Semi-Proximal Schur Complement ADMM

• Semi-Proximal Schur Complement ADMM is an advanced variant of the classical ADMM that incorporates proximal and Schur complement methods to handle multi-block convex problems.
• It decouples complex subproblems using block Gauss-Seidel sweeps and carefully designed proximal regularizations, ensuring robust convergence even under ill-conditioned scenarios.
• The method is effectively applied to quadratic SDP, structured matrix recovery, and network tomography, significantly improving computational tractability and analytic update efficiency.

Semi-Proximal Schur Complement ADMM (Alternating Direction Method of Multipliers) refers to a class of operator splitting algorithms tailored for convex optimization problems with complex block structures, where proximal terms and Schur complement-based operator lifts are systematically introduced to ensure both convergence and computational tractability, particularly in the context of large-scale and multi-block convex programs. This methodology has been primarily developed for applications such as quadratic semidefinite programming (QSDP), structured matrix recovery, and high-dimensional network tomography, and achieves robust convergence properties where direct multi-block ADMM extensions fail, by leveraging careful 2-block groupings and majorization principles (Ming et al., 2021, Li et al., 2014, Xiao et al., 2015).

1. Problem Setting and Motivation

Semi-Proximal Schur Complement ADMM is designed for composite convex optimization problems of the form

$$\min_{x_1, x_2} f_1(x_1) + f_2(x_2) \quad \text{s.t.} \quad A_1 x_1 + A_2 x_2 = 0,$$

with $f_1, f_2$ closed, proper convex (possibly indicator plus quadratic), and $A_1, A_2$ linear operators. Originating in the classical ADMM framework, which splits these constraints into block-coordinate minimizations, the semi-proximal Schur Complement variant is motivated by two considerations:

• Standard ADMM exhibits non-guaranteed convergence when extended directly to more than two blocks.
• Many practical problems feature multi-block variable groupings and quadratic or low-rank structures, where the Hessians of augmented Lagrangians are ill-conditioned or highly coupled, making direct solution or inversion computationally prohibitive (Li et al., 2014).

The methodology is thus constructed to deliver guaranteed convergence and analytic updates—via Schur-complement-based proximal terms—under mild regularity assumptions.

2. Algorithmic Structure

The core algorithmic principle involves two main features:

• Block Gauss-Seidel Sweep: Variables are grouped into two (or occasionally more) blocks, each updated sequentially, with auxiliary proximal regularizations added to decouple subproblems and guarantee unique minimizers.
• Schur Complement-Based Majorization: Proximal operators (semi-proximal terms) are chosen such that the smooth part of each subproblem's Hessian is lifted to a scaled identity or easily inverted form, via addition of a Schur complement of the coupling operator.

For a generic two-block problem, the iteration at step $k$ proceeds as: $$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$ where $H_1$, $H_2$ are positive semidefinite, self-adjoint operators, $\tau > 0$ is a step parameter, and $\mathcal{L}_\beta$ is the augmented Lagrangian (Ming et al., 2021).

For multi-block settings, symmetric Gauss-Seidel sweeps and blockwise Schur complement majorization are used to maintain physical tractability and make subblock updates reduce to analytic, efficiently solvable forms (Xiao et al., 2015).

3. Construction of Schur Complement Semi-Proximal Terms

A decisive feature of the methodology is the construction of semi-proximal operators inspired by Schur complements, which ensures that each subproblem presents a simple structure. Specifically, for a subblock with operator $M$, one augments the quadratic penalty with a majorant $f_1, f_2$0 so that $f_1, f_2$1 (where $f_1, f_2$2 is the maximal eigenvalue of $f_1, f_2$3). For example, in link-traffic tomography,

$f_1, f_2$4

where $f_1, f_2$5 is summed over block components of the routing matrix, so the $f_1, f_2$6-subproblem becomes a diagonal scaling (Ming et al., 2021). In block-structured SDP,

$f_1, f_2$7

where $f_1, f_2$8 are diagonal or low-rank plus diagonal majorants for their respective blocks (Li et al., 2014). The general calculation involves replacing each coupled block's Hessian by its Schur complement-lifted diagonalization or decoupled quadratic.

4. Global Convergence Properties

Global convergence of Semi-Proximal Schur Complement ADMM is established by verifying three critical conditions:

• $f_1, f_2$9 and $A_1, A_2$0 are closed, proper, convex functionals with blockwise unique minimizers.
• Proximal majorants $A_1, A_2$1, $A_1, A_2$2 are chosen so that $A_1, A_2$3 is strictly positive definite for each block (guaranteed via maximal eigenvalue lift).
• The step-length $A_1, A_2$4 is either in $A_1, A_2$5 or, for $A_1, A_2$6, a residual summability condition holds.

