Explicit CSI Feedback

Unlike codebook-based precoding where the UE selects a pre-defined index (PMI), explicit feedback transmits a quantized representation of the channel matrix or its eigenvectors directly. This removes the quantization error inherent in finite codebooks. The receiver estimates H (the complex channel matrix), compresses it, and feeds back coefficients representing spatial directions and gains. This allows the gNB to compute arbitrary precoders, crucial for multi-user MIMO (MU-MIMO) where nulling interference requires precise channel knowledge beyond standardized codebook granularity.

A common explicit approach transmits the dominant eigenvectors of the channel covariance matrix rather than the raw channel coefficients. The UE performs an eigenvalue decomposition of H^HH and feeds back the eigenvectors corresponding to the strongest eigenvalues. This represents the optimal beamforming directions. Compression is achieved by reporting only a subset of eigenvectors (rank indication) and applying transform coding or quantization to the eigenvector coefficients. This method directly conveys the spatial subspace structure of the channel.

In time-domain explicit feedback, the UE reports the quantized Channel Impulse Response taps. This captures the multipath profile directly, including the delay spread and complex gain of each resolvable path. The gNB can reconstruct the full frequency-domain channel via Fourier transform, enabling per-subcarrier precoding. This method is particularly relevant for OFDM systems where frequency selectivity must be precisely matched. The overhead scales with the number of significant delay taps, which can be sparse in millimeter-wave channels.

The primary engineering challenge of explicit feedback is the uplink overhead. A full MIMO channel matrix contains a large number of complex values. To make this practical, explicit schemes employ:

• Element-wise quantization: Scalar or vector quantization of individual channel coefficients.
• Transform-domain compression: Applying DFT or Karhunen-Loève Transform (KLT) to concentrate energy in fewer coefficients before quantization.
• Deep learning compression: Autoencoder architectures like CsiNet learn a compact latent representation, dramatically reducing the feedback payload while preserving reconstruction accuracy.

Explicit feedback is essential in Frequency Division Duplex (FDD) systems where channel reciprocity does not hold. In FDD, the uplink and downlink operate on different frequency bands, so the gNB cannot infer the downlink channel from uplink measurements. The UE must explicitly measure downlink CSI-RS pilots and report the quantized channel. In contrast, Time Division Duplex (TDD) systems can leverage reciprocity, reducing the need for explicit feedback. This makes explicit CSI a critical enabler for massive MIMO in FDD spectrum bands.

5G NR Type-II codebook represents a standardized middle ground between purely implicit and fully explicit feedback. It uses a linear combination of multiple DFT beams with amplitude and phase quantization per sub-band. While technically codebook-based, its high granularity approximates explicit eigenvector feedback. The UE reports beam indices, combination coefficients, and tap delays, allowing the gNB to reconstruct a detailed precoder. This hybrid approach balances feedback overhead with spatial resolution, making it the practical workhorse for high-performance MU-MIMO in 5G.

The high-resolution 5G NR codebook that underpins advanced explicit feedback. It reports a linear combination of multiple orthogonal beams with amplitude and phase quantization for each sub-band.

• Provides spatial and frequency granularity for multi-user MIMO
• Significantly higher overhead than Type-I, motivating AI-based compression
• Defined in 3GPP Release 15 and enhanced in Release 16 with frequency domain compression

The process of reducing the uplink payload required to convey explicit channel information. Modern approaches replace traditional transform coding with neural network autoencoders.

• Encoder at UE: Compresses raw channel matrix into a low-dimensional latent vector
• Decoder at gNB: Reconstructs the channel with minimal distortion
• CsiNet demonstrated that deep learning can outperform classical compressive sensing for this task

The degradation of CSI accuracy between the measurement instant and the actual downlink transmission. In explicit feedback, the processing delay at the UE plus the feedback latency creates a window of staleness.

• Critical in high-mobility scenarios (vehicular, high-speed rail)
• Drives the need for CSI prediction algorithms
• Measured by the coherence time of the channel relative to the feedback interval

A seminal deep learning framework that treats explicit CSI feedback as an autoencoder problem. The UE encoder compresses the channel matrix, and the gNB decoder reconstructs it.

• Uses convolutional neural networks to exploit spatial correlation
• Subsequent variants (CsiNet-LSTM, CsiNet+) add temporal memory for time-varying channels
• Demonstrates superior NMSE performance compared to compressive sensing baselines at equivalent compression ratios

A property in Time Division Duplex (TDD) systems where the downlink channel can be inferred from uplink measurements. This reduces reliance on explicit feedback but is imperfect due to hardware mismatches.

• Assumes identical propagation in both directions
• Requires calibration of transmit/receive RF chains
• In Frequency Division Duplex (FDD), reciprocity does not hold, making explicit feedback essential

The feedback index that the UE sends to recommend a specific precoding matrix for downlink beamforming. In explicit feedback, the PMI is derived from the quantized channel estimate.

• Part of the broader CSI report alongside CQI and RI
• Codebook-based PMI limits feedback to predefined matrices
• Explicit feedback can provide richer spatial information beyond codebook constraints