 2016
work page 2016
-
[8]
Deterministic guarantees for Burer-Monteiro factorizations of smooth semidefinite programs
Nicolas Boumal, Vladislav Voroninski, and Afonso S Bandeira. Deterministic guarantees for Burer-Monteiro factorizations of smooth semidefinite programs. Communications on Pure and Applied Mathematics , 73(3):581--608, 2020
work page 2020
Show all 39 references
-
[9]
Harnessing structures in big data via guaranteed low-rank matrix estimation: Recent theory and fast algorithms via convex and nonconvex optimization
Yudong Chen and Yuejie Chi. Harnessing structures in big data via guaranteed low-rank matrix estimation: Recent theory and fast algorithms via convex and nonconvex optimization. IEEE Signal Processing Magazine , 35(4):14--31, 2018
2018
-
[10]
Noisy matrix completion: Understanding statistical guarantees for convex relaxation via nonconvex optimization
Yuxin Chen, Yuejie Chi, Jianqing Fan, Cong Ma, and Yuling Yan. Noisy matrix completion: Understanding statistical guarantees for convex relaxation via nonconvex optimization. SIAM journal on optimization , 30(4):3098--3121, 2020
2020
-
[11]
Nonconvex optimization meets low-rank matrix factorization: An overview
Yuejie Chi, Yue M Lu, and Yuxin Chen. Nonconvex optimization meets low-rank matrix factorization: An overview. IEEE Transactions on Signal Processing , 67(20):5239--5269, 2019
2019
-
[12]
Polynomial time guarantees for the Burer-Monteiro method
Diego Cifuentes and Ankur Moitra. Polynomial time guarantees for the Burer-Monteiro method. Advances in Neural Information Processing Systems , 35:23923--23935, 2022
2022
-
[13]
Metric regularity, tangent sets, and second-order optimality conditions
Roberto Cominetti. Metric regularity, tangent sets, and second-order optimality conditions. Applied Mathematics and Optimization , 21:265--287, 1990
1990
-
[14]
Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random measurements
Emmanuel J Candes and Yaniv Plan. Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random measurements. IEEE Transactions on Information Theory , 57(4):2342--2359, 2011
2011
-
[15]
On the simplicity and conditioning of low rank semidefinite programs
Lijun Ding and Madeleine Udell. On the simplicity and conditioning of low rank semidefinite programs. SIAM Journal on Optimization , 31(4):2614--2637, 2021
2021
-
[16]
On squared-variable formulations
Lijun Ding and Stephen J Wright. On squared-variable formulations. arXiv preprint arXiv:2310.01784 , 2023
2023 arXiv
-
[17]
Analysis on symmetric cones
Jacques Faraut and Adam Kor \'a nyi. Analysis on symmetric cones . Oxford university press, 1994
1994
-
[18]
Nonlinear Programming: Sequential Unconstrained Minimization Techniques
Anthony V Fiacco and Garth P McCormick. Nonlinear Programming: Sequential Unconstrained Minimization Techniques . SIAM, 1990
1990
-
[19]
Optimality conditions for nonconvex semidefinite programming
Anders Forsgren. Optimality conditions for nonconvex semidefinite programming. Mathematical Programming , 88:105--128, 2000
2000
-
[20]
No spurious local minima in nonconvex low rank problems: A unified geometric analysis
Rong Ge, Chi Jin, and Yi Zheng. No spurious local minima in nonconvex low rank problems: A unified geometric analysis. In International Conference on Machine Learning , pages 1233--1242. PMLR, 2017
2017
-
[21]
How to escape saddle points efficiently
Chi Jin, Rong Ge, Praneeth Netrapalli, Sham M Kakade, and Michael I Jordan. How to escape saddle points efficiently. In International conference on machine learning , pages 1724--1732. PMLR, 2017
2017
-
[22]
Optimality conditions for nonlinear semidefinite programming via squared slack variables
Bruno F Louren c o, Ellen H Fukuda, and Masao Fukushima. Optimality conditions for nonlinear semidefinite programming via squared slack variables. Mathematical Programming, Series B , 168:177--200, 2018
2018
-
[23]
The effect of smooth parametrizations on nonconvex optimization landscapes
