---
title: Enhanced WMMSE (eWMMSE) Methods Overview
url: https://www.emergentmind.com/topics/enhanced-wmmse-ewmmse
type: topic
---

# Enhanced WMMSE (eWMMSE) Methods Overview

Enhanced WMMSE (eWMMSE) denotes a family of weighted minimum mean-square error methods that retain the classical rate–MSE reformulation but extend or modify it to handle additional constraints, multi-channel structure, or lower-complexity computational realizations. In the narrowest explicit usage, recent papers use the name eWMMSE for multi-channel joint channel-and-power allocation, including per-user total-power constraints and, in one formulation, minimum-rate QoS constraints [2506.03813], [2509.06395]. In a broader and commonly implied sense, closely related works present “enhanced WMMSE” variants without using the exact acronym, for example by eliminating matrix inversions, replacing exact beamformer subproblem solutions by projected-gradient updates, or unfolding WMMSE into trainable architectures for fixed-latency execution [2006.08448], [2205.08877].

## 1. Classical foundation and the scope of enhancement

Classical WMMSE addresses weighted sum-rate maximization by introducing receiver variables and MSE weights so that a nonconvex rate problem can be rewritten as a blockwise-convex weighted-MSE problem. In the downlink MISO formulation with beamformers \(\boldsymbol v_i\), receive coefficients \(u_i\), and weights \(w_i\), the weighted sum-rate objective is
\[
\begin{aligned}
\max_{\boldsymbol V}\quad & \sum_{i=1}^{N} \alpha_i \log_2\!\bigl(1+\mathrm{SINR}_i\bigr) \\
\text{s.t.}\quad & \operatorname{Tr}(\boldsymbol V \boldsymbol V^H) \le P,
\end{aligned}
\]
with
\[
\mathrm{SINR}_i = \frac{|\boldsymbol h_i^H \boldsymbol v_i|^2} {\sum_{j=1,\, j\neq i}^{N} |\boldsymbol h_i^H \boldsymbol v_j|^2 + \sigma^2},
\]
and the equivalent WMMSE surrogate is
\[
\begin{aligned}
\min_{\boldsymbol u,\boldsymbol w,\boldsymbol V}\quad & \sum_{i=1}^{N} \alpha_i \bigl(w_i e_i - \log_2 w_i \bigr) \\
\text{s.t.}\quad & \operatorname{Tr}(\boldsymbol V \boldsymbol V^H) \le P.
\end{aligned}
\]
The standard iteration updates \(u_i\), \(w_i\), and then \(\boldsymbol v_i\) through a Lagrangian-based beamformer step involving \((\boldsymbol A+\mu \boldsymbol I)^{-1}\), where \(\mu\) is chosen to satisfy the power constraint [2006.08448].

The principal computational burden of classical WMMSE lies in the transmit update. In the Shi-style implementation highlighted in the unfolding literature, one iteration requires a matrix inversion, an eigendecomposition, and a bisection search over the power-control multiplier \(\mu\) [2006.08448]. In the MU-MIMO setting, inverse-based updates also appear in the receive filter and weight blocks, and these operations are described as implementation-hostile because they are expensive, serial, hard to parallelize, and inconvenient for deep unfolding and low-latency hardware [2205.08877].

Theoretically, later comparative work places WMMSE inside a broader algorithmic taxonomy. In particular, WMMSE is shown to be a special case of weighted sum-rate maximization via fractional programming, and specific WMMSE and WSR-FP instances can be mapped exactly to a minorization-maximization formulation, WSR-MM [2311.04546]. This suggests that “enhancement” is not tied to a single canonical modification: it may target the surrogate function, the equivalent transform, the constraint treatment, or the numerical realization of the transmit update.

