Source-linked AI summary
PARAFAC-Based Channel Estimation for Intelligent Reflective Surface Assisted MIMO System
Gilderlan T. de Araújo, André L. F. de Almeida
TL;DR
Channel-state-information acquisition in IRS-assisted communications is challenged by passive surfaces and their large number of elements. The paper uses structured pilots and IRS phase shifts to cast received signals as a PARAFAC tensor and proposes two channel-estimation algorithms. The methods have similar performance but trade lower complexity against greater flexibility in training parameters.
Problem
Channel-state-information acquisition is an open issue in IRS-assisted communications, where passive operation and the large number of IRS elements are obstacles to estimation.
Source: §I · p. 1Method
The paper uses a structured time-domain pilot and IRS phase-shift protocol to model received signals with PARAFAC and estimate both channels using closed-form or iterative methods.
Source: §I · p. 1, §II · p. 2Results
The two algorithms exhibit similar performance, while the closed-form LSKRF method has lower complexity and BALS works with more flexible training-parameter choices.
Source: §V · p. 4Takeaways & Limitations
Choosing between LSKRF and BALS involves trading lower complexity for less restrictive training-parameter conditions.
Source: §IV · p. 4, §V · p. 4
Abstract
from arXiv · showhide
Intelligent reflective surface (IRS) is an emergent technology for future wireless communications. It consists of a large 2D array of passive scattering elements that control the electromagnetic properties of radio-frequency waves so that the reflected signals add coherently at the intended receiver or destructively to reduce co-channel interference. The promised gains of IRS-assisted communications depend on the accuracy of the channel state information. In this paper, we propose two novel channel estimation methods for an IRS-assisted multiple-input multiple-output (MIMO) communication system. Assuming a structured time-domain pattern of pilots and IRS phase shifts, we show that the received signal follows a parallel factor (PARAFAC) tensor model that can be exploited to estimate the involved communication channels in closed-form or iteratively. Numerical results corroborate the effectiveness of the proposed channel estimation methods and highlight the involved tradeoffs.
I. INTRODUCTION
The introduction identifies channel-state-information acquisition as an open issue in IRS-assisted communications, where passive operation and many IRS elements complicate estimation. It presents a PARAFAC-based approach with two algorithms that trade computational complexity against training-parameter flexibility.
I. INTRODUCTION
- Passive operation and the large number of IRS elements are identified as two obstacles to channel estimation.The section frames channel-state-information acquisition as an open issue for IRS-assisted communications.Source: §I · p. 1
- Prior work includes an MVU estimator and IRS phase-shift design [7], sparse two-stage estimation [8], and beam-training or activation-based approaches [9]–[11].The cited approaches address channel estimation in IRS-assisted settings, including massive MIMO, Terahertz, and mmWave contexts.Source: §I · p. 1
- The paper models the received signal as a PARAFAC tensor under structured time-domain pilots and IRS phase shifts.It uses this algebraic structure to estimate the transmitter–IRS and IRS–receiver channels.Source: §I · p. 1
- The closed-form rank-1 approximation method is less complex, while bilinear alternating least squares operates with less restrictive training-parameter conditions.The introduction presents illustrative numerical results to evaluate both methods.Source: §I · p. 1
II. SYSTEM MODEL
The system model considers an IRS-assisted MIMO link and imposes a structured protocol in which pilots repeat across blocks while IRS phases vary between them. Under this protocol, the received signal can be represented as a PARAFAC tensor for channel estimation.
II. SYSTEM MODEL
- The model has M transmit and L receive antennas, an IRS with N individually adjustable elements, and assumes the direct BS–UT line-of-sight path is unavailable.Although framed as downlink, the model also applies to uplink by reversing transmitter and receiver roles.Source: §II · pp. 1–2
- Across K blocks of T time slots, pilots repeat from block to block while the IRS phase vector remains fixed within each block and varies between blocks.The coherence time is divided into these blocks, with Ts = KT.Source: §II · p. 2
- With the structured protocol, the received signal is organized as a three-way tensor following a PARAFAC decomposition.Its dimensions are L × T × K, and its factor matrices are G, Z, and S.Source: §II · p. 2
- The paper uses the PARAFAC model’s algebraic structure to formulate two channel-estimation methods.The received signals from each block are collected into matrices Y[k].Source: §II · p. 2
III. PARAFAC-BASED CHANNEL ESTIMATION
The estimation section defines the received tensor and its noisy mode unfoldings, then specifies semi-unitary pilot and IRS phase-shift matrices as the default design. Truncated DFT matrices are offered as a good choice for both.
