Semidefinite Relaxation (SDR)

Updated 12 January 2026

Semidefinite relaxation (SDR) is a convexification technique that lifts nonconvex quadratic or polynomial programs into semidefinite programming formulations by relaxing rank or integer constraints.
It is widely applied in combinatorial optimization, signal processing, wireless communications, and quantum information, providing efficient near-optimal solutions in applications like Max-Cut, beamforming, and MIMO detection.
Enhanced SDR methods exploit problem-specific structures to tighten bounds and reduce computational overhead, making them effective for tackling large-scale and complex optimization challenges.

Semidefinite relaxation (SDR) is a convexification technique with extensive impact across combinatorial optimization, signal processing, wireless communications, quantum information, and power systems. The central idea is to embed a nonconvex quadratic or polynomial program—typically involving rank or integer constraints—into the cone of positive semidefinite matrices, relax the nonconvex constraints, and exploit tractable semidefinite programming (SDP) formulations to obtain efficiently computable bounds and often near-exact solutions.

1. Foundational Principles and Mathematical Formulation

SDR converts nonconvex quadratic programs into convex SDPs by lifting vector variables into matrix variables and relaxing rank or integrality constraints. For a prototypical quadratic optimization,

$\min_{x \in \mathbb{R}^n} \quad x^\top Q x + c^\top x, \quad\text{s.t.}\quad x \in \mathcal{S},$

where $\mathcal{S}$ is nonconvex (e.g., $x \in \{0,1\}^n$ or $|x_i|=1$ ), SDR lifts to $X = x x^\top$ and re-expresses objectives and constraints linearly in $X$ , while omitting nonconvex rank or discretization requirements: $\min_{X \in \mathbb{S}^n} \quad \operatorname{Tr}(Q X) + c^\top x, \quad\text{s.t.}\quad X \succeq 0, \text{ (linear constraints possibly derived from %%%%5%%%%)}.$ This transformation leverages the convexity of the semidefinite cone and the linear structure inherent in trace formulations (Tavakoli et al., 2023). It underpins hierarchies like Lovász-Schrijver for binary integer programs (Paparella, 2012), Lasserre for polynomial optimization in commutative and noncommutative settings (Tavakoli et al., 2023), and Navascués-Pironio-Acín (NPA) for quantum correlations.

2. SDR for Discrete and Combinatorial Optimization

SDR has become a standard tool for hard combinatorial problems, including Max-Cut, Boolean quadratic programming, quadratic assignment (QAP), and scheduling. In Boolean QP, e.g.,

$\min_{x \in \{0,1\}^n} x^\top C x + 2d^\top x,$

the lifted SDR admits $X \succeq 0$ , $X_{ii} = X_{0i}$ , and $X_{00}=1$ , yielding much tighter bounds than canonical LP relaxations. The Lovász-Schrijver relaxation projects the feasible region into a spectrahedron, cutting off many fractional LP vertices (Paparella, 2012). Enhanced SDR variants exploit eigenvalue structure to penalize non-rank-one solutions, as in concave penalization $-\lambda \langle X, X \rangle$ , majorization-minimization descent, and incorporation of sparsity information (Cerone et al., 2020).

For QAP, SDR lifts permutation matrix variables into the full PSD matrix algebra, and dual certificate analysis provides algebraic exactness conditions based solely on the signal-to-noise ratio between input data and spectral gaps (Ling, 2024). In air traffic scheduling, the SDR captures capacity constraints and delay logic by embedding the MILP in a single semidefinite constraint parametrized by system parameters [0609145].

SDR for the traveling salesman problem (TSP) exhibits limitations: recent work shows that natural SDP relaxations can have unbounded integrality gaps, especially on special cut-metric instances—the gap can grow linearly with problem size $n$ , indicating that the SDP solution may be arbitrarily far from integer-feasible solutions (Gutekunst et al., 2017).

3. SDR in Signal Processing, Wireless, and Communications

Semidefinite relaxation underpins both optimal and robust design in communications:

Downlink beamforming under channel uncertainties is convexified via SDR and S-lemma; the resulting SDP is tight and always recovers globally optimal rank-one beamformers under standard feasibility conditions (Chang et al., 2012).
Rank-two Alamouti relay beamforming is formulated as a fractional QCQP and its SDR, together with Gaussian randomization, yields constant-factor approximation guarantees with provable tightness for a small number of users and constraints (Wu et al., 2016).