## 2. Named eWMMSE in multi-channel resource allocation

The most explicit use of the term eWMMSE appears in multi-channel wireless-network resource allocation. In these formulations, the system has \(D\) transceiver pairs and \(M\) orthogonal channels, each pair may access multiple channels simultaneously, and the original joint channel-and-power allocation problem is reduced to a continuous power-allocation problem by identifying channel occupancy with whether \(p_i^m>0\) [2506.03813]. The received signal on channel \(m\) is written as
\[
y_{i}^m = h_{i,i}^m s_{i} c_{i}^m+\sum_{j \neq i}h_{i, j}^m s_{j} c_{j}^m+n_{i}^m,
\]
with rate
\[
R_i^m(\mathbf{C},\mathbf{P})=\log_2 \left(1+\text{SINR}_{i}^{m}\right),
\]
and the simplified optimization becomes
\[
\begin{aligned}
\underset{\mathbf{P}}{\operatorname{maximize}} \quad & \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_i R_i^m(\mathbf{P}) \\
\text{subject to} \quad & p_i^m \ge 0,\quad \forall i,m, \\
& \sum_{m=1}^{M} p_i^m \le P_{\max},\quad \forall i.
\end{aligned}
\]
Introducing \(v_i^m=\sqrt{p_i^m}\), receive coefficients \(u_i^m\), and weights \(w_i^m\), the paper rewrites the problem as
\[
\begin{aligned}
\underset{\mathbf{W,U,V}}{\operatorname{minimize}} \quad & \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_i (w_i^m e_i^m - \log(w_i^m)) \\
\text{subject to} \quad & \sum_{m=1}^{M}(v_i^m)^2 \le P_{\max},\quad \forall i,
\end{aligned}
\]
with updates
\[
(u_{i}^{m})_{\text{opt}} = \frac{|h_{i,i}^m|v_i^m}{\sum_{j=1}^{D}|h_{i,j}^m v_j^m|^2+(\sigma_i^m)^2},
\qquad
(w_i^m)_{\text{opt}}=(e_i^m)^{-1},
\]
and
\[
v_{i}^{m}=\frac{w_{i}^{m} u_{i}^{m} h_{i,i}^{m}} {\sum_{j=1}^{D} w_{j}^{m}\left|u_{j}^{m} h_{j, i}^{m}\right|^{2}+\lambda_{i}}.
\]
Here \(\lambda_i\) is the multiplier enforcing the total-power budget across channels [2506.03813].

A second, more constrained named eWMMSE formulation adds explicit per-user minimum-rate QoS constraints. The corresponding weighted sum-rate problem is
\[
\begin{aligned}
\underset{\mathbf{P}}{\operatorname{maximize}} \quad & \sum_{m=1}^{M}  \sum_{i=1}^{D}  \alpha_{i} R_{i}^m(\mathbf{P}) \\
\text { subject to } \quad & p_i^m\geq0, \\
&\sum_{m=1}^{M} p_{i}^m\leq P_{\max}, \\
& \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min }, \quad \forall i \in \mathcal{D},
\end{aligned}
\]
and the paper states the rate–WMMSE identity as
\[
R_{i}^{m}=\log w_{i}^{m}-w_{i}^{m} e_{i}^{m} +1,
\qquad
w_i^m=(e_i^m)^{-1}.
\]
The transmit update acquires an additional QoS multiplier \(\mu_i\):
\[
v_{i}^{m}(\lambda_{i},\mu_i)= \frac{(1+\mu_i)w_{i}^{m} u_{i}^{m} h_{i,i}^{m}} {(1+\mu_i)w_{i}^{m} |u_{i}^{m} h_{i,i}^{m}|^2+ \sum_{j \neq i} w_{j}^{m}\left|u_{j}^{m} h_{j, i}^{m}\right|^{2}+\lambda_{i}},
\]
with projected-ascent update
\[
\mu_{i}^{t} = \left[\mu_{i}^{t-1}+\rho_{i}\left(\sum_{m=1}^{M} w_{i}^{m} e_{i}^{m}-\sum_{m=1}^{M} \log w_{i}^{m}-M+R_i^{\min}\right)^{t}\right]^{+},
\]
while \(\lambda_i\) is found by bisection so that the power constraint holds [2509.06395].

