A preconditioned augmented Lagrangian method for solving semidefinite programming problems
Pith reviewed 2026-05-20 14:43 UTC · model grok-4.3
The pith
A weighted penalty preconditioner from tangent space projection accelerates the augmented Lagrangian method for ill-conditioned semidefinite programs.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central discovery is that incorporating a weight matrix derived from the projection onto the tangent space of the feasible region into a weighted penalty function within the augmented Lagrangian subproblems substantially speeds up the solution process for ill-conditioned semidefinite programming problems. Combining this preconditioned ALM with the previously developed SDPF method results in the SDPF+ solver that efficiently addresses large-scale convex SDPs possibly featuring nonlinear objective functions.
What carries the argument
The weighted penalty function in the ALM subproblem, with its weight matrix coming from the tangent space projection operator onto the feasible region.
Load-bearing premise
The projection onto the tangent space produces a weight matrix that can be applied cheaply and stays numerically stable for the ill-conditioned SDPs that occur in applications.
What would settle it
Running the preconditioned ALM and the standard ALM on the same set of ill-conditioned SDP test problems and finding no consistent reduction in iteration counts or runtime would indicate the preconditioning does not accelerate convergence as claimed.
Figures
read the original abstract
In this work, we propose a preconditioned augmented Lagrangian method (ALM) for solving semidefinite programming (SDP) problems. The preconditioner is implemented via a weighted penalty function in the ALM subproblem, with the weight matrix derived from the projection operator onto the tangent space of the feasible region. This simple yet effective modification significantly accelerates ALM, particularly for ill-conditioned SDPs. By combining the preconditioned ALM with our previously developed feasible method SDPF, we develop SDPF+, an SDP solver capable of handling convex problems with possibly nonlinear objective functions. Extensive numerical experiments demonstrate the efficiency and robustness of SDPF+, showing that it can generally outperform other solvers on large-scale SDPs whose optimal solutions exhibit low-rank structure.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper proposes a preconditioned augmented Lagrangian method (ALM) for semidefinite programming (SDP) problems. The preconditioner is realized through a weighted penalty function in the ALM subproblem, with the weight matrix obtained from the projection operator onto the tangent space of the feasible region. The method is combined with the SDPF feasible method to create SDPF+, suitable for convex SDPs possibly with nonlinear objectives. The authors present extensive numerical experiments to illustrate the efficiency and robustness of SDPF+, particularly its ability to outperform other solvers on large-scale SDPs with low-rank optimal solutions.
Significance. If the performance claims hold under more detailed scrutiny, the work could meaningfully advance SDP solvers by accelerating ALM on ill-conditioned instances and delivering competitive results on large low-rank problems. The numerical results appear consistent with the claimed acceleration and outperformance on the tested instances, with the per-iteration overhead of the tangent-space projection remaining controlled in the low-rank regime; these constitute concrete strengths of the manuscript.
major comments (1)
- Numerical Experiments section: the description of the test-set construction, stopping criteria, and any statistical assessment of performance differences is insufficient. This information is load-bearing for the central claim that SDPF+ generally outperforms other solvers on large-scale low-rank SDPs.
minor comments (2)
- Abstract: the claim of 'extensive numerical experiments' would benefit from an explicit cross-reference to the relevant tables or figures.
- Preconditioner construction: the discussion of numerical stability of the derived weight matrix across ill-conditioned instances could be expanded with a brief remark on observed condition numbers or failure modes.
Simulated Author's Rebuttal
We thank the referee for the careful review and the positive assessment of the potential impact of our work on accelerating ALM for ill-conditioned SDPs. We address the single major comment below and will incorporate the suggested improvements in the revised manuscript.
read point-by-point responses
-
Referee: Numerical Experiments section: the description of the test-set construction, stopping criteria, and any statistical assessment of performance differences is insufficient. This information is load-bearing for the central claim that SDPF+ generally outperforms other solvers on large-scale low-rank SDPs.
