⚡ electrical engineering

PARAFAC-Based Channel Estimation for Beyond Diagonal Reconfigurable Surfaces

This paper proposes a pilot-assisted tensor framework and a PARAFAC-based alternating least-squares (PALS) receiver for group-connected Beyond Diagonal Reconfigurable Intelligent Surfaces (BD-RIS) that maintains a fixed interconnection topology to enable practical, low-complexity channel estimation with superior accuracy compared to conventional methods.

Original authors: Gilderlan Tavares de Araújo, Bruno Sokal, André L. F. de Almeida

Published 2026-03-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gilderlan Tavares de Araújo, Bruno Sokal, André L. F. de Almeida

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 have a conversation with a friend in a huge, noisy stadium. You can't shout directly because the crowd is too loud, so you use a team of mirrors (called a Reconfigurable Intelligent Surface, or RIS) to bounce your voice around the obstacles and reach your friend.

In the old days, these mirrors were simple: they could only tilt slightly to reflect light in one direction, like a standard bathroom mirror. But the new technology, called BD-RIS (Beyond Diagonal RIS), is like a team of mirrors that can also talk to each other. They can pass signals between themselves, creating a much richer, more powerful path for your voice. This is amazing for speed and clarity, but it creates a huge problem: How do you figure out exactly how the sound is bouncing?

This is the "Channel Estimation" problem. To make the mirrors work, you first need to "train" them by sending test signals (pilots) and measuring the result.

The Problem: The "Reconfiguring Nightmare"

Previous methods for training these advanced mirror teams were like trying to solve a puzzle where you have to completely rebuild the entire mirror wall every single time you send a test signal.

Imagine you have 100 mirrors. To train them, you have to physically rewire how they connect to each other for the first test, then rewire them again for the second test, and again for the third.
This is slow, expensive, and practically impossible to build in real life. It's like trying to tune a radio by replacing the whole radio every time you change the station.

The Solution: The "Fixed Skeleton, Moving Skin"

The authors of this paper propose a clever new way to train these mirrors using a mathematical concept called PARAFAC (which sounds scary, but think of it as a "Lego Blueprint").

Instead of rebuilding the whole wall every time, they propose a two-timescale approach:

The Skeleton (Fixed): The way the mirrors are connected to each other (the "topology") stays exactly the same throughout the entire training session. Think of this as the metal frame of a puppet. It never changes.
The Skin (Moving): Only the "phase shifts" (the angle or timing of the reflection) change for each test. Think of this as the puppet's skin or clothes changing colors, while the bones underneath stay put.

By keeping the "skeleton" fixed, you don't need to do expensive rewiring. You just tweak the settings on the existing hardware.

The Magic Trick: The "Unfolding Puzzle"

How do they figure out the channel if they aren't changing the connections? They use a mathematical trick called Tensor Decomposition.

Imagine you have a 3D block of data (Time × Space × Signal).

Old methods tried to look at this block as a giant, messy jumble.
The new method realizes the data has a hidden, neat structure (like a stack of identical Lego bricks).
They use an algorithm called PALS (PARAFAC Alternating Least Squares). You can think of PALS as a smart robot that looks at the messy data and says, "Ah! I see a pattern. If I separate the 'Time' part from the 'Space' part, I can figure out exactly how the signal traveled from the transmitter to the mirrors, and then from the mirrors to the receiver."

Why This Matters (The Results)

The paper proves three main things:

It's Accurate: Even though they aren't rewiring the mirrors constantly, the new method is just as good at finding the signal path as the complex, expensive methods. It matches the "state-of-the-art" performance.
It's Cheap: Because the hardware connections stay fixed, the system is much easier to build and control. It reduces the "brain power" needed to run the mirrors.
It's Practical: It solves the biggest hurdle stopping these advanced mirrors from being used in real 6G networks.

The Analogy Summary

The Old Way: Trying to tune a piano by replacing the entire frame and strings for every single note you play. (Too hard, too slow).
The New Way (This Paper): Keeping the piano frame solid and permanent, but just adjusting the hammers (the phase shifts) to hit the right notes. You get the same beautiful music, but you don't have to rebuild the piano every time.

In a nutshell: This paper gives us a blueprint to use super-advanced, interconnected smart mirrors for future wireless networks without needing impossible hardware. It's a "fix the software, keep the hardware simple" approach that makes high-speed 6G much closer to reality.

1. Problem Statement

In Beyond Diagonal Reconfigurable Intelligent Surface (BD-RIS) assisted MIMO systems, channel estimation is a critical bottleneck. Unlike traditional diagonal RISs, BD-RISs feature rich inter-element coupling (group-connected or fully connected architectures) that enables significant performance gains but introduces two major challenges:

High Dimensionality: The cascaded channel structure involves a massive number of parameters, making pilot design and parameter recovery difficult.
Hardware Complexity: Existing tensor-based estimation methods often require reconfiguring the entire scattering matrix (inter-element connections) at every training block. This frequent hardware reconfiguration is costly, difficult to implement, and creates a substantial control overhead.