The two named eWMMSE variants therefore define a narrow technical meaning of the acronym: WMMSE extended to scalar multi-channel power allocation, with implicit channel assignment via zero/nonzero power, coupled per-user total-power constraints, and optionally explicit QoS constraints. In the QoS-constrained version, the paper states that if the original problem is feasible, eWMMSE is guaranteed to converge to a locally optimal solution, whereas if the problem is infeasible, it produces a suboptimal allocation that minimizes QoS constraint violations while maximizing the total achievable rate [2509.06395].

## 3. Computationally enhanced WMMSE: inverse-free and unfoldable realizations

A large strand of the literature enhances WMMSE by targeting its computational bottlenecks rather than altering the underlying rate–MSE principle. In the MISO downlink, one influential approach replaces the exact beamformer subproblem solution by projected gradient descent. For fixed \((\boldsymbol u,\boldsymbol w)\), the beamformer subproblem is convex in \(\boldsymbol V\), so the update is approximated by
\[
\widetilde{\boldsymbol V}^{(k)} = \boldsymbol V^{(k-1)} - \gamma \nabla f\!\left(\boldsymbol V^{(k-1)}\right),
\qquad
\boldsymbol V^{(k)} = \Pi_{\mathcal C}\!\left\{\widetilde{\boldsymbol V}^{(k)}\right\},
\]
with
\[
\nabla f(\boldsymbol v_i^{(k)}) = -2\alpha_i w_i u_i \boldsymbol h_i + 2 \boldsymbol A \boldsymbol v_i^{(k)},
\]
and projection onto the transmit-power ball
\[
\Pi_{\mathcal C}\{\boldsymbol V\} = \begin{cases} \boldsymbol V, & \operatorname{Tr}(\boldsymbol V\boldsymbol V^H)\le P,\\[1mm] \dfrac{\boldsymbol V}{\|\boldsymbol V\|}\sqrt P, & \text{otherwise.} \end{cases}
\]
This removes the matrix inversion, eigendecomposition, and bisection steps from the classical transmit update and yields complexity \(\mathcal O(L M^3)\) for classical WMMSE versus \(\mathcal O(L K M^2)\) for the unfolded/projected-gradient version [2006.08448].

The same work then unfolds \(L\) outer WMMSE iterations into a neural architecture with \(K\) projected-gradient substeps per layer and trainable step sizes
\[
\boldsymbol \Gamma = [\boldsymbol \gamma_1,\ldots,\boldsymbol \gamma_L],\qquad \boldsymbol \gamma_l=[\gamma_l^{(1)},\ldots,\gamma_l^{(K)}].
\]
Importantly, the trainable parameters are only the PGD step sizes; the paper explicitly states that it does not introduce momentum or Nesterov extrapolation [2006.08448]. At \(10\) dB SNR with \(K=4\), the trainable unfolded WMMSE outperforms classical WMMSE for \(L\le 4\) and performs essentially equally for \(L=5,6\); at \(L=6\), both reach about \(98\%\) of the converged WMMSE sum-rate. At \(20\) dB, increasing the number of PGD steps from \(K=4\) to \(K=8\) restores competitiveness over a larger range of depths [2006.08448].