Authors: We agree that additional detail in the Numerical Experiments section would strengthen the presentation and better support the central performance claims. In the revision we will expand the description of the test-set construction to clarify how the large-scale low-rank convex SDP instances were generated (including the specific random models and parameter ranges used), provide the precise stopping criteria and tolerances applied to SDPF+ as well as to the competing solvers, and include statistical summaries such as performance profiles or average runtime ratios across the instance collection. These additions will make the experimental evidence more transparent without altering the reported results. revision: yes
Circularity Check
No significant circularity detected
full rationale
The paper presents an algorithmic construction for a preconditioned ALM where the weight matrix is explicitly built from the tangent-space projection operator onto the feasible region. This modification is justified by reference to standard ALM convergence theory together with new numerical experiments on ill-conditioned instances. The subsequent combination with the authors' earlier SDPF method to produce SDPF+ is a straightforward composition that does not alter the independent status of the preconditioner derivation or the reported acceleration. No equation is shown to be equivalent by construction to a fitted parameter, no uniqueness theorem is imported from the authors' own prior work to forbid alternatives, and the performance claims rest on external experimental benchmarks rather than internal self-reference. The derivation chain therefore remains self-contained against the stated assumptions and observed results.
Axiom & Free-Parameter Ledger
axioms (1)
- standard math Standard convergence theory for augmented Lagrangian methods continues to apply after the introduction of the tangent-space-derived weighting.
Reference graph
Works this paper leans on
-
[1]
E. Abbe. Community detection and stochastic block model s: recent developments. Journal of Machine Learning Research , 18(177):1–86, 2018
work page 2018
- [2]
-
[3]
A. I. Barvinok. Problems of distance geometry and convex properties of quadratic maps. Discrete & Computational Geometry , 13(2):189–202, 1995
work page 1995
-
[4]
H. H. Bauschke, P. L. Combettes, H. H. Bauschke, and P. L. C ombettes. Correction to: convex analysis and monotone operator theory in Hilbert spaces. Springer, 2017
work page 2017
-
[5]
A. Beck and M. Teboulle. A fast iterative shrinkage-thre sholding algorithm for linear inverse problems. SIAM J. on Imaging Sciences , 2(1):183–202, 2009
work page 2009
-
[6]
N. Boumal. An introduction to optimization on smooth manifolds . Cambridge Univer- sity Press, 2023
work page 2023
-
[7]
K. Bredies and H. Sun. A proximal point analysis of the pre conditioned alternating direction method of multipliers. Journal of Optimization Theory and Applications , 173:878–907, 2017
work page 2017
-
[8]
S. Burer. On the copositive representation of binary and continuous nonconvex quadratic programs. Mathematical Programming, 120(2):479–495, 2009. 23
work page 2009
-
[9]
S. Burer and C. Choi. Computational enhancements in low- rank semidefinite program- ming. Optimisation Methods and Software , 21(3):493–512, 2006
work page 2006
-
[10]
S. Burer and R. D. Monteiro. A nonlinear programming alg orithm for solving semidef- inite programs via low-rank factorization. Mathematical Programming, 95(2):329–357, 2003
work page 2003
-
[11]
R. E. Burkard, S. E. Karisch, and F. Rendl. QAPLIB–a quad ratic assignment problem library. Journal of Global optimization , 10:391–403, 1997
work page 1997
-
[12]
Y. Cui, D. Sun, and K.-C. Toh. On the r-superlinear conve rgence of the KKT residuals generated by the augmented Lagrangian method for convex com posite conic program- ming. Mathematical Programming, 178(1):381–415, 2019
work page 2019
-
[13]
P. Giselsson and S. Boyd. Diagonal scaling in douglas-r achford splitting and admm. In 53rd IEEE Conference on Decision and Control , pages 5033–5039. IEEE, 2014
work page 2014
- [14]
-
[15]
Q. Han, C. Li, Z. Lin, C. Chen, Q. Deng, D. Ge, H. Liu, and Y. Ye. A low-rank ADMM splitting approach for semidefinite programming. INFORMS J. on Computing, to appear, 2025
work page 2025
-
[16]
D. Hou, T. Tang, and K.-C. Toh. A low-rank augmented Lagr angian method for doubly nonnegative relaxations of mixed-binary quadratic progra ms. Operations Resarch, to appear., 2025
work page 2025
-
[17]
D. Hou, T. Tang, and K.-C. Toh. Rinnal+: a riemannian alm solver for sdp-rlt relax- ations of mixed-binary quadratic programs: Di hou et al. Mathematical Programming Computation, pages 1–57, 2026
work page 2026
- [18]
-
[19]
S. Kontogiorgis and R. R. Meyer. A variable-penalty alt ernating directions method for convex optimization. Mathematical Programming, 83(1):29–53, 1998
work page 1998
-
[20]
J. B. Lasserre. Global optimization with polynomials a nd the problem of moments. SIAM J. on Optimization , 11(3):796–817, 2001
work page 2001
-
[21]
X. Li, D. Sun, and K.-C. Toh. QSDPNAL: a two-phase augmen ted Lagrangian method for convex quadratic semidefinite programming. Mathematical Programming Compu- tation, 10(4):703–743, 2018
work page 2018
-
[22]
Y. Liu, Y. Xu, and W. Yin. Acceleration of primal–dual me thods by preconditioning and simple subproblem procedures. Journal of Scientific Computing , 86(2):21, 2021. 24
work page 2021
-
[23]
U. Marteau-Ferey, F. Bach, and A. Rudi. Non-parametric models for non-negative functions. Advances in Neural Information Processing Systems , 33:12816–12826, 2020
work page 2020
- [24]
-
[25]
M. Nakata, B. J. Braams, K. Fujisawa, M. Fukuda, J. K. Per cus, M. Yamashita, and Z. Zhao. Variational calculation of second-order reduced d ensity matrices by strong n-representability conditions and an accurate semidefinit e programming solver. The Journal of Chemical Physics , 128(16), 2008
work page 2008
- [26]
- [27]
- [28]
-
[29]
G. Pataki. On the rank of extreme matrices in semidefinit e programs and the mul- tiplicity of optimal eigenvalues. Mathematics of Operations Research , 23(2):339–358, 1998
work page 1998
-
[30]
T. Pock and A. Chambolle. Diagonal preconditioning for first order primal-dual algo- rithms in convex optimization. In 2011 International Conference on Computer Vision , pages 1762–1769, 2011
work page 2011
-
[31]
J. Povh and F. Rendl. Copositive and semidefinite relaxa tions of the quadratic assign- ment problem. Discrete Optimization, 6(3):231–241, 2009
work page 2009
- [32]
-
[33]
F. Rendl. Semidefinite programming and combinatorial o ptimization. Applied Numer- ical Mathematics, 29(3):255–281, 1999
work page 1999
-
[34]
R. T. Rockafellar. Augmented Lagrangians and applicat ions of the proximal point algorithm in convex programming. Mathematics of Operations Research , 1(2):97–116, 1976
work page 1976
-
[35]
N. J. A. Sloane. Challenge problems: Independent sets i n graphs. http://www. research. att. com/˜ njas/doc/graphs. html , 2005
work page 2005
-
[36]
R. Sotirov. An efficient semidefinite programming relaxa tion for the graph partition problem. INFORMS Journal on Computing , 26(1):16–30, 2014
work page 2014
- [37]
-
[38]
D. Sun, Y. Yuan, G. Zhang, and X. Zhao. Accelerating prec onditioned ADMM via degenerate proximal point mappings. SIAM J. on Optimization , 35(2):1165–1193, 2025
work page 2025
-
[39]
