⚡ electrical engineering

TenSIM: Tensor-Based Channel Estimation for MIMO Systems with Stacked Intelligent Metasurfaces

This paper proposes TenSIM, a tensor-based channel estimation framework for stacked intelligent metasurface (SIM) assisted MIMO systems that leverages distinct PARAFAC and Tucker models for odd and even layer configurations to decouple channels, ensure identifiability, and outperform unstructured baselines in accuracy and scalability.

Original authors: André L. F. de Almeida, George C. Alexandropoulos

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: André L. F. de Almeida, George C. Alexandropoulos

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 Picture: Building a Better "Smart Mirror" for Wi-Fi

Imagine you are trying to send a message from a radio tower (the Transmitter) to a house (the Receiver). In a normal city, buildings block the signal, and the message gets lost or scrambled.

To fix this, engineers are building Stacked Intelligent Metasurfaces (SIMs). Think of a SIM not as a single mirror, but as a multi-layered stack of smart glass sheets.

Single-layer mirrors (old tech): Can only reflect a signal once. It's like holding up a single piece of foil to bounce a flashlight beam.
Stacked mirrors (SIMs): These are like a stack of 5 or 10 smart glass sheets. As the light passes through each layer, the glass can twist, turn, and focus the beam. This allows for much more powerful control over the signal, like a high-tech lens that can bend light in complex ways.

The Problem: To make these smart mirrors work, the computer needs to know exactly how the signal is traveling through every single layer of glass. This is called Channel Estimation.

The Challenge: Because there are so many layers interacting with each other, figuring out the path is like trying to solve a giant, tangled knot. If you try to untangle it by pulling on every string at once (the old way), it takes forever and often fails.

The Solution: TenSIM (The "Tensor" Detective)

The authors of this paper propose a new method called TenSIM. Instead of trying to untangle the whole knot at once, they realized the knot has a specific shape depending on whether the stack has an odd or even number of layers.

They use a mathematical tool called Tensor Decomposition (think of it as a special kind of 3D puzzle solver) to break the problem down into manageable pieces.

1. The Odd-Layer Stack (The "PARAFAC" Puzzle)

If your stack has an odd number of layers (e.g., 3, 5, 7), there is a perfect middle layer.

The Strategy: You only need to change the settings on that one middle layer while keeping the others fixed.
The Analogy: Imagine a sandwich with 5 slices of bread. You only wiggle the middle slice of bread. Because the layers on the left and right are fixed, the signal passing through them acts like a single, solid block.
The Result: This creates a clean, simple mathematical structure (called PARAFAC). It's like solving a puzzle where the pieces fit together in a straight line.
- Pros: It's fast, cheap to compute, and very stable even if the layers are far apart.
- Cons: It's slightly less precise than the other method if the layers are very close together.

2. The Even-Layer Stack (The "Tucker" Puzzle)

If your stack has an even number of layers (e.g., 2, 4, 6), there is no single middle layer. There are two layers right in the center.

The Strategy: You have to wiggle both of those middle layers at the same time.
The Analogy: Imagine a sandwich with 4 slices of bread. You have to wiggle the two middle slices together. The space between those two slices becomes a complex interaction zone.
The Result: This creates a richer, more complex mathematical structure (called Tucker). It's like solving a puzzle where the pieces are interlocked in a 3D grid.
- Pros: If you have enough data and the layers are close together, this method is the most accurate. It can reconstruct the signal path with high precision.
- Cons: It is computationally heavy (takes more brainpower) and gets unstable if the layers are spaced too far apart.

How They Tested It (The "Training" Phase)

To teach the system how to work, they send special test signals (pilots) through the stack.

The Innovation: Instead of changing every layer of the stack for every test signal (which would be slow and require too much power), they only change the middle layer(s).
The Benefit: This saves a massive amount of time and energy while still giving the computer enough information to solve the puzzle.