III. PARAFAC-BASED CHANNEL ESTIMATION
- The estimation goal is to recover the BS–IRS channel H and IRS–UT channel G from the received signal tensor.The noisy tensor is defined as Y + N, with N the additive-noise tensor.Source: §III · p. 2
- The three noisy mode unfoldings of the received tensor are formed by unfolding the noise tensor in the corresponding modes.These are denoted Yi, i = 1, 2, 3, with corresponding noise unfoldings Ni.Source: §III · p. 2
- By default, the pilot matrix X and IRS phase-shift matrix S are semi-unitary; truncated DFT matrices are suggested for both.The stated conditions are X^H X = I_M and S^H S = I_N.Source: §III · p. 3
A. Closed-form solution
The closed-form method filters the received-signal unfolding using the IRS and pilot matrices, then estimates the channels through rank-1 matrix approximations. The factors are recovered from dominant singular vectors.
A. Closed-form solution
- Bilinear filtering of Y3 with the IRS and pilot matrices produces W, a noisy Khatri-Rao-structured virtual MIMO channel.The filtered noise is denoted ˜N3.Source: A. Closed-form solution · p. 3
- The channel estimates are obtained by solving a least-squares problem that minimizes a set of rank-1 matrix approximation errors.The problem can be solved with the LSKRF algorithm [18],.Source: A. Closed-form solution · p. 3
- For each IRS element, dominant left and right singular vectors of ˜Wn provide the corresponding estimates of gn and hn.This reduces channel estimation to N rank-1 matrix approximation subproblems; truncated SVD returns the dominant singular vectors.Source: A. Closed-form solution · p. 3
B. Iterative solution
The iterative BALS method alternates least-squares estimates of G and H using noisy matrix unfoldings. It monitors reconstruction error and typically converges in fewer than 10 iterations.
B. Iterative solution
- BALS alternates least-squares estimates of G and H using the noisy versions of the matrix unfoldings.The procedure repeats the updates until convergence.Source: B. Iterative solution · p. 3
- Column-orthogonal X and S allow the left pseudo-inverses in the channel updates to be replaced by lower-complexity matrix products.The paper summarizes the method in Algorithm 2.Source: B. Iterative solution · p. 3
- Figure 3 plots NMSE against SNR for the estimated channels H and G.The supplied figure caption identifies the channels whose NMSE is shown.Source: Figure 3 · p. 4
- BALS usually converges in fewer than 10 iterations, with stopping based on a reconstruction-error change of at most 10^-6.The error is computed from the reconstructed PARAFAC model formed from the current channel estimates.Source: B. Iterative solution · pp. 3–4
C. Design requirements
The methods have different training-design requirements: LSKRF needs semi-unitary or column-orthogonal pilot and IRS matrices, while BALS requires full-column-rank update matrices.
C. Design requirements
- LSKRF requires semi-unitary or column-orthogonal S and X, implying K ≥ N and T ≥ M.These conditions are needed for its bilinear filtering step.Source: C. Design requirements · p. 4
- BALS requires full-column-rank update matrices, implying K min(T, L) ≥ N; this permits fewer blocks than LSKRF but is more computationally complex.The paper identifies a tradeoff between BALS's more flexible operating conditions and LSKRF's usually lower computational complexity.Source: C. Design requirements · p. 4
D. Ambiguities
Both channel estimates are unique up to compensating scalar ambiguities, which disappear in the cascaded-channel estimate; knowledge of the IRS matrix also removes permutation ambiguity.
D. Ambiguities
- The estimated rows of H and columns of G can carry compensating scaling factors, so the channel estimates are unique only up to scalar ambiguities.The factors compensate each other across the two channel matrices.Source: D. Ambiguities · p. 4
- Forming the cascaded channel estimate G H^T removes the scalar ambiguities, and known S prevents permutation ambiguity.Source: D. Ambiguities · p. 4
IV. NUMERICAL RESULTS
The simulations evaluate channel-estimation accuracy using NMSE and compare the cascaded-channel estimates and runtimes of LSKRF and BALS. Both methods provide satisfactory performance, with a tradeoff between computational complexity and training-parameter flexibility.
IV. NUMERICAL RESULTS
- Channel-estimation accuracy is evaluated using normalized mean square error (NMSE), with results averaged over 3000 Monte Carlo runs.The simulations use T = 4, L = 2, K = 50, M = 3, and N ∈ {10, 40}.Source: §IV · p. 4
- Both algorithms provide satisfactory NMSE performance, but performance degrades as the number of IRS elements increases.The evaluation includes NMSE of the estimated cascaded channel, whose results follow the behavior observed for the channel estimates.Source: §IV · p. 4, Figure 4 · p. 4
- BALS is more complex than LSKRF, and its runtime grows faster as the number of IRS elements increases.BALS can operate with less restrictive choices of K and T, yielding a tradeoff between complexity and operating conditions.Source: §IV · p. 4, Figure 5 · p. 4
V. CONCLUSION
The conclusion presents two PARAFAC-based channel-estimation methods with similar performance but different complexity and training-parameter requirements.
V. CONCLUSION
- The two proposed algorithms exhibit similar channel-estimation performance.Source: §V · p. 4
- The closed-form LSKRF method has lower complexity but more restrictive requirements on training parameters K and T.Source: §V · p. 4
- The iterative BALS method is more computationally complex but supports more flexible choices of K and T.Source: §V · p. 4