Null-Space Flow Matching for MIMO Channel Estimation in Latency-Constrained Systems

Abstract: Accurate yet low-latency channel state information (CSI) acquisition is essential for multiple-input multiple-output (MIMO) communication systems. While advanced deep generative models, such as score-based and diffusion models, enable high-fidelity CSI reconstruction from limited pilot observations, they often suffer from high inference latency. To achieve accurate CSI estimation under stringent latency constraints, this paper proposes a null-space flow matching (FM) framework that decomposes pilot-limited MIMO channel estimation into a range-space reconstruction problem and a null-space generation problem. Specifically, the range-space component of the channel is directly recovered from noisy pilot observations, while only the ambiguous null-space component is iteratively refined using an FM-based generative prior. To further improve the robustness of the proposed framework, we introduce a power-law time schedule to better allocate the limited number of refinement steps, along with a noise-aware adaptive correction strategy to suppress channel noise on the refinement trajectory. Experimental results demonstrate that our method achieves a competitive normalized mean square error (NMSE) even under a strict latency budget of around 3 ms, while delivering superior estimation accuracy and faster inference than both model-based and generative baselines.

• The paper introduces a novel null-space flow matching method that decomposes MIMO channel estimation into range-space recovery and null-space generative refinement to drastically reduce latency.
• The method employs a power-law non-uniform time schedule and noise-aware adaptive correction, achieving NMSE below -20 dB in ~3 ms and outperforming diffusion models by a 10-fold reduction in inference time.
• Practical implications include robust performance in low-SNR and low-pilot-density settings, making it suitable for real-time 6G wireless systems.

Null-Space Flow Matching for Latency-Constrained MIMO Channel Estimation

Motivation and Problem Setting

Estimation of channel state information (CSI) in MIMO systems is a foundational task for advanced wireless communication, especially as antenna counts and user mobility increase in emerging 6G networks. The tight constraints on pilot overhead and channel coherence time, exacerbated by high-mobility scenarios, pose a severe bottleneck for acquiring accurate CSI within operational latency bounds. Traditional model-based approaches, such as LMMSE estimators and Gaussian mixture models, are limited by parametric assumptions and inflexibility in complex environments. Recent deep generative models, including VAEs, score-based, and diffusion models, deliver high-fidelity channel reconstructions but suffer prohibitive inference latency. The central task is to reconcile high-quality CSI estimation with stringent latency constraints, particularly when rapid adaptation is required.

Null-Space Flow Matching Framework

This paper introduces a methodological advance by decomposing the pilot-limited MIMO channel estimation problem into range-space and null-space subproblems. The range-space component—directly observable from pilot transmissions—is recovered from noisy pilot data via pseudo-inverse projection, guaranteeing measurement consistency. In contrast, the null-space component, representing underdetermined channel features invisible to the measurement matrix, is iteratively generated using a flow matching (FM) generative prior. FM, in contrast to stochastic dynamics of score-based and diffusion models, learns a deterministic velocity field for sample transport, facilitating larger step sizes and reducing discretization error, thereby improving inference efficiency.

Notably, the paper introduces a power-law non-uniform time schedule, allocating finer refinement steps to early stages of inference where discretization error is most impactful. Further, a noise-aware adaptive correction strategy modulates measurement consistency enforcement based on the reliability of the FM prediction versus pilot observation, suppressing noise injection into generative trajectories. The method ensures pilot consistency throughout inference while leveraging a learned velocity field (U-Net architecture) for null-space refinement.

Numerical Results and Comparative Analysis

The experimental evaluation employs realistic channel models (3GPP TR 38.901, CDL-C) on a 64x16 antenna configuration, covering diverse SNR and pilot density regimes. The performance metric is NMSE, averaged over 20,000 test samples.

Strong Numerical Results:

• Under a strict time budget ($\sim$3 ms), the proposed method consistently achieves NMSE below -20 dB, maintaining measurement-consistent generation even in latency-constrained settings.
• When compared to diffusion posterior sampling [8], which requires $\sim$30 ms to reach -20 dB NMSE, the method attains an approximate 10-fold latency reduction and remains on the Pareto front of the accuracy-latency tradeoff.
• Across all SNR and pilot density operating points, null-space flow matching outperforms both single-step model-based estimators and full-space generative baselines, even with ample computational budget.
• At pilot densities as low as $\alpha=0.125$, NMSE remains at -10.61 dB, demonstrating robust underdetermined regime performance.
• The method’s advantage is most pronounced in low-SNR and low-pilot-density conditions, attributed to pilot consistency preservation and noise-aware correction.
• Iterative baselines (SGM, DPS, FM-PGD) exhibit greater sensitivity to the latency budget, incurring sharp NMSE degradation as coherence-time constraints tighten.
• Full-space FM (FM-PGD) delivers inferior accuracy-latency tradeoff relative to null-space FM, underscoring the benefit of avoiding redundant generative refinement of pilot-determined range-space features.

Implications and Theoretical Significance

The null-space flow matching paradigm embodies a principled approach to ill-posed linear inverse problems in latent-variable generative modeling by explicitly decoupling measurement-determined and unconstrained channel subspaces. The method advances generative channel estimation by focusing refinement capacity on genuinely uncertain features, ensuring superior pilot consistency and reducing inference complexity. The adaptive correction mechanism, grounded in uncertainty quantification, introduces a theoretically sound solution to noise injection concerns inherent in iterative generative inference.

Practically, the approach is immediately applicable to real-time MIMO systems facing rapidly varying channel conditions and severe latency constraints. The flow matching trajectory’s geometric efficiency and deterministic ODE formulation position the method as a strong candidate for deployment in hardware-constrained environments and future 6G architectures. The results indicate applicability to low-pilot-density and low-SNR regimes, critical for massive MIMO and highly mobile network slices.

Future Directions

Further research is warranted to extend null-space flow matching to wideband and multi-user contexts, potentially integrating environment-aware or multimodal sensing [17] for improved generalization. Exploration of adaptive time scheduling algorithms and tighter integration with hardware accelerators may further reduce latency and computational overhead. The incorporation of transfer learning and federated training on distributed channel datasets may address non-stationary propagation environments in edge applications.

Conclusion

Null-space flow matching enables accurate, measurement-consistent, and low-latency MIMO channel estimation by decomposing the problem into range-space recovery and null-space generative refinement, driven by flow-matching priors. Power-law time scheduling and noise-aware correction substantially enhance robustness under tight coherence-time budgets. Empirical results establish the approach as superior in both accuracy and latency tradeoff compared to model-based and generative baselines. Theoretical and practical implications are significant for the design of next-generation real-time wireless communication systems (2604.22005).