What the Results Show

The paper ran thousands of computer simulations to compare their new method against old methods. Here is what they found:

Better Accuracy: Both TenSIM methods (Odd and Even) were much better at guessing the signal path than the old "brute force" methods.
The Trade-off:
- TenSIM-PARAFAC (Odd layers) is the reliable workhorse. It's fast, doesn't crash easily, and works well even if the layers are spaced out. It's great for large systems.
- TenSIM-Tucker (Even layers) is the precision surgeon. If you have a small stack with layers very close together and plenty of data, it gives the most accurate picture. But if the layers are far apart, it gets confused.
Robustness: Even if the computer doesn't know the exact settings of the smart glass (due to hardware errors), TenSIM can still figure it out, provided they send enough test signals.

The Bottom Line

The paper introduces TenSIM, a smart way to "listen" to signals passing through multi-layered smart mirrors. By realizing that odd and even stacks behave differently, they created two specialized tools:

Use the Odd-Layer tool if you want speed, stability, and simplicity.
Use the Even-Layer tool if you need maximum precision and have a tightly packed stack.

This approach allows 6G networks to use these powerful new mirrors without getting bogged down by complex math or slow processing times.

Technical Summary: TenSIM: Tensor-Based Channel Estimation for MIMO Systems with Stacked Intelligent Metasurfaces

Problem Statement
Stacked Intelligent Metasurfaces (SIMs) represent a promising architecture for 6G and beyond, offering richer electromagnetic wave manipulation than conventional single-layer metasurfaces through cascaded inter-layer coupling. However, realizing these gains requires accurate and scalable channel estimation. The end-to-end response of a SIM-assisted Multiple-Input Multiple-Output (MIMO) system involves a highly coupled multi-factor inverse problem, comprising transmitter-to-SIM channels, SIM-to-receiver channels, inter-layer coupling matrices, and layer-dependent phase profiles.

Existing channel estimation methods face significant challenges:

High Dimensionality and Ambiguity: Direct estimation with naive pilot designs leads to high-dimensional bilinear/multilinear ambiguities and poor identifiability.
Training Overhead: Treating the composite channel as an unstructured matrix requires a training overhead proportional to the square of the SIM aperture ( $N^2$ ), which is impractical for large-scale deployments.
Structural Ignorance: Current approaches often fail to exploit the specific algebraic structure induced by the SIM cascade, particularly the distinction between odd and even numbers of layers.
Scalability and Robustness: Learning-based methods suffer from high data/computation demands, while existing structured estimators often rely on restrictive assumptions (e.g., strong channel sparsity) or fail to address the specific identifiability conditions of SIM cascades.

Methodology
The paper proposes TenSIM, a unified tensor-based channel estimation framework that leverages the structural properties of SIM cascades. The core methodology involves:

Parity-Dependent Training Protocol: The authors introduce a reduced-complexity training strategy where only the most informative central layers are reconfigured across pilot blocks, while outer layers remain fixed.
- Odd-Layer SIM ( $L = 2M + 1$ ): Only the single central layer ( $M+1$ ) is varied.
- Even-Layer SIM ( $L = 2M$ ): The two central layers ( $M$ and $M+1$ ) are jointly varied.
Tensor Modeling: Based on the training protocol, the received pilot signals are modeled as third-order tensors. The framework derives two distinct tensor decompositions based on the parity of the layer count:
- Odd-Layer Case: The cascade induces a PARAFAC (Parallel Factor) model. The received tensor is decomposed into effective transmitter-SIM and SIM-receiver channel factors and a programmable phase configuration matrix, with an identity core tensor.
- Even-Layer Case: The cascade induces a Tucker model. The received tensor is decomposed into the same channel factors and training matrix, but coupled through a known core tensor determined by the central inter-layer propagation matrix ( $W_{M+1}$ ).
Estimation Algorithms: The authors develop Alternating Least Squares (ALS) estimators tailored to each model:
- TenSIM-PARAFAC: Iteratively updates the effective channel factors using Khatri-Rao product structures.
- TenSIM-Tucker: Iteratively updates factors using Kronecker product structures involving the known core tensor unfoldings.
- Blind/Imperfect Training: The framework is extended to jointly estimate the training matrix when it is unknown or imperfect at the receiver, cycling through channel and training factor updates.
Identifiability Analysis: The paper establishes rank-based conditions for unique identifiability. It demonstrates that under mild diversity conditions (sufficient pilot blocks $K$ and full-rank training designs), the channel factors are identifiable up to trivial scaling ambiguities (diagonal scaling for PARAFAC, global scalar for Tucker).