The core problem is to achieve accurate estimation of individual channel components (Tx-RIS and RIS-Rx) with low training overhead while maintaining hardware feasibility (i.e., avoiding full reconfiguration of the BD-RIS topology).

2. Methodology

The authors propose a pilot-assisted tensor framework operating under a two-timescale protocol. The methodology relies on three main pillars:

A. Structured Low-Rank PARAFAC Modeling

Instead of designing the BD-RIS scattering tensor independently for each block, the authors model the training structure using a PARAFAC (Parallel Factor Analysis) decomposition.

Fixed Topology: The interconnection topology (spatial factors) is kept fixed across all training blocks.
Dynamic Factors: Only low-dimensional phase coefficients (time-dependent factors) are updated per block.
Mathematical Formulation: The received signal tensor $\mathbf{Y}$ is modeled as a block-diagonal core tensor decomposed into factor matrices. Specifically, the scattering tensor for each group $q$ is modeled as:
$\mathbf{S}^{(q)} = \mathbf{I}_{3, \bar{R}} \times_1 \bar{\mathbf{P}}_1 \times_2 \bar{\mathbf{P}}_2 \times_3 \mathbf{P}^{(q)}_3$
Here, $\bar{\mathbf{P}}_1$ and $\bar{\mathbf{P}}_2$ are fixed spatial factor matrices (e.g., truncated DFT), while $\mathbf{P}^{(q)}_3$ contains the block-specific phase shifts.

B. PARAFAC Alternating Least Squares (PALS) Receiver

Based on the structured model, the authors derive a PALS receiver to decouple and estimate the individual channel matrices ( $\mathbf{G}$ and $\mathbf{H}$ ).

The algorithm alternates between updating estimates of $\mathbf{G}$ and $\mathbf{H}$ using Least Squares (LS) on the mode-unfoldings of the received tensor.
The updates utilize the Khatri-Rao product structure inherent in the PARAFAC model to separate the channel factors.
The process iterates until convergence, recovering the individual channels up to a scaling ambiguity (which does not affect the combined channel optimization).

C. Identifiability and Complexity Analysis

The paper derives necessary and sufficient conditions for channel identifiability:

Training Overhead: The minimum number of blocks $K$ required is $K \geq \bar{N}Q / M_R$ . This is significantly lower than existing methods (like BTALS or BTKF) which often require $K \geq \bar{N}^2 Q$ .
Design Complexity: The number of parameters to be adjusted is reduced from $\bar{N}^2 K$ (full tensor reconfiguration) to $(2\bar{N} + K)\bar{R}$ (factor matrix updates), where $\bar{R}$ is the PARAFAC rank.

3. Key Contributions

Hardware-Friendly Training Design: The paper introduces a novel training strategy where the BD-RIS interconnection topology remains constant across blocks. Only phase shifts are updated, drastically reducing hardware control complexity and implementation cost compared to state-of-the-art methods.
PALS Receiver: Development of a low-complexity, iterative receiver that exploits the PARAFAC structure to recover individual channel factors ( $\mathbf{G}$ and $\mathbf{H}$ ) rather than just the cascaded channel.
Theoretical Guarantees: The authors provide rigorous identifiability conditions and derive the minimum number of training blocks required for unique channel recovery, a result previously missing in the literature for this specific setup.
Complexity Reduction: Demonstrates a mathematical reduction in the number of control variables, making the system scalable as the number of RIS elements increases.

4. Numerical Results

Simulations were conducted comparing the proposed PALS method against:

LS (Least Squares): Conventional estimation.
BTALS/BTKF: State-of-the-art tensor-based methods that require full reconfiguration.

Key Findings:

Accuracy: PALS achieves Normalized Mean Square Error (NMSE) performance comparable to BTALS/BTKF and significantly outperforms conventional LS. It successfully recovers both individual channel factors with high fidelity.
Scalability: Performance improves as the number of groups increases (due to rank/diversity gain).
Complexity Trade-off: While maintaining state-of-the-art accuracy, PALS reduces the BD-RIS design complexity by orders of magnitude. Figure 1 in the paper illustrates that as the number of blocks ( $K$ ) increases, the complexity gap between PALS and competing methods widens significantly in favor of PALS.

5. Significance

This work bridges the gap between theoretical performance and practical implementation in BD-RIS systems.

Practicality: By eliminating the need for frequent, full-topology reconfiguration, the proposed method makes BD-RIS deployment feasible for real-world hardware constraints.
Efficiency: It achieves high estimation accuracy with reduced training overhead and control complexity.
Future Impact: The framework provides a scalable solution for next-generation wireless networks (6G) utilizing intelligent surfaces, offering a clear path toward implementing high-performance BD-RIS without prohibitive hardware costs.

In summary, the paper proposes a PARAFAC-based framework that transforms BD-RIS channel estimation from a hardware-intensive problem into a manageable, low-complexity task without sacrificing estimation accuracy.