Sampling-Free Diffusion Transformers for Low-Complexity MIMO Channel Estimation

Diffusion model-based channel estimators have shown impressive performance but suffer from high computational complexity because they rely on iterative reverse sampling. This paper proposes a sampling-free diffusion transformer (DiT) for low-complexity MIMO channel estimation, termed SF-DiT-CE. Exploiting angular-domain sparsity of MIMO channels, we train a lightweight DiT to directly predict the clean channels from their perturbed observations and noise levels. At inference, the least square (LS) estimate and estimation noise condition the DiT to recover the channel in a single forward pass, eliminating iterative sampling. Numerical results demonstrate that our method achieves superior estimation accuracy and robustness with significantly lower complexity than state-of-the-art baselines.

💡 Research Summary

The paper addresses the high computational burden of diffusion‑model‑based MIMO channel estimators, which traditionally rely on iterative reverse sampling (often dozens or hundreds of neural function evaluations) to denoise a noisy observation and recover the channel matrix. To overcome this bottleneck, the authors propose a Sampling‑Free Diffusion Transformer for Channel Estimation (SF‑DiT‑CE), a framework that requires only a single forward pass of a lightweight diffusion transformer (DiT) at inference time.

Key technical contributions are as follows:

1. Angular‑domain sparsity exploitation – The complex MIMO channel matrix H is transformed into the angular domain using unitary DFT matrices at the transmitter and receiver. In this domain the channel exhibits strong sparsity and low‑dimensional manifold structure, which simplifies learning and improves generalization. The complex angular‑domain channel is then represented as a two‑channel real‑valued image (real and imaginary parts) suitable for convolutional and transformer processing.

2. VE (Variance‑Exploding) forward noise model – The authors adopt the VE diffusion formulation, where the forward process adds Gaussian noise with increasing variance while leaving the signal unscaled (X_t = X + σ_t ε). This choice aligns perfectly with the least‑squares (LS) estimate of the channel (b_H^LS = H + N P^H), which can be interpreted as the clean channel corrupted by additive white Gaussian noise of variance σ². Consequently, the training corruption exactly matches the statistical nature of the LS input used at inference, eliminating the model‑data mismatch that plagues VP‑based approaches.

3. X‑prediction objective – Instead of predicting the noise (ε‑prediction) or the diffusion velocity (V‑prediction), the network is trained to directly output the clean channel image X_0 from a noisy input X_t and the corresponding noise level σ_t. This “clean‑signal” prediction leverages the low‑dimensional manifold assumption of MIMO channels and reduces learning difficulty. The loss is formulated as a velocity loss (V‑loss), which is algebraically equivalent to an X‑prediction loss but provides stable gradients.

4. Lightweight DiT architecture – The diffusion transformer consists of only two transformer blocks, each with a hidden dimension of 128, patch size 4, and eight attention heads. Input tokens are created by patchifying the image and embedding both spatial position (2‑D sinusoidal embedding) and the scalar noise level (sinusoidal embedding → scale‑shift parameters). Conditional layer‑normalization and gated residual connections allow the model to adapt its denoising strength according to σ_t while keeping the parameter count modest.

5. Sampling‑free inference pipeline – At test time the procedure is:

   • Compute the LS estimate from the received pilots.
   • Transform the LS estimate to the angular domain and form the real‑imaginary image.
   • Feed this image together with the known noise variance σ² into the trained DiT.
   • Obtain the denoised angular‑domain channel in a single forward pass.
   • Convert back to the spatial domain to produce the final CSI estimate. This eliminates the iterative reverse diffusion steps entirely, reducing the number of neural function evaluations from O(10‑100) to 1.

6. Experimental validation – Simulations use 3GPP CDL‑C and CDL‑D channel models with a (64 × 16) uniform linear array at 40 GHz. Training data consist of 10,000 channel realizations per profile; testing uses 100 realizations. The proposed SF‑DiT‑CE achieves up to 5.6 dB NMSE improvement over the LS baseline and outperforms LMMSE, prior VE/VP diffusion estimators, and recent GAN‑based methods, especially at low SNR. In terms of computational cost, the single‑pass design yields more than a ten‑fold reduction in inference latency compared with existing diffusion approaches, making it suitable for real‑time wireless systems.

In summary, the paper introduces a novel combination of VE‑aligned noise modeling, direct clean‑signal prediction, angular‑domain sparsity, and a compact diffusion transformer to deliver high‑accuracy, low‑complexity MIMO channel estimation. The sampling‑free paradigm removes the primary obstacle to deploying diffusion‑based priors in practical communication receivers and opens avenues for extensions to massive MIMO, multi‑user scenarios, and hardware‑efficient implementations.