Weighted Minimum Mean Square Error

• WMMSE is a framework that converts nonconvex sum-rate and utility maximization problems into structured convex surrogate problems using alternating minimization.
• The method optimizes receive filters, nonnegative weights, and transmit variables with closed-form updates, ensuring provable convergence and low complexity.
• Recent extensions apply WMMSE to large-scale MIMO, reconfigurable intelligent surfaces, integrated sensing and communications, and deep learning architectures.

The weighted minimum mean square error (WMMSE) framework is a fundamental methodology for solving nonconvex optimization problems arising in wireless communications, signal processing, and networked systems. WMMSE enables the transformation of difficult sum-rate or utility maximization, often involving signal-to-interference-plus-noise ratio (SINR) or mutual information objectives, into structured (block-)convex surrogate problems that can be efficiently solved by alternating minimization. Modern research extends WMMSE to large-scale MIMO, RIS- and antenna-reconfigurable systems, integrated sensing and communications, and deep learning architectures, often with provable convergence, low complexity, and scalable implementations.

1. Fundamental Principles and Problem Equivalence

The canonical WMMSE approach is predicated on the equivalence between problems of the form

$$\max_{\mathbf{x}} \sum_k w_k \log_2(1 + \mathrm{SINR}_k(\mathbf{x}))$$

subject to power or other system constraints, and an augmented minimum weighted mean-square-error objective: $$\min_{\mathbf{w},\mathbf{u},\mathbf{v}} \sum_k w_k \, e_k(\mathbf{u},\mathbf{v}) - \log w_k$$ where $e_k$ is the estimation MSE of user $k$. This equivalence allows for an alternating block-coordinate minimization, typically over three blocks: receive filters (equalizers), nonnegative weights, and transmit/precoding variables (including power-control or beamforming vectors). At each step, closed-form optimal updates are available for all blocks except