T. Tang and K.-C. Toh. Exploring chordal sparsity in sem idefinite programming with sparse plus low-rank data matrices. arXiv preprint arXiv:2410.23849 , 2024
-
[40]
T. Tang and K.-C. Toh. A feasible method for general conv ex low-rank SDP problems. SIAM J. on Optimization , 34(3):2169–2200, 2024
work page 2024
-
[41]
T. Tang and K.-C. Toh. Solving graph equipartition SDPs on an algebraic variety. Mathematical Programming, 204(1):299–347, 2024
work page 2024
-
[42]
K.-C. Toh, M. J. Todd, and R. H. T¨ ut¨ unc¨ u. SDPT3—a Matl ab software package for semidefinite programming, version 1.3. Optimization Methods and Software , 11(1- 4):545–581, 1999
work page 1999
-
[43]
J. Wang and L. Hu. Solving low-rank semidefinite program s via manifold optimization. Journal of Scientific Computing , 104(1):1–33, 2025
work page 2025
-
[44]
S. Wang and C. Ding. Local convergence analysis of augme nted Lagrangian method for nonlinear semidefinite programming. Computational Optimization and Applications , 87(1):39–81, 2024
work page 2024
-
[45]
Y. Xu. Iteration complexity of inexact augmented Lagra ngian methods for constrained convex programming. Mathematical Programming, 185(1):199–244, 2021
work page 2021
-
[46]
L. Yang, D. Sun, and K.-C. Toh. SDPNAL+: a majorized semi smooth Newton- CG augmented Lagrangian method for semidefinite programmin g with nonnegative constraints. Mathematical Programming Computation, 7(3):331–366, 2015
work page 2015
-
[47]
X.-Y. Zhao, D. Sun, and K.-C. Toh. A Newton-CG augmented Lagrangian method for semidefinite programming. SIAM J. Optimization , 20(4):1737–1765, 2010. 26 A Proof of Theorem 1 In the following analysis, we always assume k≥ 2. From (19) and (20), we have that ∂zg(zk)−H ∗ (γ k) = ( ∂X δSn + (X k) +∇ f (X k)−A ∗ (λ k)−B ∗(µ k), ∂ yδP (yk) + µ k ) . (41) Becaus...
work page 2010
-
[48]
blk is a K× 2 cell array that specifies the block structure of the variables, wh ere K is the number of blocks. For each k∈ [K], blk{k,1} is a character indicating the type of the k-th block, which can be one of the following: • ’s’: symmetric positive semidefinite matrix variable, 28 • ’l’: nonnegative vector variable, • ’f’: free vector variable. blk{k,2}...
-
[49]
At is a K× 1 or K× 2 cell array that describe the linear operator A in (SDPG1). It has two columns when the user wants to make use of the sparse plus low- rank structure of the constraint matrices Ai’s described in [39]. Suppose the k-th block is a matrix variable with dimension n and the constraint matrices {Ai}i∈ [m] have sparse plus low-rank decomposit...
-
[50]
C is either a cell array or a function handle that describes the object ive function f in (SDPG1). When the SDP problem has linear objective function f (X) = ∑K i=1 ⟨ Ci, X i ⟩ such that the matrix blocks have sparse plus low-rank decomposition Ci = Cs i + Uidiag(di)U ⊤ i , we have that C{i, 1} = { Ci if blk{i,1}̸ = ’s’ Cs i if blk{i,1} = ’s’. (58) C{i, 2...
-
[51]
b is an m-dimensional vector which is exactly the vector b in (SDPG1)
-
[52]
The data structure of Bt is the same as At
Bt,l,u are used to denote the inequality constraint, l≤B (X)≤ u, in (SDPG1). The data structure of Bt is the same as At
-
[53]
The user can refer to the main function sdpfplus.m for the complete list of available parameters
par is a struct-type array that contains the parameters and options for the so lver. The user can refer to the main function sdpfplus.m for the complete list of available parameters. Although there are many parameters, most of them are intended f or solver development and have been pre-tuned. Here, we highlight only a few important param eters that are co...
work page 2001
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.