Key Contributions

Unified Framework: Introduction of TenSIM, which unifies channel estimation for SIM-assisted MIMO systems by distinguishing between odd and even layer architectures.
Structural Exploitation: Derivation of parity-dependent tensor models (PARAFAC for odd, Tucker for even) that decouple the Tx-SIM and SIM-Rx channel factors while explicitly accounting for inter-layer wave coupling.
Algorithmic Development: Design of practical ALS estimators for both models, along with constructive sufficient conditions for full-column-rank training designs.
Complexity and Trade-off Analysis: Quantification of computational complexity, highlighting that TenSIM-PARAFAC offers lower complexity ( $O(N^3)$ ) and better scalability, while TenSIM-Tucker ( $O(N^4)$ ) provides richer modeling at the cost of higher complexity and sensitivity to inter-layer spacing.
Robustness Evaluation: Demonstration that the framework remains effective under imperfect or blind SIM training conditions when additional pilot diversity is available.

Results
Extensive numerical experiments validate the proposed framework:

Performance vs. Baselines: Both TenSIM-PARAFAC and TenSIM-Tucker significantly outperform unstructured Least Squares (LS) baselines by exploiting the tensor structure. They achieve lower Normalized Mean Squared Error (NMSE) with significantly fewer training blocks ( $K \ll N^2$ ).
SNR and Diversity: Estimation accuracy improves monotonically with Signal-to-Noise Ratio (SNR) and training diversity ( $K$ ).
Architecture Comparison:
- TenSIM-PARAFAC (Odd): Demonstrates superior scalability, lower computational complexity, and stronger robustness to variations in inter-layer spacing. It performs well even when coupling is weak.
- TenSIM-Tucker (Even): Provides more accurate channel reconstruction when sufficient training and strong layer coupling are available. However, its performance degrades as inter-layer spacing increases due to the ill-conditioning of the core propagation matrix.
Convergence: TenSIM-Tucker exhibits SNR-independent convergence behavior, whereas TenSIM-PARAFAC converges faster at higher SNRs.
Imperfect Training: The framework successfully handles scenarios where the SIM training matrix is imperfectly known, provided sufficient pilot diversity is used to compensate for the uncertainty.

Significance and Claims
The paper claims that TenSIM provides a unified and physically interpretable approach for channel estimation in SIM-assisted MIMO systems. Its primary significance lies in:

Explicit Identifiability: It provides rigorous conditions under which the cascaded channel factors can be uniquely recovered, addressing a fundamental gap in current SIM literature.
Physical Interpretability: The framework maps the physical structure of the SIM (odd vs. even layers) directly to specific algebraic tensor models, offering a clear link between electromagnetic design and signal processing.
Practical Trade-offs: It explicitly characterizes the trade-offs between estimation accuracy, computational complexity, and robustness to physical parameters (like inter-layer spacing), guiding system designers in choosing between odd and even layer configurations based on their specific deployment constraints.
Scalability: By reducing the training overhead from $O(N^2)$ to a manageable level through structured tensor decomposition, TenSIM makes large-scale SIM deployment more feasible.

The authors conclude that while the even-layer (Tucker) approach offers a richer model, the odd-layer (PARAFAC) approach often offers a more favorable balance of estimation reliability and implementation cost, particularly for large SIM apertures or widely spaced layers. The work bridges the gap between electromagnetic layer programmability and identifiable communication-layer models.