For general MU-MIMO, matrix-valued receive filters and weights make the inverse-removal problem harder. A later matrix-inverse-free WMMSE variant addresses this by combining gradient descent for \(\mathbf U\) and \(\mathbf V\) with Newton–Schulz iteration for \(\mathbf W\):
\[
\mathbf U_k^{+} = \mathbf U_k - \gamma_u \nabla_u f(\mathbf U_k),
\qquad
\mathbf W_k^{+} = \mathbf W_k\big(2\mathbf I - \mathbf E_k\mathbf W_k\big),
\qquad
\mathbf V_k^{+} = \mathbf V_k - \gamma_v \nabla_v f(\mathbf V_k).
\]
The paper describes this as the first MU-MIMO WMMSE variant free from matrix inverses and other non-parallelizable operations, proves convergence to a stationary point of the original WSR problem, and states that “matrix-inverse-free” means that no direct inverse operation is executed, not that inversion is irrelevant mathematically [2205.08877]. In fully loaded scenarios, unfolded inverse-free configurations such as \((4,2,8)\) and \((8,2,4)\) reach about \(99\%\) of the WMMSE-at-convergence WSR, while IAIDNN reaches about \(91\%\) in the same regime [2205.08877].

These inverse-free and projected-gradient formulations establish a broad sense in which eWMMSE denotes a computational enhancement of WMMSE: the objective and block structure are preserved, but the dominant update is converted from inverse-heavy exact minimization into multiplication-dominated first-order or iterative-inverse primitives that are more compatible with GPUs, ASICs, FPGAs, and deep unfolding.

## 4. Acceleration, surrogate redesign, and reduced-complexity WSR updates

Another enhancement line starts from the observation that WMMSE, WSR-FP, and WSR-MM differ less than their literatures suggest. Comparative analysis shows that WMMSE is a special case of WSR-FP, and that specific WMMSE and WSR-FP instances coincide exactly with WSR-MM through a BCA-to-MM mapping [2311.04546]. This enables new transmit-block designs that target the same bottleneck as classical WMMSE but are derived through surrogate-function flexibility.

In that framework, WSR-MM+ and its BCA counterpart WSR-FP+ replace the exact transmit-block solution by an additional quadratic majorization. For the MIMO case, choosing
\[
\eta_k\ge \lambda_{\max}\!\left(\sum_{j=1}^K \omega_j \mathbf H_{j,k}^{\mathsf H}\mathbf A_j\mathbf H_{j,k}\right)
\]
leads to the subproblem
\[
\min_{\{\mathbf W_i\}\in\mathcal W}\quad \sum_{k=1}^K \eta_k \|\mathbf W_k-\mathbf Q_k\|_F^2,
\]
with solution
\[
\mathbf W_k= \mathbf Q_k \min\left\{ \sqrt{\frac{P_k}{\|\mathbf Q_k\|_F^2}},\,1 \right\}.
\]
For MISO, the analogous update is
\[
\mathbf w_k= \mathbf q_k \min\left\{ \sqrt{\frac{P}{\sum_{k=1}^K \|\mathbf q_k\|^2}},\,1 \right\}.
\]
These updates eliminate repeated inverse or pseudo-inverse evaluations during multiplier search. The resulting MISO complexity drops from \(\mathcal O\big(I(KM^2+M^3)\big)\) for WSR-MM to \(\mathcal O(KM^2)\) for WSR-MM+, while the MIMO complexity drops from \(\mathcal O\big(I(KM^3+M^3)\big)\) to \(\mathcal O(KM^3)\) [2311.04546]. Although the paper does not name this eWMMSE, WSR-FP+ is explicitly the BCA counterpart of WSR-MM+, so it is naturally interpreted as an enhanced WMMSE-like alternating method.

A related acceleration perspective comes from fractional programming. The quadratic transform paper identifies WMMSE as a special case of QT for \(\sum_i \mu_i\log(1+\mathrm{SINR}_i)\), proves a local objective-gap rate of \(O(1/k)\) for conventional QT and its WMMSE specialization, and then shows that a nonhomogeneous QT exposes the method as projected gradient ascent, enabling Nesterov extrapolation. Under Lipschitz-gradient assumptions, the extrapolated QT satisfies
\[
f(\underline{x}^*) - f(\underline{x}^k) \le \frac{2C\,[f(\underline{x}^*)-f(\underline{x}^0)]}{(k+1)^2},
\]
yielding an \(O(1/k^2)\) rate versus \(O(1/k)\) for conventional QT/WMMSE [2312.05726]. This supplies a rate-theoretic interpretation of accelerated eWMMSE-like variants.

