🔢 mathematics

Vectorized Generalized Nearest Neighbor Decoding for In-block Memory Channel

This paper extends generalized nearest neighbor decoding to vectorized form for in-block memory channels, providing analytical characterizations, closed-form optimality conditions, and a joint design framework that demonstrates significant performance gains over conventional baselines in realistic communication scenarios.

Original authors: Yuhao Liu, Xinwei Li, Shuqin Pang, Hao Wu, Wenyi Zhang

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuhao Liu, Xinwei Li, Shuqin Pang, Hao Wu, Wenyi Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a secret message to a friend across a noisy, chaotic room. In the ideal world of textbook physics, the room is empty, and your voice travels in a straight line. But in the real world, the room is full of echoes, people shouting, and the air itself might be shifting.

This paper tackles a specific problem in communication: How do you decode a message when the noise isn't just random static, but has a pattern that changes over short bursts of time?

Here is a breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Block" of Noise

Usually, engineers treat every single word (or bit) of a message as an independent event. They assume the noise hitting your ear at this exact second has nothing to do with the noise hitting your ear a millisecond later.

But in reality, noise often comes in blocks.

The Analogy: Imagine you are listening to a song on a radio that has a bad connection. The static doesn't just pop in and out randomly; it might be a low hum that lasts for the entire duration of a specific musical phrase, or a sudden burst of wind that distorts a whole sentence.
The Paper's Term: This is called an "In-Block Memory" (IBM) channel. The noise inside one "block" of time is related to itself, even if the noise in the next block is totally different.

2. The Old Way: Looking at One Brick at a Time

For a long time, the best way to decode these messages was to look at each "brick" (each symbol) individually and guess what it was based on the noise right next to it. This is like trying to figure out what a blurry photo says by squinting at one pixel at a time.

The paper points out that this is inefficient. If you know that a whole sentence was distorted by the same wind gust, you should look at the whole sentence together to figure out what the wind did, rather than guessing each word in isolation.

3. The New Solution: The "Vectorized" Decoder

The authors propose a new decoding method called Vectorized Generalized Nearest Neighbor Decoding (Vec-GNND).

The Analogy: Instead of looking at one brick, imagine you have a mold that fits the entire sentence.
- Old Method: You try to fit a single, small, rigid piece of clay to every single letter. If the wind blew the letters sideways, your rigid piece doesn't fit.
- New Method (Vec-GNND): You have a flexible, smart mold. When the wind distorts the sentence, your mold stretches and shifts all at once to match the distortion. It looks at the shape of the whole block of data.

This new decoder uses two special tools:

A Processing Function (The Lens): It looks at the received signal and "cleans" it up, figuring out how the noise warped the shape.
A Scaling Function (The Ruler): It adjusts the size of the message to match what it should have been, compensating for the noise.

4. The "Secret Sauce": Learning the Shape of the Noise

The paper does something clever: it doesn't just guess how to fix the message. It calculates the perfect way to fix it based on the statistics of the noise.

The Analogy: Imagine you are a detective trying to reconstruct a shattered vase.
- If you know the vase was dropped on a soft carpet (one type of noise), you know the pieces will be scattered gently.
- If you know it was dropped on concrete (another type of noise), the pieces will be scattered violently.
- The paper's math acts like a detective who instantly calculates: "Based on the pattern of the shards I see right now, the vase was dropped on this specific type of surface. Therefore, I will glue the pieces together exactly this way."

They derived a mathematical formula that tells the decoder exactly how to stretch, shrink, and rotate the received signal to get the most accurate message back.

5. The Results: Why It Matters

The authors tested their new "smart mold" against the old "one-brick-at-a-time" method in two specific scenarios:

Block Noncoherent Channels: Where the whole message gets rotated by an unknown angle (like a compass spinning).
Phase Noise Channels: Where the signal gets wobbly and distorted over time (like a shaky camera).

The Finding:
In these scenarios, the new Vec-GNND method significantly outperformed the old methods.

The Analogy: In the high-noise tests, the old method was like trying to read a sign in a hurricane by looking at one letter at a time—you missed everything. The new method was like holding the sign steady and reading the whole word at once. It recovered much more information, especially when the noise was strong.

Summary

This paper introduces a smarter way to listen to noisy signals. Instead of treating every tiny piece of data as an isolated event, it treats groups of data as a single, connected unit. By using a mathematical "mold" that adapts to the specific shape of the noise in real-time, it can recover messages much more accurately than previous techniques, particularly in environments where noise comes in predictable bursts.

Technical Summary: Vectorized Generalized Nearest Neighbor Decoding for In-block Memory Channels

Problem Statement
Modern communication systems often operate over channels that deviate from the idealized additive white Gaussian noise (AWGN) model. Practical environments frequently exhibit statistical dependence within finite transmission blocks due to phenomena such as bandlimitation, multipath propagation, phase noise, and mobility. These scenarios are modeled as in-block memory (IBM) channels, where symbols within a block are statistically dependent, while blocks themselves are independent.

