Traditional wireless communication systems model the propagation channel between a transmitter and a receiver as a two-dimensional matrix. While this abstraction works well for basic multi-antenna setups, next-generation wireless architectures—such as Massive MIMO, millimeter-wave (mmWave), Terahertz (THz) bands, and Reconfigurable Intelligent Surfaces (RIS)—introduce an explosion of physical dimensions. Signals are now distributed across massive antenna arrays, hundreds of subcarriers, multiple time slots, and distinct polarization states.
When we flatten this multi-dimensional physical reality into a standard two-dimensional matrix, we destroy the structural dependencies connecting these domains. As a result, matrix-based estimation methods require massive pilot overhead or complex compressed sensing algorithms that are prone to local minima.
By contrast, treating the wireless environment as a high-order tensor preserves its geometric and physical properties. This structural alignment allows researchers to uniquely untangle overlapping signals, eliminate pilot overhead, and achieve closed-form algebraic solutions for channel estimation.
Preserving the Physical Geometry
The core insight of this research direction is that wireless signals naturally span multiple domains simultaneously. Instead of flattening the received data, we can organize it directly as a high-order tensor.
For instance, Panqi Chen et al. model the wireless channel as a fourth-order tensor capturing time, spatial location, frequency, and polarization. In multi-carrier modulation systems, Christos Chatzichristos structures received signals across antennas, time, and frequency as a three-way tensor. Even in generative modeling, Ushnish Sengupta et al. demonstrate that representing wireless channel responses as multi-channel tensors allows diffusion models to accurately synthesize complex spatial and spectral dependencies.
Why does this multi-dimensional representation make channel estimation easier? In high-frequency bands like mmWave and THz, signals do not scatter omnidirectionally; instead, they propagate along a small number of discrete physical paths. In tensor algebra, a single propagation path maps perfectly to a rank-one tensor (an outer product of vectors representing the spatial steering vector, frequency response, and time delay).
Consequently, a multi-path wireless channel can be represented as a low-rank tensor, where each rank-one component corresponds to an individual physical path. This mathematical alignment allows us to use tensor decomposition to isolate individual paths from the aggregate received signal.
Matrix Flattening (Loses Structure):
[ Space × Time × Frequency ] ---> [ Large 2D Matrix ] ---> Requires heavy pilot overhead
Tensor Modeling (Preserves Structure):
[ Space × Time × Frequency ] ---> [ 3D Low-Rank Tensor ] ---> Unique path decomposition
Uniqueness and Parameter Separation
In matrix algebra, a low-rank matrix decomposition ($H = AB^T$) is never unique without strict external constraints (such as orthogonality in SVD). We can always insert an arbitrary invertible matrix $T$ ($A T T^{-1} B^T = AB^T$) without changing the observed matrix. This creates an inherent ambiguity that complicates the extraction of actual physical parameters like Angles of Arrival (AoA) or delays.
Higher-order tensors do not suffer from this limitation. Under very mild conditions, Canonical Polyadic Decomposition (CPD) and Parallel Factor Analysis (PARAFAC) are uniquely identifiable up to simple scaling and permutation. This means that if we can represent the received signal as a low-rank tensor, decomposing it will yield the exact physical factor matrices representing the channel.
This mathematical uniqueness serves as an elegant engine for solving hardware and multi-user interference problems:
- Joint Multi-User Estimation: Zhou Zhou et al. employ a layered pilot transmission scheme where users transmit pilot symbols using a common beamforming vector. This structures the received signal into a third-order tensor, allowing the base station to jointly estimate the channels of multiple users simultaneously by decomposing the tensor into its rank-one path components.
- Hardware Calibration Defiance: In array processing, Junli Liang uses a three-way tensor trilinear model for 2-D Direction of Arrival (DOA) estimation. Because the tensor decomposition is inherently unique, the algorithm separates signal components and restores steering vectors without needing prior knowledge of the array's mutual coupling coefficients.
- Dual-Polarization and Directionality: Cheng Qian et al. (2018) model double-directional massive MIMO channels with dual-polarized antennas as a low-rank four-way tensor, utilizing PARAFAC to recover underlying spatial parameters directly.
- Joint Sensing and Communication: Exploiting this structural uniqueness allows wireless signals to double as radar. Christina Chaccour et al. organize uplink THz communication signals into a third-order tensor across measurements, snapshots, and subcarriers. Applying tensor decomposition extracts high-resolution spatial parameters to map the surrounding environment without requiring dedicated sensing pilots.