A further complementary development is A-MMMSE, which compares directly against conventional WMMSE and its enhanced variant, reduced-WMMSE. A-MMMSE replaces the expensive precoder block by projected gradient descent and introduces a two-stage warm start that first solves the unweighted sum-MSE problem and then switches to weighted WMMSE. The paper states that A-MMMSE matches the WSR performance of both conventional WMMSE and reduced-WMMSE, achieves up to \(5\times\) speedup over WMMSE, up to \(2\times\) speedup over reduced-WMMSE in large-user scenarios, and an additional \(3\times\) GPU acceleration over reduced-WMMSE [2510.20507]. This suggests that “enhancement” can also mean implementation-oriented acceleration around WMMSE-type fixed points rather than a new equivalent transform.

## 5. Learning-based enhanced WMMSE and graph-structured unfolding

Deep unfolding and graph neural parameterization have produced a distinct class of enhanced WMMSE methods in which a small number of trainable layers approximate or improve the behavior of many classical iterations. In UWMMSE for SISO power allocation, the WMMSE recursion is preserved but the \(w\)-update is modulated by GNN outputs:
\[
\mathbf{a}^{(k)}=\Psi(\mathbf{H};\theta_a^{(k)}),\qquad \mathbf{b}^{(k)}=\Psi(\mathbf{H};\theta_b^{(k)}),
\]
\[
w_i^{(k)}= \frac{a_i^{(k)}}{1-u_i^{(k)}h_{ii}v_i^{(k-1)}+b_i^{(k)}},
\qquad
v_i^{(k)}= \alpha\!\left( \frac{u_i^{(k)}h_{ii}w_i^{(k)}}{\sum_j h_{ji}^2 (u_j^{(k)})^2 w_j^{(k)}} \right).
\]
The architecture is permutation equivariant through the graph construction, and with \(K=4\) layers the reported average sum-rates are \(83.21\) for UWMMSE, \(82.94\) for WMMSE, and \(76.49\) for truncated WMMSE, while test-time runtimes are \(2.0\) ms/sample for UWMMSE and \(16\) ms/sample for WMMSE [2009.10812].

A related knowledge-driven D2D allocator uses a WMMSE-unrolled GNN with two cascaded message-passing modules per layer, one aligned with \((u,w)\) updates and one with the \(v\)-update. In the reported normalized performance table, UWGNN attains \(102.817\%\), \(102.996\%\), and \(102.256\%\) relative performance for user counts \(10\), \(30\), and \(50\), respectively, while a prior UWMMSE baseline degrades to \(92.168\%\) and \(90.497\%\) at \(30\) and \(50\) users [2307.05882]. This indicates that graph-structured enhancement is particularly valuable for size-varying and topology-varying interference networks.

Stability analysis of unfolded WMMSE adds an important counterweight to pure performance claims. For UWMMSE under bounded additive perturbations \(H\mapsto \tilde H=H+E\), the paper derives a recursive bound on \(|v_i^{(k)}-\tilde v_i^{(k)}|\) in terms of numerator and denominator perturbation terms, previous-layer errors, and bounded GNN-output changes [2110.07471]. Empirically, over \(800\)k test nodes, about \(66\%\) have an upper-bound margin close to \(0\), about \(30\%\) have margin greater than \(2.0\), and about \(4\%\) violate the bound, which the paper attributes to higher-order terms [2110.07471]. A common misconception is therefore that learning-based eWMMSE variants are merely faster black-box surrogates; this literature instead presents structured, often interpretable WMMSE modifications with explicit perturbation analyses.

Taken together, the unfolding literature suggests a specific meaning of enhanced WMMSE in finite-latency settings: preserve the \(u\)–\(w\)–\(v\) semantics of WMMSE, but train a small number of channel-conditioned parameters so that truncated execution is no longer a crude approximation of the asymptotic algorithm.

