betaFeedback

WorksEngineering

Structured PARAFAC Channel Estimation for Flexible Intelligent Metasurface Receivers

Abstract

Flexible intelligent metasurfaces (FIMs) create geometric diversity through aperture morphing, but their configuration-dependent responses complicate uplink channel estimation. This letter considers multiple users sharing a common angular response and develops a structured PARAFAC model across calibrated FIM configurations. Its known identity factor and rank-one pilot-weighted channel factor enable deterministic, noniterative algebraic factor recovery through per-element rank-one approximations, reference-state normalization, global rank-one projection, and a least-squares morphing-factor update. The recovered spatial and morphing factors are then jointly matched to the calibrated manifold to estimate the common angles, user gains, and configuration-dependent channels. Simulations demonstrate accurate recovery and show that out-of-plane morphing resolves the azimuth-sign ambiguity of an undeformed planar aperture; an angular-spread experiment quantifies the mismatch when the common-response assumption is relaxed.

Introduction

FLEXIBLE intelligent metasurfaces (FIMs) introduce controllable geometric degrees of freedom by adjusting the positions of their constituent elements, thereby reshaping the effective aperture and modifying the spatial response [1, 2]. This geometric reconfigurability provides additional propagation and spatial-diversity opportunities, but it also complicates channel acquisition because the array response varies with the adopted morphing configuration. Accurate channel estimation must therefore account for the coupling between the underlying propagation parameters and the prescribed FIM geometry.

Recent studies have investigated channel estimation for FIM and related geometry-reconfigurable architectures. Sparse angular recovery with FIM-shape optimization was considered in [3], while model-based interpolation, kernel methods, sparse parametric recovery, and neural-operator learning were studied in [4]. For movable-antenna MIMO systems, tensor decomposition was employed to exploit measurements collected over antenna movements [5], and an alternating-least-squares PARAFAC formulation was introduced for point-to-point MIMO systems equipped with morphing FIM arrays [6]. Tensor-based receiver designs have been investigated in UAV-aided MIMO systems, including information monitoring receivers [7]. Tensor-based channel estimation has also been widely explored in systems incorporating reconfigurable intelligent surfaces (RISs) [9, 10], including PARAFAC models for RIS-assisted multiple-input multiple-output (MIMO) systems [12, 11] and low-rank tensor recovery for RIS-assisted millimeter-wave orthogonal frequency division multiplexing (OFDM) [13]. For multi-RIS uplink MIMO systems, a PARATUCK2-based framework was developed for joint channel and symbol estimation, followed by subspace-based geometric parameter extraction [14]. Structured tensor factorizations have also been investigated for beyond-diagonal RIS architectures [15]. These studies establish the value of tensor modeling, while motivating estimators that explicitly exploit the structure of multiuser uplinks observed across calibrated morphing states.

To address this need, we study uplink channel estimation for multiple users observed over prescribed, calibrated FIM morphing configurations. We focus on the common-angular-response regime, in which users share a dominant arrival direction as seen by the FIM; this model also approximates compact user clusters unresolved by the effective aperture. Relative to the ALS-based point-to-point FIM-PARAFAC formulation in [6], we target the multiuser uplink and exploit the structure it induces: the known identity factor decouples the model unfolding into per-element rank-one matrix approximations, while the pilot-weighted channel factor obeys a global rank-one constraint. We use these properties for reference-state normalization, global rank-one projection, and a single least-squares morphing-factor refinement, avoiding alternating PARAFAC updates. We then jointly match the recovered spatial and morphing factors to the calibrated FIM manifold to estimate the common angles, user gains, and configuration-dependent channels. Simulations demonstrate accurate factor recovery and show how morphing resolves the azimuth-sign ambiguity inherent to the undeformed planar aperture. A

This work was supported in part by the National Key Research and Development Program of China under Grant 2024YFE0200103, the China Scholarship Council (Grant No. 202508110189), the NSFC under Grant 62001008, and the Beijing Municipal Natural Science Foundation under Grants 4212002 and L192034. This work is partially supported by the National Institute of Science and Technology (INCT-SigNals) sponsored by Brazil’s National Council for Scientific and Technological Development (CNPq) (406517/2022-3), and FUNCAP (INCT-25255-82587.32.41/64). The research of André L. F. de Almeida is partially supported by CNPq (303356/2025-1). nonzero-angular-spread experiment additionally quantifies the model mismatch as this assumption is relaxed.

