⚡ electrical engineering

Discrete Diffusion for Codebook-Based Beam Candidate Generation

This paper proposes a history-conditioned discrete denoising diffusion probabilistic model to generate high-quality beam candidates for limited-probing mmWave systems, demonstrating superior performance in signal-to-noise ratio, beam-miss probability, and probe regret compared to existing baselines, particularly under tight probing budgets.

Original authors: Amirhossein Azarbahram, Onel L. A. López

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amirhossein Azarbahram, Onel L. A. López

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

The Big Problem: Finding a Needle in a Moving Haystack

Imagine you are trying to talk to a friend using a super-powerful flashlight in a pitch-black forest. Your flashlight has a very narrow beam (like a laser pointer). To have a good conversation, you need to point the light exactly at your friend's eyes.

However, there are two big problems:

The Haystack: There are hundreds of possible directions you could point the light (the "codebook").
The Moving Target: Your friend is running around, and trees (buildings) might suddenly block the light.

In the world of 5G and 6G (millimeter-wave internet), this is exactly what happens. The internet is super fast, but the signal is so fragile that if you miss the target by a tiny bit, or if someone walks in front of the signal, the connection drops.

The Challenge: The phone (User Equipment) and the tower (Base Station) can only check a few directions at a time because checking all of them takes too long and uses up too much battery. They have to guess which directions are best based on what happened in the past.

The Old Way vs. The New Way

The Old Way (Discriminative Models):
Imagine a student taking a multiple-choice test. The old AI methods look at the past data and say, "I am 99% sure the answer is Option A." They pick one single best guess.

The Flaw: If the student is wrong, they are completely lost. Also, in a forest, sometimes Option A is good, but Option B is almost as good. If the student only looks at Option A, they miss the backup plan.

The New Way (Discrete Diffusion):
This paper introduces a new method called D3PM-BM. Instead of guessing one answer, this AI acts like a creative brainstorming session.

Think of it like a detective trying to find a suspect.

The Old Detective: "The suspect is definitely wearing a red hat." (High confidence, but if they are wrong, the case is closed).
The New Detective (Diffusion): "The suspect is likely wearing a red hat, but maybe a blue one, or maybe a green one. Let's generate a list of the top 10 most likely outfits and check all of them."

How the "Diffusion" Magic Works

The term "Diffusion" sounds complicated, but here is the simple analogy: The "Blur and Sharpen" Game.

The Blur (Forward Process): Imagine taking a clear photo of the best flashlight direction and slowly adding static noise to it until it looks like a blurry mess of random pixels.
The Sharpen (Reverse Process): Now, imagine teaching a computer to look at that blurry mess and guess what the original clear photo was.
- The computer starts with pure randomness (static).
- It takes a step back, removing a little bit of noise.
- It takes another step, removing more noise.
- It repeats this until it "hallucinates" a clear image of the best direction.

Because this process is random, if you run it 10 times, you get 10 slightly different clear images.

Run 1: "Maybe the light should go Left."
Run 2: "Maybe the light should go Left-Slightly-Up."
Run 3: "Maybe the light should go Right."

Why is this better? Instead of betting everything on one direction, the system generates a shortlist of candidates. It says, "We don't know for sure, but these 5 directions are the most promising. Let's check all 5 quickly."

The "History" Ingredient

The paper also emphasizes that this AI is history-conditioned.
Think of it like a GPS that learns your driving habits.

If you always turn left at 5th Street, the GPS doesn't just look at the map; it looks at your past turns.
Similarly, this AI looks at the history of signal feedback. "Last time we checked the left, the signal was weak. Two seconds ago, the right was strong. The user is moving north."

It uses a special "Transformer" (a type of AI brain) to read this history like a story, understanding the flow of time and movement, rather than just looking at a single snapshot.

The Results: Why Should We Care?

The researchers tested this new method against the old ones in a simulated city with moving cars and people.

Better Connection: The new method kept the internet connection stronger (higher Signal-to-Noise Ratio).
Fewer Dropped Calls: It was much better at avoiding "beam misses" (pointing the light at a tree instead of a person).
The "Low Budget" Superpower: The biggest win happened when the system was allowed to check very few directions (a tight budget). In these tight situations, the old methods panicked and picked the wrong direction. The new "brainstorming" method picked a diverse list of good options, ensuring that even if they couldn't check everything, they checked the right things.

Summary Analogy

Imagine you are a chef trying to find the perfect spice blend for a soup, but you can only taste 3 spoons at a time.

Old AI: Tastes the soup, thinks "It needs salt," and adds salt. If it was wrong, the soup is ruined.
New AI (Diffusion): Tastes the soup, thinks "It might need salt, or maybe pepper, or maybe a pinch of cumin." It generates a list of 5 likely combinations. It then tastes those 5 combinations. Even if it can only taste 3, it has a much higher chance of finding the perfect flavor because it explored more possibilities.

In short: This paper teaches the internet how to be less rigid and more creative, generating a diverse list of "best guesses" to keep our connections fast and stable, even when we are moving fast or the signal is blocked.

1. Problem Statement

The paper addresses the challenge of beam management in millimeter-wave (mmWave) communication systems under limited probing constraints.

