ManiSDP

ManiSDP aims to solve the following low-rank semidefinite program (SDP) via manifold optimization: $$\inf_{X\succeq0}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }\mathcal{A}(X)=b,{\ }\mathcal{B}(X)=d,$$ where the linear constraints $\mathcal{A}(X)=b$ are arbitrary while the linear constraints $\mathcal{B}(X)=d$, if present, are assumed to define certain manifold structure. Here, low-rank means the SDP admits a low-rank optimal solution.

Dependencies

• MATLAB
• Manopt

Usage

Note: The optimal setting of parameters in ManiSDP is highly problem-dependent. We encourage the users to find the optimal parameters (defined in options) via preliminary experiments before running large-scale cases. In our experiences, the parameters (sigma0, theta, TR_maxiter, TR_maxinner, tao) have a significant influence on the performance of ManiSDP.

ManiSDP accepts SeDuMi format data.

So far, ManiSDP supports four types of SDPs.

SDPs with only unit diagonal constraints

$$\inf_{X\succeq0}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }X_{ii}=1,{\ }\text{for}{\ }i=1,\ldots,n.$$

clear options;
options.tol = 1e-8;
[sol, opt, data] = ManiSDP_onlyunitdiag(C, options);

Input
C: cost matrix
options:
- tol (=1e-8 by default): tolerance of residues
- p0 (=40 by default): initial value of the factorization size p
- AL_maxiter (=20 by default): maximum number of iterations of the augmented Lagrangian method
- theta (=1e-1 by default): threshold for estimating matrix ranks
- delta (=8 by default): number of negative eigenvalues used to construct a descent direction
- alpha (=0.5 by default): step size for escaping from saddle points
- tolgrad (=1e-8 by default): tolerance of stopping gradient norms used in the Riemannian trust-region method
- TR_maxiter (=40 by default): maximum number of iterations of the Riemannian trust-region method
- TR_maxinner (=100 by default): maximum Hessian calls per trust-region iteration
- line_search (=0 by default): whether (=1) or not (=0) to use line search to decide the step size for escaping from saddle points

Output
sol: optimal solution
opt: optimum
data: auxiliary data

SDPs with unit diagonal constraints

$$\inf_{X\succeq0}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }\mathcal{A}(X)=b,{\ }X_{ii}=1,{\ }\text{for}{\ }i=1,\ldots,n.$$

clear options;
options.tol = 1e-8;
[sol, opt, data] = ManiSDP_unitdiag(At, b, c, n, options);

Input
At, b, c: SeDuMi format data
n: size of the PSD matrix
options:
- tol (=1e-8 by default): tolerance of residues
- p0 (=2 by default): initial value of the factorization size p
- AL_maxiter (=300 by default): maximum number of iterations of the augmented Lagrangian method
- sigma0 (=1e-3 by default): initial value of the penalty parameter
- sigma_min (=1e-2 by default): minimum value of the penalty parameter
- theta (=1e-3 by default): threshold for estimating matrix ranks
- delta (=8 by default): number of negative eigenvalues used to construct a descent direction
- alpha (=0.1 by default): step size for escaping from saddle points
- tolgrad (=1e-8 by default): tolerance of stopping gradient norms used in the Riemannian trust-region method
- TR_maxiter (=4 by default): maximum number of iterations of the Riemannian trust-region method
- TR_maxinner (=25 by default): maximum Hessian calls per trust-region iteration
- tao (=1 by default): factor for self-adaptively updating the penalty parameter
- line_search (=0 by default): whether (=1) or not (=0) to use line search to decide the step size for escaping from saddle points

Output
sol: optimal solution
opt: optimum
data: auxiliary data

SDPs with the unit trace constraint

$$\inf_{X\succeq0}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }\mathcal{A}(X)=b,{\ }\sum_{i}X_{ii}=1.$$

For the case that $X$ has a constant trace $c$, one can scale $X$ by the factor $\frac{1}{c}$ to match the unit trace case.

clear options;
options.tol = 1e-8;
[sol, opt, data] = ManiSDP_unittrace(At, b, c, n, options);