Injecting Physics via Vandermonde Structures
To bypass the computational overhead and local minima traps of iterative optimization algorithms (like Alternating Least Squares), researchers inject the physical geometry of antenna arrays directly into the tensor's mathematical structure.
When using Uniform Linear Arrays (ULAs), the phase shifts across antenna elements form a geometric progression. This progression translates mathematically into a Vandermonde matrix. By designing structured, conjugate-symmetric training sequences, Cheng Qian et al. (2019) force the received signal tensor to possess a Vandermonde factor matrix. Because Vandermonde matrices have highly rigid algebraic properties, the authors derive closed-form, non-iterative algebraic solutions for channel estimation, identifying more signal paths than there are receive antennas.
Physical Domain: Uniform Antenna Array (Equal Spacing Δd)
↓
Signal Domain: Geometric Phase Shifts (e^-jφ, e^-j2φ, e^-j3φ)
↓
Algebraic Domain: Vandermonde Factor Matrix ---> Closed-form, single-pass estimation
Building on this concept, Panqi Chen et al. constrain three of the four factor matrices in their channel tensor to have a Vandermonde structure. They introduce a single-pass algorithm (VSD-Fort) that exploits this multi-Vandermonde structure to accurately resolve channel parameters even for high-rank tensors representing up to 420 sub-paths.
Taming Cascaded Channels in RIS Systems
Reconfigurable Intelligent Surfaces (RIS) introduce a unique architectural challenge: they consist of passive reflective elements without active radio chains. Consequently, they cannot process or transmit their own pilot signals. The base station only observes a "cascaded" channel (Base Station $\rightarrow$ RIS $\rightarrow$ User), which mathematically manifests as a product of multiple channel matrices, leading to a severe parameter explosion.
Tensor modeling elegantly untangles this multiplicative cascade. By structuring multi-user signals across different RIS phase-shift configurations as a three-way tensor, Chongwen Huang et al. and Chau Yuen et al. (2021) demonstrate that the cascaded link can be uniquely decoupled. PARAFAC decomposition unfolds the tensor, enabling the iterative estimation of the separate base-station-to-RIS and RIS-to-user matrices without requiring active processing at the surface.
Passive RIS Problem:
[Base Station] ----(Matrix H1)----> [Passive RIS] ----(Matrix H2)----> [User]
↓
Observed: Cascaded Product (H1 × H2)
Tensor Solution:
Stack multiple RIS phase configurations into a 3D Tensor Train ---> Uniquely separates H1 and H2
For ultra-massive surfaces where near-field effects dominate, Tierui Gong et al. decompose the electromagnetic tensor Green's function into local surface offsets and a central displacement vector, simplifying holographic MIMO estimation via Kronecker and Hadamard products. To understand the fundamental limits of these systems, Zhen Qin et al. prove that the cascaded channel fits a Tensor Train model. Identifying this specific multiplication-form structure reveals the true, lower degrees of freedom in the channel, allowing the authors to establish tight theoretical bounds on estimation accuracy.
Multi User Decoupling and Massive Access
In massive machine-type communications (such as 6G IoT networks), the base station faces a dual challenge: it must simultaneously determine which devices are active (device detection) and what their channels look like (channel estimation), often without the luxury of pre-allocated orthogonal pilots.
Tensor frameworks handle this by stacking multi-user signals into coupled tensors. Xiaodan Shao et al. (2021) map transmitted data sequences into high-order rank-one tensors. In multi-user environments, tensor decompositions can suffer from permutation and scaling ambiguities across different data blocks. The authors resolve this by utilizing shared channel parameters across multiple data subblocks to act as an "anchor," providing a consistent reference that enables blind, simultaneous device detection and channel estimation.
To scale this to massive random access without manual parameter tweaking, Xiaodan Shao et al. (2022) combine coupled tensor modeling with sparse Bayesian learning, exploiting the angular sparsity between the base station and the RIS.
Finally, to prevent overfitting in highly complex environments, Lei Cheng et al. formulate the tensor decomposition within a variational inference framework. This approach successfully addresses the mathematical coupling of overlapping multi-user signals, allowing the receiver to automatically learn environmental complexity and decouple individual user channels from a single non-standard tensor observation.
Key Foundational Work
Channel Estimation for Millimeter-Wave Multiuser MIMO Systems via PARAFAC Decomposition (2016) — contributed 4 concept elements to this direction.
TWO L-SHAPED ARRAY-BASED 2-D DOAS ESTIMATION IN THE PRESENCE OF MUTUAL COUPLING (2011) — contributed 1 concept elements to this direction.