CSI Compression

A seminal deep learning framework that treats CSI compression as an encoder-decoder problem. The user equipment (UE) uses a convolutional neural network to compress the channel matrix into a low-dimensional codeword vector, which is transmitted over the feedback link. The base station then uses a mirror decoder network to reconstruct the full channel.

• Encoder: Compresses spatial-frequency CSI into a compact latent representation
• Decoder: Reconstructs high-dimensional channel from the received codeword
• NMSE performance: Achieves superior reconstruction quality compared to compressive sensing baselines at compression ratios of 4x to 16x
• Variants: CsiNet+ uses convolutional long short-term memory layers to capture temporal correlation across multiple time slots

A model-based approach that exploits the inherent sparsity of the channel in the angular-delay domain. Rather than sampling the full channel matrix, the UE transmits a small set of linear measurements. The base station solves an ℓ₁-norm minimization problem to recover the sparse channel representation.

• Sparsity basis: Channel is sparse in the Discrete Fourier Transform (DFT) domain due to limited scattering clusters
• Measurement matrix: Random Gaussian or structured random projections reduce dimensionality
• Reconstruction algorithms: Orthogonal Matching Pursuit (OMP), Basis Pursuit Denoising (BPDN), and Approximate Message Passing (AMP)
• Limitation: Performance degrades at very low compression ratios where the restricted isometry property (RIP) is violated

Techniques that jointly optimize compression and finite-rate quantization to account for the limited capacity of the feedback link. Standard autoencoders assume infinite-precision codewords, but real systems require discretizing the latent vector into a finite number of bits.

• Uniform quantization: Simple scalar quantization of each latent dimension with trainable step sizes
• Vector quantization (VQ): Maps the continuous latent vector to the nearest entry in a learned codebook of discrete embeddings
• VQ-VAE for CSI: Combines vector quantization with variational autoencoders to learn a discrete latent space optimized for reconstruction fidelity
• End-to-end training: Uses straight-through estimator or Gumbel-Softmax reparameterization to backpropagate through non-differentiable quantization operations

Exploits the temporal correlation of the channel by encoding only the difference between consecutive CSI snapshots. Since the channel evolves smoothly in low-to-moderate mobility scenarios, the residual signal is sparser and requires fewer bits to encode than the full channel matrix.

• Differential encoding: Transmits Δh(t) = h(t) − h(t−1) instead of h(t)
• Recurrent architectures: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks learn to predict and compress the temporal residual
• Transformer-based: Self-attention mechanisms capture long-range dependencies across multiple time steps for improved prediction accuracy
• Mobility robustness: Performance degrades gracefully as Doppler spread increases, with adaptive compression rates based on estimated velocity

A structured feedback approach where the precoding matrix indicator (PMI) is selected from a hierarchical codebook with varying spatial granularity. The UE first selects a wideband beam group, then refines with subband-specific beam selection and co-phasing coefficients.

• Type-I codebook: Low-resolution, single-beam selection with low overhead, suitable for single-user MIMO
• Type-II codebook: High-resolution, multi-beam linear combination with amplitude and phase quantization per subband for multi-user MIMO
• Enhanced Type-II (eType-II): Adds frequency-domain compression via Discrete Fourier Transform (DFT) basis to reduce subband reporting overhead
• Port selection codebook: Extends Type-II to scenarios where the UE selects a subset of CSI-RS ports before beam combination

Applies Slepian-Wolf theorem principles to compress correlated CSI across multiple base stations without inter-BS communication. Since nearby UEs experience correlated shadowing and scattering environments, their channel matrices exhibit statistical dependencies that can be exploited for joint compression.

• Wyner-Ziv coding: Encodes CSI with side information available only at the decoder (base station), not the encoder (UE)
• Distributed autoencoders: Each UE independently compresses its channel; a central decoder jointly reconstructs all channels by modeling cross-user correlation
• Application: Cell-free massive MIMO systems where a central processing unit fuses compressed CSI from geographically distributed access points
• Rate-distortion trade-off: Achieves lower aggregate feedback rate than independent compression when inter-user correlation exceeds 0.3

The foundational deep learning framework for CSI compression, using an autoencoder structure where the encoder at the user equipment compresses the channel matrix and the decoder at the base station reconstructs it. Key aspects:

• Uses convolutional neural networks to extract spatial features
• Achieves significantly higher reconstruction quality than compressive sensing at low compression ratios
• The original 2018 paper demonstrated a 4x to 16x reduction in feedback overhead

A high-resolution 5G NR codebook that provides detailed spatial and frequency granularity for multi-user MIMO precoding. Unlike Type-I, it combines multiple orthogonal beams with amplitude and phase scaling per sub-band. Compression is critical here because:

• Reporting full Type-II coefficients creates massive uplink overhead
• 3GPP Release 16 introduced frequency-domain compression via Discrete Fourier Transform basis subset selection
• Enhanced Type-II (eType-II) further reduces overhead using tap-based time-domain compression

A feedback mechanism where the receiver reports quantized channel matrix coefficients or eigenvectors directly, rather than selecting a pre-defined index from a codebook. This approach:

• Provides the base station with richer channel information for advanced precoding
• Requires aggressive compression to be practical, making it a primary use case for neural autoencoders
• Enables non-codebook-based transmission in reciprocity-challenged FDD systems

The standard performance metric for evaluating CSI compression and reconstruction quality. NMSE quantifies the error between the original channel matrix H and the reconstructed matrix Ĥ, normalized by the power of H. Key considerations:

• Lower NMSE indicates better preservation of spatial structure
• Typical targets for massive MIMO are below -10 dB
• Must be evaluated jointly with cosine similarity to ensure beamforming direction is preserved

The phenomenon where CSI becomes outdated between measurement and transmission due to node mobility. Compression interacts with aging because:

• Compressed feedback introduces additional latency from encoding and decoding
• Neural networks can be trained to perform joint compression and prediction, forecasting future channels while reducing overhead
• This is critical in vehicular and high-speed rail scenarios where Doppler shifts exceed 500 Hz

A model-driven technique that unrolls iterative optimization algorithms (like ISTA or AMP) into neural network layers. Benefits for CSI compression:

• Combines the interpretability of classical signal processing with the learning capacity of deep networks
• Requires fewer trainable parameters than pure black-box autoencoders
• Converges in a fixed number of iterations, enabling deterministic latency for real-time feedback