In the transition toward 6G, wireless networks are scaling from "Massive MIMO" to "Extremely Large-Scale MIMO" (XL-MIMO) and holographic surfaces. These systems use hundreds or even thousands of antennas to focus energy into precise "pencil beams." However, a fundamental mathematical wall stands in the way: the matrix inversion.
Most optimal signal processing techniques—whether for detecting a user's signal (uplink) or precoding data to avoid interference (downlink)—rely on the Minimum Mean Square Error (MMSE) criterion. Mathematically, this requires inverting a matrix that represents the channel state. As Zhe Wang and Jiayi Zhang note in their tutorial on XL-MIMO, the computational cost of this inversion scales cubically ($O(N^3)$) with the number of antennas or users. In a world of sub-millisecond latency requirements, $O(N^3)$ is a death sentence for real-time hardware.
Researchers have converged on a shared mission: finding ways to achieve near-optimal performance while bypassing the need for exact, high-dimensional matrix inversions.
The Polynomial Approximation Path
One of the most intuitive ways to avoid an inversion is to replace it with a series of simpler operations. If we think of a matrix inversion as $1/x$, we can approximate it using a Taylor-like expansion.
Axel Müller, Abla Kammoun, and Emil Björnson pioneered the use of Truncated Polynomial Expansion (TPE). Instead of computing $(H^H H + \sigma^2 I)^{-1}$ directly, TPE expresses the inverse as a sum of matrix powers: $w_0 I + w_1 G + w_2 G^2 \dots$, where $G$ is the Gram matrix.
This works because, in massive MIMO, the channel matrices become "hardened"—their eigenvalues concentrate in a predictable way. By using Random Matrix Theory to pick the optimal weights ($w$), Nafiseh Shariati and colleagues showed in their PEACH algorithm that you can get 99% of the performance of an exact inversion using only a few terms. This shifts the complexity from cubic to quadratic, which is far more manageable for FPGAs. Similarly, Michael Wu demonstrated that a Neumann series approximation could enable uplink speeds of 600 Mb/s by replacing inversions with pipelined matrix-vector multiplications.
Exploiting Statistical and Physical Structure
Another strategy is to stop trying to invert the "instantaneous" channel every few milliseconds and instead look at the bigger picture.
The physical environment (buildings, streets) changes much slower than the phase of a radio wave. David Neumann and Wolfgang Utschick proposed the Optimal Bilinear Equalizer (OBE), which performs interference suppression based on long-term channel covariance matrices. Because these statistics are stable, the "heavy" math can be done once and reused for many data packets. Zhe Wang extended this to cell-free systems, showing that precomputed statistical matrices can replace real-time inversions without losing spectral efficiency.
Researchers also exploit the mathematical structure of the matrices themselves. Chuan Zhang and Xiaohu You introduced a Tridiagonal Matrix Approximation (TMA). By recognizing that most of the "energy" in a correlated massive MIMO channel is concentrated near the diagonal, they can ignore the distant off-diagonal elements, turning a massive inversion into a simple, hardware-friendly tridiagonal solve.
Learning the Inverse through Deep Unfolding
The most recent frontier involves "unfolding" classical iterative algorithms into neural networks. Traditional solvers like Conjugate Gradient (CG) or Approximate Message Passing (AMP) take many steps to converge, and each step might still involve a matrix operation.
Yi Wei and Ming-Min Zhao developed LcgNet, which takes the Conjugate Gradient algorithm and replaces its fixed parameters with learnable weights. By training the network, it learns to "jump" to the solution in far fewer iterations than the standard algorithm.
Similarly, Hengtao He and Jun Zhang proposed AMP-GNN, which uses Graph Neural Networks to refine the messages passed in the AMP algorithm. These "model-driven" deep learning approaches provide a middle ground: they use the known physics of the wireless channel (the model) to keep the network small and fast, while using learning to bypass the heavy computational bottlenecks of classical matrix algebra.
Why It Works: The Blessing of Dimensionality
The reason these approximations work so well is a phenomenon often called the "Blessing of Dimensionality." In small MIMO systems (e.g., 2x2), every element of the matrix is critical. But as the number of antennas grows to infinity, the Gram matrix $H^H H$ tends toward a diagonal matrix (channel hardening).
As Charles Jeon showed with Gram-Matrix Interpolation, the channel becomes so predictable across frequencies that you don't even need to compute the matrix for every subcarrier—you can compute it for a few and interpolate the rest. By leaning into the structure that emerges at scale, researchers are turning the "massive" nature of 6G from a computational burden into a simplifying advantage.