While nearest-neighbor (NN) decoding is optimal for memoryless AWGN channels, it is often computationally intractable or suboptimal for general IBM channels, particularly when the channel characteristics are uncertain or mismatched. Existing mismatched decoding frameworks, such as the Generalized Nearest Neighbor Decoding Rule (GNNDR), have been primarily confined to symbol-wise memoryless channels. A direct element-wise application of GNNDR to IBM channels fails to exploit the shared statistical structure (e.g., common phase rotations or correlated noise) within a block, leading to significant performance losses.

Methodology
This work extends the GNNDR framework to IBM channels by proposing a Vectorized Generalized Nearest Neighbor Decoding (Vec-GNND) architecture. The core methodology involves:

Block-Level Metric Formulation: Unlike element-wise decoding, Vec-GNND employs a block-level metric that utilizes vector-valued processing functions ( $g$ ) and matrix-valued scaling functions ( $f$ ). The decoding rule is defined as:
$\hat{m} = \arg \min_{m \in \mathcal{M}} \sum_{l=1}^{L} \| g(y_l, v_l) - f(y_l, v_l)x_l(m) \|_2^2$
where $y_l$ is the received block vector, $v_l$ is the channel state information (CSI), and $x_l(m)$ is the transmitted codeword block.
Generalized Mutual Information (GMI) Analysis: Leveraging GMI as an operational lower bound on mismatch capacity, the authors derive an analytical characterization of the achievable rate for Gaussian codebooks. The analysis utilizes the moment-generating function technique and the strong law of large numbers to establish the GMI expression for arbitrary processing and scaling functions.
Optimization of Processing and Scaling: The paper derives the optimal $g$ and $f$ functions that maximize the GMI. This involves:
- Computing the conditional mean ( $\mu$ ) and conditional covariance ( $\Sigma$ ) of the input given the output and CSI.
- Performing a spectral decomposition of the conditional covariance matrix.
- Introducing a truncation technique for eigenvalues that exceed the power constraint $P$ , ensuring the scaling function remains well-defined and information is not discarded via singular projections.
Joint Codebook-Metric Design: The authors formulate a joint design perspective where the input covariance matrix of the Gaussian codebook and the decoding metric are optimized together. For fixed covariances, the metric optimization yields a closed-form solution. For the specific case of diagonal covariance matrices, the paper derives first-order self-consistent optimality conditions, revealing a water-filling-like structure.

Key Contributions
The paper makes the following specific contributions:

Vectorized Architecture: It introduces Vec-GNND as a block-level mismatched decoding architecture that preserves intra-block statistical structures through matrix-valued scaling and vector-valued processing.
Optimal Metric Characterization: It provides a closed-form analytical expression for the optimal Vec-GNND and its corresponding GMI for general IBM channels with Gaussian codebooks. Notably, it relaxes previous assumptions by removing strict power constraints on conditional variance through a novel truncation method.
Restricted Variants: The work characterizes the achievable GMI for several restricted, lower-complexity configurations, including:
- Constant scalar and matrix scaling.
- CSI-dependent scalar and matrix scaling.
- Linear processing functions.
Joint Design Framework: It establishes a framework for joint codebook-metric design, reducing the problem to input-covariance optimization and providing first-order conditions for diagonal covariances.
Numerical Validation: The proposed scheme is evaluated on block noncoherent AWGN channels and phase noise channels.

Results
Numerical evaluations demonstrate consistent performance gains of the optimal Vec-GNND over conventional scaling-based baselines and element-wise GNNDR:

Block Noncoherent AWGN: In this scenario, all symbols in a block share a common unknown phase. The optimal Vec-GNND exploits this structure, achieving a GMI that grows unboundedly with SNR (at a rate of $\frac{B_x-1}{B_x} \text{SNR}$ ). In contrast, element-wise GNNDR, which ignores the shared phase, saturates at a finite constant ( $\approx 0.4287$ bits/channel use) regardless of SNR. The performance gap widens as the block length $B_x$ increases.
Phase Noise Channels: For channels with correlated phase noise within a block, the optimal Vec-GNND and its restricted variants (e.g., linear processing) significantly outperform the identity decoding rule (standard Euclidean distance). The performance gap increases with both block length and diffusion intensity, highlighting the robustness of Vec-GNND against severe codeword rotations.

Significance and Claims
The paper claims that the primary contribution is not merely the abstraction of IBM channels, but the identification of a GMI-optimal block-level metric that effectively exploits intra-block statistical structures unavailable to element-wise decoders. The proposed Vec-GNND offers a clear signal-processing interpretation: it performs decoding along the principal directions of the channel's conditional covariance, thereby preserving maximal information.

The work highlights that while the optimal metric requires evaluating conditional means and covariances (which may be computationally intensive for general channels), the framework provides a theoretical benchmark and a structured approach for designing practical receivers. The authors note that for specific cases like linear processing, the scheme connects to the Bussgang decomposition and worst-case uncorrelated noise theorems, offering a heuristic justification for its efficacy. The paper concludes that Vec-GNND holds substantial potential for realistic communication scenarios involving block memory, though future work is needed to extend these results to general memory channels, reduced-dimension processing, and discrete constellations.