In multiple-input multiple-output (MIMO) detection, SDR is applied to complex QCQPs with discrete argument (constellation) constraints. Enhanced SDR methods explicitly impose angular constraints via convex polytopes or probability-simplex lifts and adopt block-diagonal or separable matrix factorizations for analytical tightness and computational efficiency. Necessary and sufficient PSD conditions for tightness have been derived for M-PSK constellations, based on eigenvalue spectra and geometric properties of the constraint (Lu et al., 2017, Jiang et al., 2021). For high-order QAM (e.g., 16-QAM, 64-QAM), mathematically distinct SDR formulations—polynomial-inspired (PI-SDR), bound-constrained (BC-SDR), and virtually-antipodal (VA-SDR)—have been shown to be equivalent, supporting the selection of computationally simplest forms without performance loss (0809.4529).

Turbo receivers for LDPC-coded MIMO can be integrated into a joint SDR framework, which embeds both code and symbol detection in a single SDP. This approach provides improved soft information and bit error rates, with complexity reduction strategies exploiting single-SDR solutions across iterative decoding rounds (Wang et al., 2018).

4. Quantum Information and Polynomial Optimization

Semidefinite relaxation enables systematic analysis of quantum correlations and polynomial optimization hierarchies (Tavakoli et al., 2023). In the quantum setting, nonlocality and entanglement properties are cast as feasibility or optimization problems over sets defined by quantum moment matrices. Relaxing to an SDP over the quantum moment matrix (commuting or noncommuting monomials), subject to physical and symmetry constraints, produces convergent outer approximations to target sets (e.g., quantum correlations, separable states).

The Lasserre and NPA hierarchies are representative: each level increases the degree of moment matrix considered and tightens the relaxation at the expense of exponential growth in matrix size and computational burden. Strong duality and explicit polynomial or operator certificates are available in the SDP duals. Exploitation of problem symmetries (e.g., via group representations) is essential for scaling to large quantum or polynomial systems.

5. Enhanced and Specialized SDRs

Standard SDRs may be loose in the presence of structure such as modulus and argument constraints, leading to advances in enhanced SDR (Lu et al., 2019, Xu et al., 2023). This class incorporates convex envelopes of nonconvex sets—argument and modulus coupling, phase difference constraints, discrete and continuous angle constraints—via linear and second-order cone inequalities directly in the lifted matrix space. These augmented SDPs close substantial relaxation gaps, improve suboptimal solution extraction, and generalize to both discrete and continuous phase and amplitude regime problems in applications such as radar waveform design and beamforming.

Global branch-and-bound algorithms for complex quadratic programs utilize enhanced SDR bounds to partition the feasible space, systematically closing the dual gap and rapidly converging to $\epsilon$ -optimal solutions. Empirical results demonstrate order-of-magnitude improvements in solution quality and computational performance against tailored combinatorial or general-purpose algorithms (Lu et al., 2019).

6. Scalability, Tightness, and Limitations

Advanced first-order and interior-point methods (SDPAD-LR, chordal decomposition, low-rank factorization) support practical application of SDR to extremely large-scale problems, including Markov Random Field (MRF) MAP inference and network optimal power flow. For large transmission networks (e.g., up to 82,000 buses), carefully structured SDR formulations together with robust chordal decomposition and scaling can yield globally optimal solutions within minutes to hours, with optimality gaps generally below 1% except in severely stressed cases (Eltved et al., 2018, Huang et al., 2014).

SDR is tight in a range of problems—rank-one solutions are numerically and sometimes analytically guaranteed for MISO wireless power transfer (Lang et al., 2017), robust beamforming (Chang et al., 2012), and under specific spectral gap conditions for QAP (Ling, 2024), MIMO detection (Lu et al., 2017, Jiang et al., 2021), and certain multicasting scenarios (Wu et al., 2016). However, SDR may not be tight for all problems; as demonstrated in TSP (Gutekunst et al., 2017), even sophisticated SDP relaxations can fail to provide meaningful bounds on certain hard combinatorial instances.

7. Equivalence Results and Block-Structured SDRs

Recent analyses have unified disparate SDR formulations under equivalence theorems that exploit separable or block-diagonal structure in the constraints. These findings substantiate the mathematical equivalence of supposedly distinct forms, facilitate significant computational savings via dimension reduction, and clarify tightness conditions. Applications span SDRs for high-order QAM detection (0809.4529), enhanced complex and real SDRs for PSK detection (Jiang et al., 2021), and general block-structured convexification in polynomial and power flow problems.

References: