⚡ electrical engineering

Score-Based Conditional Flow Models for MIMO Receiver Design with Superimposed Pilots

This paper proposes a conditional flow matching receiver (CFM-Rx) that leverages unsupervised, deterministic generative modeling to achieve robust joint channel estimation and data detection in MIMO systems with superimposed pilots, effectively overcoming pilot contamination and eliminating the need for labeled training data.

Original authors: Ruhao Zhang, Yupeng Li, Yitong Liu, Shijian Gao, Jing Jin, Hongwen Yang, Jiangzhou Wang

Published 2026-02-26

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Ruhao Zhang, Yupeng Li, Yitong Liu, Shijian Gao, Jing Jin, Hongwen Yang, Jiangzhou Wang

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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 listen to a friend speaking to you at a very loud, chaotic party. This is exactly what happens inside a modern cell phone when it tries to receive data from a cell tower.

In the world of 5G and beyond, phones use multiple antennas (MIMO) to talk to the tower. To understand the message, the phone needs to know exactly how the "air" (the channel) is distorting the sound. Usually, the tower sends a special "test tone" (called a pilot) so the phone can figure out the distortion.

The Problem: The "Overhead" Tax
Traditionally, the tower stops sending your data to send these test tones. It's like your friend pausing their story every few seconds to say, "Can you hear me?" This wastes time and space, slowing down your internet.

To fix this, engineers came up with Superimposed Pilots (SIP). Instead of pausing, the tower whispers the test tone while it's shouting your data. It's like your friend telling a story while simultaneously humming a tune. The problem? The story and the tune get mixed up. The phone hears a jumbled mess and struggles to separate the "test tone" from the "story," leading to errors and dropped calls.

The Old Solution: The "Supervised" Student
Previous AI solutions tried to solve this by acting like a student who memorized thousands of examples of "jumbled messes" and their correct "stories."

The Flaw: This student needs a massive library of labeled examples (which are hard to get). If the party gets louder, or the friend changes their accent, or the room layout changes, the student gets confused and has to relearn everything from scratch. It's rigid and expensive.

The New Solution: CFM-Rx (The "Intuitive Detective")
This paper introduces a new receiver called CFM-Rx. Instead of memorizing examples, it acts like a brilliant detective who understands the physics of the situation.

Here is how it works, using simple analogies:

1. The "Reverse Movie" Trick (Generative Modeling)

Imagine a video of a shattered glass vase reassembling itself.

Old AI (Diffusion Models): Tries to reassemble the vase by randomly guessing where every piece goes, checking, and trying again. It's slow and jittery. It might take 100 tries to get it right.
CFM-Rx (Flow Matching): Instead of guessing randomly, it calculates the exact path every piece needs to take to get back to the whole vase. It's like a smooth, deterministic movie playing in reverse. It knows exactly where to go, step-by-step, without the jitter. This makes it fast and stable.

2. The "Two-Step Dance" (Predictor-Corrector)

When the phone receives the jumbled signal (Story + Hum), CFM-Rx doesn't just guess. It does a two-step dance:

The Predictor (The Intuition): "Based on what I know about how radio waves usually behave, the story probably looks something like this." It uses a pre-trained "brain" (a neural network) that has learned the general rules of radio channels without needing specific examples of your data.
The Corrector (The Reality Check): "Wait, the actual sound I'm hearing is slightly different. Let me adjust my guess to match the real sound." It uses the math of the current signal to nudge the guess into the right place.

It repeats this dance a few times (about 30 steps), and suddenly, the "story" (your data) and the "hum" (the pilot) are perfectly separated.

3. Why It's a Game Changer

No Homework Needed: Unlike the old AI, CFM-Rx doesn't need a massive library of labeled "jumbled messages." It learns the structure of the radio channel itself. This means it works even if you change the antenna setup or the modulation type (the "accent" of the data).
Speed: Because it uses a smooth "movie" path instead of random guessing, it solves the problem in milliseconds. This is crucial for real-time 5G applications like self-driving cars or remote surgery.
Robustness: Even when the "party" is incredibly noisy (low signal), CFM-Rx can still separate the story from the hum better than traditional methods.

The Bottom Line
Think of CFM-Rx as a smart, intuitive listener who doesn't need to memorize every possible conversation. Instead, it understands the laws of physics and sound, allowing it to instantly untangle a messy signal, separate the pilot from the data, and deliver your message clearly—even when the connection is terrible. It's faster, smarter, and requires less "training" than the AI receivers we've used before.

1. Problem Statement

The paper addresses the critical challenge of Joint Channel Estimation and Data Detection (JCED) in Multiple-Input Multiple-Output (MIMO) systems utilizing Superimposed Pilots (SIP).

Context: SIP improves spectral efficiency by transmitting pilot and data symbols simultaneously on the same time-frequency resources, eliminating the overhead of orthogonal pilots.
Challenge: This superposition causes severe pilot contamination and data interference, making the separation of signals difficult.
Limitations of Existing Solutions:
- Conventional Model-Driven Methods (e.g., LMMSE): Struggle with the non-linear interference and require accurate channel covariance statistics, which are hard to obtain in dynamic environments.
- Supervised Deep Learning (DL): Requires massive labeled datasets (transmitted bits vs. received signals), which are costly to generate and store. These models often suffer from poor generalization to new channel conditions, pilot patterns, or modulation schemes and require frequent retraining.
- Diffusion Models (SDE-based): While effective for inverse problems, they rely on stochastic differential equations (SDEs) requiring hundreds of iterative sampling steps. This results in high latency and computational complexity, making them unsuitable for real-time 5G/6G receiver constraints (sub-millisecond processing times).