Context: mmWave systems rely on highly directional beams to overcome path loss. However, these links are sensitive to blockage and user mobility, requiring frequent beam adaptation.
The Constraint: In practical systems, a Base Station (BS) cannot sweep the entire beam codebook at every time slot due to excessive overhead. Instead, it can only probe a small subset of beams ( $P \ll K$ , where $K$ is the codebook size).
The Decision Problem: The BS must select which $P$ beams to probe based on partial, noisy, and quantized feedback from previous slots. The goal is to maximize the long-term average Signal-to-Noise Ratio (SNR) by ensuring the "oracle" (optimal) beam is included in the probed set, or at least that the best probed beam is close to the oracle.
Challenges:
- Partial Observability: The true channel state is hidden; only quantized feedback from a subset of beams is available.
- Combinatorial Action Space: Selecting a subset of $P$ beams from $K$ options creates a massive search space, making optimal dynamic programming intractable.
- Temporal Dynamics: Beam quality evolves over time due to mobility, requiring the model to exploit historical probing sequences.
- Limitations of Existing Methods: Traditional discriminative models (e.g., standard classifiers) often output a single "best" prediction, lacking the diversity needed to cover multiple plausible beam directions in dynamic environments.

2. Methodology: D3PM-BM

The authors propose D3PM-BM (History-Conditioned Discrete Denoising Diffusion Probabilistic Model for Beam Management). This is a generative framework designed to produce a diverse set of high-quality beam candidates.

A. System Formulation

Input: A history of probing actions and feedback ( $H_t$ ) over $L$ previous slots.
Output: An ordered list of $S$ candidate beam indices.
Objective: Learn a conditional distribution $p(\text{beam} | H_t)$ that captures the probability of promising beams, rather than just predicting a single deterministic label.

B. Architecture Components

Hierarchical History Encoder:
- Token Formation: Converts beam indices (embeddings) and scalar feedback (normalized via MLP) into tokens.
- Within-Slot Aggregation: Uses attention-style pooling to aggregate $P$ probe measurements into a single slot representation.
- Temporal Modeling: A Transformer encoder processes the sequence of slot representations to capture long-term temporal dependencies, outputting a context vector $c_t$ .
Discrete Diffusion Model (D3PM):
- Unlike standard DDPMs that operate on continuous Gaussian noise, this model uses Discrete Diffusion (categorical diffusion) suitable for beam indices.
- Forward Process: Gradually corrupts the "oracle" beam index (clean data) by mixing it with a uniform distribution over the codebook.
- Reverse Process: A neural network (denoiser) learns to reconstruct the clean beam index from the corrupted state, conditioned on the history context $c_t$ .
Soft Oracle Label Training:
- Instead of training on a single "hard" label (the top-1 beam), the model uses sparse soft labels.
- The target distribution assigns probability mass to the top- $M$ strongest beams, weighted by their relative SNR using a temperature parameter.
- Benefit: This preserves information about near-optimal alternatives, stabilizing training and encouraging the model to learn a diverse set of high-quality candidates rather than collapsing to a single mode.
Inference and Candidate Ranking:
- Sampling: The model generates $S_{gen}$ samples by running the reverse diffusion process from random noise.
- Ranking: Candidates are ranked based on a composite score combining:
  - Frequency: How often a beam appears in the samples.
  - Confidence: The log-probability of the sample at the final denoising step.
- The top $S$ distinct beams form the proposal list, from which the BS selects $P$ beams to probe.

3. Key Contributions

Problem Formulation: Defined predictive beam selection as a history-dependent, partially observable sequential decision problem under strict probing budgets.
Novel Framework (D3PM-BM): Developed the first application of Discrete Denoising Diffusion Probabilistic Models specifically for codebook-based beam candidate generation. It learns a conditional distribution over discrete beam indices directly from partial probing histories.
Training Innovation: Introduced a temperature-scaled soft oracle label mechanism. This allows the model to supervise on multiple near-optimal beams simultaneously, improving robustness against label noise and enhancing candidate diversity.
Offline-Online Workflow: Proposed a practical workflow where the model is trained offline on logged interaction traces and deployed online as a candidate generator, decoupling data collection from model learning.
Complexity Analysis: Demonstrated that short diffusion chains can recover most performance benefits, offering a favorable accuracy-complexity tradeoff for real-time deployment.

4. Numerical Results

The authors evaluated D3PM-BM against strong baselines (EMA, UCB, Transformer-based discriminative predictor, and ODE-LSTM) using the DeepMIMO dataset.

Performance Metrics: Average SNR, Oracle Miss Probability (probability the best beam is not probed), and Conditional Probe Regret (SNR loss when the oracle is missed).
Key Findings:
- Superior SNR: D3PM-BM consistently achieved the highest average SNR, particularly in low-probing regimes (small $P$ ), where candidate quality is most critical.
- Reduced Miss Probability: It significantly lowered the probability of missing the oracle beam compared to discriminative models.
- Lower Regret: Even when the oracle beam was missed, D3PM-BM selected beams with SNRs much closer to the oracle, minimizing performance loss.
- Diversity vs. Accuracy: While discriminative models (like TRM) had slightly better Top-1 accuracy, D3PM-BM achieved significantly higher Top-m inclusion rates (for $m > 1$ ). This confirms that the generative approach provides a broader coverage of plausible high-SNR beams.
- Robustness: The method remained robust under noisy feedback and coarse quantization levels.
- Efficiency: Performance saturated quickly with short diffusion chains, suggesting that inference latency can be minimized without significant performance degradation.

5. Significance

This paper represents a significant shift in mmWave beam management from discriminative prediction (guessing the single best beam) to generative candidate generation (producing a diverse set of high-probability beams).

Theoretical Impact: It demonstrates the viability of Discrete Diffusion Models for combinatorial decision problems in wireless communications, moving beyond continuous domains.
Practical Impact: By improving the quality of the "candidate set" under tight overhead constraints, the method enables more reliable mmWave links in high-mobility scenarios without increasing the probing overhead.
Future Direction: It opens the door for using advanced Generative AI (GenAI) to handle uncertainty and stochasticity in wireless channel dynamics, moving beyond point estimates to probabilistic decision support.