System Model and FIM-PARAFAC Formulation

As illustrated in Figure 1, we consider a narrowband block-fading uplink with KK single-antenna users and an NN-element FIM receiver. In the undeformed reference state, the N=NxNzN=N_xN_z elements form a planar array in the xzxz-plane, indexed by n=(p−1)Nz+qn=(p-1)N_z+q, with xp=(p−1)dxx_p=(p-1)d_x and zq=(q−1)dzz_q=(q-1)d_z. During training, the FIM cycles through MM prescribed, calibrated morphing configurations. Element positions remain fixed along xx and zz and vary only along yy; thus, the position of element nn in configuration mm is rn,m=[xp,yn,m,zq]T\mathbf{r}_{n,m}=[x_p,y_{n,m},z_q]^{T}, where yn,my_{n,m} is the element’s out-of-plane displacement. The displacement profile of the mmth configuration is collected as ym=[y1,m,…,yN,m]T∈RN\mathbf{y}_m=[y_{1,m},\ldots,y_{N,m}]^{T}\in\mathbb{R}^{N}.

FIM-assisted uplink training and structured PARAFAC formulation

Figure 1. FIM-assisted uplink training and structured PARAFAC formulation.

Consider a far-field plane wave with elevation angle θ\theta and azimuth angle ϕ\phi, with unit direction vector u(θ,ϕ)=[sin⁡θcos⁡ϕ,sin⁡θsin⁡ϕ,cos⁡θ]T\mathbf{u}(\theta,\phi)=[\sin\theta\cos\phi,\sin\theta\sin\phi,\cos\theta]^{T}. Let λ\lambda denote the carrier wavelength and κ=2π/λ\kappa=2\pi/\lambda the wavenumber. The response of the nnth FIM element under the mmth morphing configuration is [am(θ,ϕ)]n=ejκrn,mTu(θ,ϕ)[\mathbf{a}_m(\theta,\phi)]_n=e^{\mathrm{j}\kappa\mathbf{r}_{n,m}^{T}\mathbf{u}(\theta,\phi)}. Let the undeformed UPA steering vector be aUPA(θ,ϕ)=ax(θ,ϕ)⊗az(θ)\mathbf{a}_{\mathrm{UPA}}(\theta,\phi)=\mathbf{a}_x(\theta,\phi)\otimes\mathbf{a}_z(\theta), and the morphing-induced phase vector be em(θ,ϕ)=ejκsin⁡θsin⁡ϕym\mathbf{e}_m(\theta,\phi)=e^{\mathrm{j}\kappa\sin\theta\sin\phi\mathbf{y}_m}, where the exponential is applied elementwise. The complete FIM steering vector is therefore am(θ,ϕ)=aUPA(θ,ϕ)⊙em(θ,ϕ)\mathbf{a}_m(\theta,\phi)=\mathbf{a}_{\mathrm{UPA}}(\theta,\phi)\odot\mathbf{e}_m(\theta,\phi). The first morphing configuration is selected as the undeformed reference state, i.e.,

y1=0N,e1(θ,ϕ)=1N.(1)\mathbf{y}_1=\mathbf{0}_N,\qquad\mathbf{e}_1(\theta,\phi)=\mathbf{1}_N. \tag*{(1)}

During uplink training, user kk transmits a known pilot sequence sk∈CT\mathbf{s}_k\in\mathbb{C}^{T}, and all pilot sequences are collected in S=[s1,…,sK]T∈CK×T\mathbf{S}=[\mathbf{s}_1,\ldots,\mathbf{s}_K]^{T}\in\mathbb{C}^{K\times T}. Let αk∈C\alpha_k\in\mathbb{C} denote the complex channel gain of user kk, and let (θk,ϕk)(\theta_k,\phi_k) denote its angular parameters. Under the single-dominant-path model, the channel between user kk and the FIM under configuration mm is hk,m=αkam(θk,ϕk)\mathbf{h}_{k,m}=\alpha_k\mathbf{a}_m(\theta_k,\phi_k). The received pilot matrix under the mmth morphing configuration is therefore

Ym=∑k=1Khk,mskT+Wm,m=1,…,M,(2)\mathbf{Y}_m=\sum_{k=1}^{K}\mathbf{h}_{k,m}\mathbf{s}_k^{T}+\mathbf{W}_m,\qquad m=1,\ldots,M, \tag*{(2)}

where Wm∈CN×T\mathbf{W}_m\in\mathbb{C}^{N\times T} denotes additive noise.