## 6. Application-specific extensions, limitations, and terminology

The eWMMSE idea has also expanded into application-specific formulations. In integrated sensing and communications, a mutual-information-based WMMSE framework maximizes a weighted sum of communication MI and sensing MI under clutter by introducing receive beamformers \(\mathbf B_k,\mathbf B_\tau\), MSE matrices \(\mathbf E_k,\mathbf E_\tau\), and weight matrices \(\mathbf W_k,\mathbf W_\tau\), so that the sensing term enters the same log-det/MMSE machinery as the communication terms [2310.12686]. In movable-antenna networks, standard WMMSE transceiver updates are embedded inside a block-coordinate framework with majorization-minimization updates for antenna positions, and the planar movement mode reduces computation time by approximately \(30\%\) at a little performance expense [2404.09734]. In RIS-assisted MU-MISO, WMMSE is made compression-aware by recomputing the effective channel from the actual decompressed RIS phase shifts before updating the AP beamformer, and the paper states that accounting for phase-shift compression errors during beamforming significantly improves the sum-rate performance even when the number of control bits is lower than the number of RIS elements [2510.05438].

Robust incomplete-CSI design in FDD provides another eWMMSE-like extension. There the classical rate objective is replaced by a training-aware lower bound based on LMMSE channel estimation, yielding the augmented weighted average MSE
\[
\bar\xi_k = u_k \bar\varepsilon_k - \log_2 u_k,
\qquad
\bar\xi_k^{\mathrm{MMSE}} = 1-\bar R_k,
\]
and a fully closed-form precoder update
\[
\mathbf P^{\mathrm{unconst}} = \left( \mathbf Z+\hat{\mathbf H}\mathbf G^\ast\mathbf U\mathbf G\hat{\mathbf H}^H +\frac{\mathrm{tr}(\mathbf U\mathbf G\mathbf G^\ast)}{P_{\mathrm{dl}}}\mathbf I \right)^{-1} \hat{\mathbf H}\mathbf G^\ast\mathbf U.
\]
The paper reports runtime approximately \(10\) times smaller than SIWMMSE at \(40\) dB while retaining strong robustness under few-pilot incomplete-CSI scenarios [2312.01888]. By contrast, in holographic MIMO RS-NOMA ISAC, a paper uses the exact term “Enhanced WMMSE (E-WMMSE)” but only as a partially specified baseline, explicitly stating that it extends conventional WMMSE to handle RS-NOMA with sensing constraints while not providing a full rigorous derivation, convergence proof, or explicit sensing-augmented WMSE formulation [2512.19699].

Three limitations recur across the literature. First, enhanced WMMSE methods remain local or stationary-point methods for nonconvex problems; none of the surveyed variants claim global optimality. Second, learning-based variants depend on training distributions, fixed layer counts, and often fixed system dimensions, although graph-based designs mitigate this to some extent [2006.08448], [2009.10812]. Third, the term itself is not standardized: some papers reserve eWMMSE for multi-channel JCPA with or without QoS constraints, while others describe projected-gradient, inverse-free, accelerated, or application-specific designs as enhanced WMMSE in substance but not in name [2509.06395], [2205.08877].

In contemporary usage, eWMMSE is therefore best understood as a technical category rather than a single algorithm. At its core lies the classical WMMSE equivalence between weighted sum-rate and weighted MSE. What changes across the literature is the surrounding architecture: multi-channel power coupling, QoS pricing, matrix-inverse-free realizations, projected-gradient beamformer steps, extrapolated surrogate updates, graph-conditioned unfoldings, or sensing- and geometry-aware extensions. The unifying principle is not the acronym but the preservation of WMMSE structure while improving feasibility, latency, hardware-friendliness, or task specificity.

Source: https://www.emergentmind.com/topics/enhanced-wmmse-ewmmse