Under these conditions, the full sequence of iterates converges to a primal-dual solution, as shown by direct application of monotonicity and paramonotonicity of the respective subdifferentials, standard Lyapunov-type descent, and boundedness arguments (Ming et al., 2021, Li et al., 2014, Xiao et al., 2015).

5. Implementation and Computational Complexity

Implementational details reflect the structure of the primal and dual sub-block updates. Dominant computational costs per iteration are:

• Singular value decomposition (SVD) of an $A_1, A_2$7 matrix for nuclear norm or spectral norm projections; $A_1, A_2$8 complexity, reducible to $A_1, A_2$9 via rank-$k$0 truncation.
• Matrix-vector operations involving the routing matrix $k$1 and its adjoint, $k$2.
• Elementwise thresholding/projection, $k$3.
• Storage for $k$4 blocks, with exploitation of matrix sparsity or low-rankness where possible (Ming et al., 2021).

For large semidefinite or conic programming tasks, block-diagonal and sparse factorizations are precomputed, and iterative solvers (e.g., for partial eigenvalue decompositions) are tolerated within the convergence theory when residual errors are summable (Xiao et al., 2015).

Parameter selection guidelines include:

• Penalty $k$5, with step-length $k$6 set to $k$7 (the golden ratio) or within $k$8 for over-relaxation.
• Majorants are chosen minimally above coupling operator spectra.
• Cross-validation is used for regularization parameters in data-motivated tasks (Ming et al., 2021).

6. Representative Applications

The methodology has been applied to:

• Internet Traffic Tomography: The Sparsity Low-Rank Recovery (SLRR) problem exploits both traffic sparsity and low-rank, yielding analytic updates for large-scale networks using SCB-spADMM, with empirical results on benchmark (Abilene, GEANT) and large commercial datasets (Huawei Origin-Destination) (Ming et al., 2021).
• Quadratic Semidefinite Programming: SCB-SPADMM achieves state-of-the-art performance for large-scale QSDP instances, including binary quadratic (BIQ) and quadratic assignment (QAP) matrices, by reducing coupled subproblems to tractable forms via Schur complement proximal construction (Li et al., 2014).
• Composite Conic Problems: General doubly non-negative SDP and cone-constrained quadratic programming benefit from symmetric Gauss-Seidel block decompositions and semi-proximal lifts (Xiao et al., 2015).

Table: Typical Operator Choices in SCB-SPADMM (see (Ming et al., 2021, Li et al., 2014))

Block	Operator	Majorant / Proximal Term
$k$9	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$0	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$1
$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$2	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$3	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$4
$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$5	Hessian + Coupling	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$6 (diagonal/low-rank)
$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$7	Hessian + Coupling	$$\begin{aligned} x_1^{k+1} &= \underset{x_1}{\arg\min}\;\; \mathcal{L}_\beta(x_1, x_2^k; y^k) + \frac{\beta}{2}\|x_1 - x_1^k\|_{H_1}^2 \ x_2^{k+1} &= \underset{x_2}{\arg\min}\;\; \mathcal{L}_\beta(x_1^{k+1}, x_2; y^k) + \frac{\beta}{2}\|x_2 - x_2^k\|_{H_2}^2 \ y^{k+1} &= y^k + \tau \beta [A_1 x_1^{k+1} + A_2 x_2^{k+1}] \end{aligned}$$8 (diagonal/low-rank)

7. Connections and Extensions

Semi-Proximal Schur Complement ADMM generalizes the classical augmented Lagrangian approach by introducing blockwise lifted regularizations based on Schur complements, encompassing:

• Block-structured composite conic programming (Xiao et al., 2015).
• Structured low-rank matrix recovery and tomography (Ming et al., 2021).
• Convex quadratic SDP with mixed polyhedral, spectral, and semidefinite constraints (Li et al., 2014).

Ongoing research investigates adaptive parameter selection, exploitation of low-rank structure, inexact subproblem solves, and implementation in distributed and parallel environments. The symmetric Gauss-Seidel block extension and relaxation parameter optimization are particularly effective in multi-block scenarios, further leading to robust convergence and practical efficiency.

References:

• "A Convergent Semi-Proximal Alternating Direction Method of Multipliers for Recovering Internet Traffics from Link Measurements" (Ming et al., 2021)
• "A Schur Complement Based Semi-Proximal ADMM for Convex Quadratic Conic Programming and Extensions" (Li et al., 2014)
• "A Generalized Alternating Direction Method of Multipliers with Semi-Proximal Terms for Convex Composite Conic Programming" (Xiao et al., 2015)