2. Methodology: Conditional Flow Matching Receiver (CFM-Rx)

The authors propose CFM-Rx, an unsupervised generative framework that learns directly from received signals without labeled data. It leverages Flow Matching (FM) and Optimal Transport (OT) to replace stochastic diffusion with deterministic inference.

Core Concepts:

Unsupervised Learning: Instead of learning input-output mappings (received signal $\to$ bits), the model learns the underlying wireless channel distribution as a prior. This eliminates the need for labeled training data.
Deterministic ODE vs. Stochastic SDE: Unlike diffusion models that use noisy SDEs, CFM-Rx uses a moment-consistent Ordinary Differential Equation (ODE). This ensures a smooth, constant-speed trajectory from a noise prior to the target data distribution, enabling single-pass, low-latency inference.
Conditional Flow Matching: The framework unifies flow matching with score-based modeling. It formulates a conditional velocity field that decomposes the inference process into two components:
1. Prior-Driven Drift: Guided by a pre-trained neural network (U-Net) that captures the statistical structure of the channel.
2. Likelihood-Based Correction: A gradient ascent step that steers the generated samples to be consistent with the actual received signal ( $Y$ ) and pilot symbols ( $P$ ).

Algorithmic Workflow:

The receiver employs an Iterative Predictor-Corrector Scheme:

Initialization: Start with random noise samples for the channel ( $H_1$ ) and data ( $D_1$ ).
Predictor Step (Unconditional): Use the pre-trained flow network $v_\theta(H_t, t)$ and analytical Tweedie's formula for data to evolve the state backward in time (from noise to signal).
Corrector Step (Conditional): Apply gradient ascent on the log-likelihood of the received signal ( $\nabla \log p(Y|H, D, P)$ ) to correct the estimates, ensuring physical consistency with the observation model.
Iteration: Repeat for $T$ steps (e.g., $T=30$ ) to converge to the posterior distribution $p(H, D | Y, P)$ .

3. Key Contributions

Novel Unsupervised Framework: Proposed CFM-Rx, which jointly estimates channels and detects data without labeled training data, significantly reducing dataset requirements and improving adaptability to varying system settings.
Moment-Consistent ODE Formulation: Developed a theoretical framework that unifies ODE-based flow matching with score-based modeling. This replaces the computationally expensive stochastic sampling of diffusion models with a deterministic, efficient process suitable for real-time hardware.
Conditional Velocity Decomposition: Derived a receiver-specific conditional velocity field that explicitly separates the generative prior from the measurement likelihood. This mechanism mitigates error propagation common in cascaded (estimate-then-equalize) designs.
Robustness and Generalization: Demonstrated that the learned channel prior generalizes across different modulation formats (QPSK, 8PSK, 16PSK) and unseen channel models (CDL-A, CDL-B) without retraining, a capability lacking in supervised end-to-end (E2E) models.

4. Simulation Results

Extensive simulations were conducted on a 4x4 MIMO system using 3GPP CDL channel models.

Channel Estimation Accuracy:
- CFM-Rx with SIP outperforms the state-of-the-art diffusion model (DM-JED) by 5–7 dB in Normalized Mean Square Error (NMSE) at high SNR.
- At low SNR (< 5 dB), the gain exceeds 10 dB, approaching the theoretical performance ceiling of the Oracle LMMSE estimator.
Bit Error Rate (BER) & Throughput:
- CFM-Rx achieves the lowest BER among all AI-based and conventional methods.
- Due to the elimination of pilot overhead, SIP-based CFM-Rx achieves a throughput of ~1.34 Mbit/s, compared to ~1.23 Mbit/s for Orthogonal Pilot (OP) schemes.
Convergence & Efficiency:
- Speed: CFM-Rx converges within 30 sampling steps, whereas DM-JED requires >100 steps for convergence.
- Complexity: CFM-Rx requires significantly fewer Floating Point Operations (FLOPs) ( $5.7 \times 10^7$ ) compared to E2E ( $7.6 \times 10^8$ ) and DM-JED ( $8.9 \times 10^7$ ).
- Data Efficiency: Requires only 8,000 training samples (channel realizations) compared to 100,000 labeled pairs for E2E.
Generalization: The model maintains high performance when tested on unseen channel models (CDL-A/B) and modulation orders, whereas supervised E2E models suffer severe performance degradation.

5. Significance and Impact

Real-Time Viability: By replacing stochastic SDE sampling with deterministic ODE flow matching, the paper bridges the gap between high-performance generative AI and the strict latency constraints of 5G/6G physical layer receivers.
Practical Deployment: The unsupervised nature of CFM-Rx removes the bottleneck of acquiring massive labeled datasets, making it highly scalable for diverse and dynamic wireless environments.
Future Directions: The framework lays the groundwork for advanced applications in Non-Orthogonal Multiple Access (NOMA) and Integrated Sensing and Communication (ISAC), where separating overlapping signals is paramount. It also suggests future extensions to the delay-Doppler domain (OTFS) for high-mobility scenarios.

In summary, CFM-Rx represents a paradigm shift in receiver design, moving from supervised, high-latency, data-hungry models to unsupervised, deterministic, and highly generalizable generative frameworks capable of handling the complex interference inherent in superimposed pilot systems.