Eitan Levin, Joe Kileel, and Nicolas Boumal. The effect of smooth parametrizations on nonconvex optimization landscapes. Mathematical Programming , pages 1--49, 2024
2024
-
[24]
From the simplex to the sphere: Faster constrained optimization using the hadamard parametrization
Qiuwei Li, Daniel McKenzie, and Wotao Yin. From the simplex to the sphere: Faster constrained optimization using the hadamard parametrization. arXiv preprint arXiv:2112.05273 , 2021
2021 arXiv
-
[25]
Gradient descent only converges to minimizers
Jason D Lee, Max Simchowitz, Michael I Jordan, and Benjamin Recht. Gradient descent only converges to minimizers. In Conference on Learning Theory , pages 1246--1257. PMLR, 2016
2016
-
[26]
Linear and nonlinear programming second edition
David G Luenberger. Linear and nonlinear programming second edition. Columbus, Ohio: Addison-Wesley , 1984
1984
-
[27]
Nocedal and S
J. Nocedal and S. J. Wright. Numerical Optimization . Springer, New York, second edition, 2006
2006
-
[28]
The Burer-Monteiro sdp method can fail even above the Barvinok-Pataki bound
Liam O'Carroll, Vaidehi Srinivas, and Aravindan Vijayaraghavan. The Burer-Monteiro sdp method can fail even above the Barvinok-Pataki bound. Advances in Neural Information Processing Systems , 35:31254--31264, 2022
2022
-
[29]
On the rank of extreme matrices in semidefinite programs and the multiplicity of optimal eigenvalues
G \'a bor Pataki. On the rank of extreme matrices in semidefinite programs and the multiplicity of optimal eigenvalues. Mathematics of Operations Research , 23(2):339--358, 1998
1998
-
[30]
An inexact augmented lagrangian framework for nonconvex optimization with nonlinear constraints
Mehmet Fatih Sahin, Ahmet Alacaoglu, Fabian Latorre, Volkan Cevher, et al. An inexact augmented lagrangian framework for nonconvex optimization with nonlinear constraints. Advances in Neural Information Processing Systems , 32, 2019
2019
-
[31]
First and second order analysis of nonlinear semidefinite programs
Alexander Shapiro. First and second order analysis of nonlinear semidefinite programs. Mathematical programming , 77:301--320, 1997
1997
-
[32]
Similarity and other spectral relations for symmetric cones
Jos F Sturm. Similarity and other spectral relations for symmetric cones. Linear Algebra and its applications , 312(1-3):135--154, 2000
2000
-
[33]
High-dimensional statistics: A non-asymptotic viewpoint , volume 48
Martin J Wainwright. High-dimensional statistics: A non-asymptotic viewpoint , volume 48. Cambridge University Press, 2019
2019
-
[34]
Characterization of the subdifferential of some matrix norms
G Alistair Watson. Characterization of the subdifferential of some matrix norms. Linear Algebra Appl , 170(1):33--45, 1992
1992
-
[35]
Rank optimality for the Burer-Monteiro factorization
Irene Waldspurger and Alden Waters. Rank optimality for the Burer-Monteiro factorization. SIAM journal on Optimization , 30(3):2577--2602, 2020
2020
-
[36]
Complexity of proximal augmented lagrangian for nonconvex optimization with nonlinear equality constraints
Yue Xie and Stephen J Wright. Complexity of proximal augmented lagrangian for nonconvex optimization with nonlinear equality constraints. Journal of Scientific Computing , 86(3):1--30, 2021
2021
-
[37]
A useful variant of the davis--kahan theorem for statisticians
Yi Yu, Tengyao Wang, and Richard J Samworth. A useful variant of the davis--kahan theorem for statisticians. Biometrika , 102(2):315--323, 2015
2015
-
[38]
Improved global guarantees for the nonconvex Burer-Monteiro factorization via rank overparameterization
Richard Y Zhang. Improved global guarantees for the nonconvex Burer-Monteiro factorization via rank overparameterization. arXiv preprint arXiv:2207.01789 , 2022
2022 arXiv
-
[39]
Global optimality in low-rank matrix optimization
Zhihui Zhu, Qiuwei Li, Gongguo Tang, and Michael B Wakin. Global optimality in low-rank matrix optimization. IEEE Transactions on Signal Processing , 66(13):3614--3628, 2018
2018
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.