Input
At, b, c: SeDuMi format data
n: size of the PSD matrix
options:
- tol (=1e-8 by default): tolerance of residues
- p0 (=1 by default): initial value of the factorization size p
- AL_maxiter (=300 by default): maximum number of iterations of the augmented Lagrangian method
- sigma0 (=1e1 by default): initial value of the penalty parameter
- sigma_min (=1e2 by default): minimum value of the penalty parameter
- theta (=1e-2 by default): threshold for estimating matrix ranks
- delta (=10 by default): number of negative eigenvalues used to construct a descent direction
- alpha (=0.04 by default): step size for escaping from saddle points
- tolgrad (=1e-8 by default): tolerance of stopping gradient norms used in the Riemannian trust-region method
- TR_maxiter (=3 by default): maximum number of iterations of the Riemannian trust-region method
- TR_maxinner (=40 by default): maximum Hessian calls per trust-region iteration
- tao (=1/6e3 by default): factor for self-adaptively updating the penalty parameter
- line_search (=1 by default): whether (=1) or not (=0) to use line search to decide the step size for escaping from saddle points

Output
sol: optimal solution
opt: optimum
data: auxiliary data

SDPs with arbitrary affine constraints

$$\inf_{X\succeq0}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }\mathcal{A}(X)=b.$$

clear options;
options.tol = 1e-8;
[sol, opt, data] = ManiSDP(At, b, c, n, options);

Input
At, b, c: SeDuMi format data
n: size of the PSD matrix
options:
- tol (=1e-8 by default): tolerance of residues
- p0 (=1 by default): initial value of the factorization size p
- AL_maxiter (=300 by default): maximum number of iterations of the augmented Lagrangian method
- sigma0 (=1e-2 by default): initial value of the penalty parameter
- sigma_min (=1e-1 by default): minimum value of the penalty parameter
- theta (=1e-1 by default): threshold for estimating matrix ranks
- delta (=8 by default): number of negative eigenvalues used to construct a descent direction
- alpha (=0.2 by default): step size for escaping from saddle points
- tolgrad (=1e-8 by default): tolerance of stopping gradient norms used in the Riemannian trust-region method
- TR_maxiter (=4 by default): maximum number of iterations of the Riemannian trust-region method
- TR_maxinner (=50 by default): maximum Hessian calls per trust-region iteration
- tao (=0.25 by default): factor for self-adaptively updating the penalty parameter
- line_search (=1 by default): whether (=1) or not (=0) to use line search to decide the step size for escaping from saddle points

Output
sol: optimal solution
opt: optimum
data: auxiliary data

SDPs with multi-blocks

ManiSDP supports SDPs with multi-blocks:

$$\inf_{X}{\ }\langle C, X\rangle{\ }\text{s.t.}{\ }\mathcal{A}(X)=b, {\ }X \in \mathbb{S}_+^{n_1\times\cdots\times n_t}, {\ }\mathrm{diag}(X_i) = 1, {\ }i = 1,\ldots,K.nob.$$

clear options;
options.tol = 1e-4;
K.nob = 10; % indicate the first K.nob blocks are unit-diagonal
[sol, opt, data] = ManiSDP_multiblock(At, b, c, K, options);

Solving moment relxations for polynomial optimization problems

$$\inf_{x\in\mathbb{R}^q}{\ }f(x){\ }\text{s.t.}{\ }h_i(x)=0,{\ }i=1,\ldots,l.$$

First define your polynomial optimization problem using SPOTLESS as follows:

q = 10;
x = msspoly('x', q);
f = x'*randn(q, 1);
h = x.^2 - 1;
problem.vars = x;
problem.objective = f;
problem.equality = h;
kappa = 2; % relaxation order
[SDP,info] = dense_sdp_relax(problem, kappa);
[sol, opt, data] = ManiSDP(At, b, c, n, options);

Then generate the moment relxation and solve it with ManiSDP:

kappa = 2; % relaxation order
[SDP,info] = dense_sdp_relax(problem, kappa);
options.tol = 1e-8;
[sol, opt, data] = ManiSDP(SDP.sedumi.At, SDP.sedumi.b, SDP.sedumi.c, SDP.sedumi.K.s, options);

Examples

Check the folder /example.

Note that to run example_stls.m, one has to first add the folder NearestRankDeficient in CertifiablyRobustPerception; to run example_rotationsearch.m, one has to first add the folder RotationSearch in CertifiablyRobustPerception.

The Python version for ManiSDP

• ManiSDP-Python

The Julia version for ManiSDP

Coming soon.

References

Jie Wang and Liangbing Hu, Solving Low-Rank Semidefinite Programs via Manifold Optimization, 2023

Contact

Jie Wang: [email protected]