In this work, we focus on the angularly co-located-user configuration, for which θk=θ\theta_k=\theta and ϕk=ϕ\phi_k=\phi for k=1,…,Kk=1,\ldots,K. This condition applies exactly to users sharing the same dominant arrival direction and also serves as a nominal model for a compact user cluster that the effective FIM aperture cannot resolve. Let a≜aUPA(θ,ϕ)\mathbf{a}\triangleq\mathbf{a}_{\mathrm{UPA}}(\theta,\phi), em≜em(θ,ϕ)\mathbf{e}_m\triangleq\mathbf{e}_m(\theta,\phi), and α=[α1,…,αK]T\boldsymbol{\alpha}=[\alpha_1,\ldots,\alpha_K]^T. Define the pilot-weighted channel gain vector as z≜STα\mathbf{z}\triangleq\mathbf{S}^{T}\boldsymbol{\alpha}. The signal model in (2) is equivalent to

Ym=diag⁡{em}azT+Wm.(3)\mathbf{Y}_m=\operatorname{diag}\{\mathbf{e}_m\}\mathbf{a}\mathbf{z}^{T}+\mathbf{W}_m. \tag*{(3)}

Structured FIM-PARAFAC Representation

The received matrices {Ym}m=1M\{\mathbf{Y}_m\}_{m=1}^{M} are stacked along the morphing configuration dimension to construct the third-order observation tensor Y∈CN×T×M\mathcal{Y}\in\mathbb{C}^{N\times T\times M}, with Y(:,:,m)=Ym\mathcal{Y}(:,:,m)=\mathbf{Y}_m. Its three modes correspond to the FIM elements, pilot symbols, and morphing configurations, respectively.

Define

H≜zaT∈CT×N,rank⁡(H)=1,(4)\mathbf{H}\triangleq\mathbf{z}\mathbf{a}^{T}\in\mathbb{C}^{T\times N},\qquad\operatorname{rank}(\mathbf{H})=1, \tag*{(4)}

and collect the morphing responses in E≜[e1,…,eM]T∈CM×N\mathbf{E}\triangleq[\mathbf{e}_1,\ldots,\mathbf{e}_M]^{T}\in\mathbb{C}^{M\times N}. Since the first configuration is undeformed, the reference condition in (1) gives E(1,:)=1NT\mathbf{E}(1,:)=\mathbf{1}_N^{T}.

Let I3,N∈RN×N×N\mathcal{I}_{3,N}\in\mathbb{R}^{N\times N\times N} denote the third-order identity tensor. Substituting (3) into the stacked observations gives the structured PARAFAC model

Y=I3,N×1IN×2H×3E+W.(5)\mathcal{Y}=\mathcal{I}_{3,N}\times_1\mathbf{I}_N\times_2\mathbf{H}\times_3\mathbf{E}+\mathcal{W}. \tag*{(5)}

where W\mathcal{W} is the stacked noise tensor. The factors are therefore IN\mathbf{I}_N, H\mathbf{H}, and E\mathbf{E}, with a known identity factor and the global rank-one constraint on H\mathbf{H} in (4). Under the adopted unfolding convention, its mode-1 unfolding is [Y](1)=(E⊙H)T+[W](1)[\mathcal{Y}]_{(1)}=(\mathbf{E}\odot\mathbf{H})^{T}+[\mathcal{W}]_{(1)}. Here, ⊙\odot denotes the Khatri–Rao product. Transposing this relation and defining N1≜[W](1)T\mathbf{N}_1\triangleq[\mathcal{W}]_{(1)}^{T} gives

Q≜[Y](1)T=E⊙H+N1.(6)\mathbf{Q}\triangleq[\mathcal{Y}]_{(1)}^{T}=\mathbf{E}\odot\mathbf{H}+\mathbf{N}_1. \tag*{(6)}

The known identity factor and the columnwise structure of (6) enable the deterministic closed-form factorization developed in the next section.

Proposed Structured FIM-PARAFAC Estimator

Columnwise Rank-One Factorization

Starting from the mode-1 representation in (6), let

Q=[q1,…,qN]=E⊙H+N1∈CMT×N.(7)\mathbf{Q}=[\mathbf{q}_1,\ldots,\mathbf{q}_N]=\mathbf{E}\odot\mathbf{H}+\mathbf{N}_1\in\mathbb{C}^{MT\times N}. \tag*{(7)}

Denote by hn\mathbf{h}_n and en\mathbf{e}_n the nnth columns of H\mathbf{H} and E\mathbf{E}, respectively. Since the Khatri–Rao product operates columnwise, the nnth column of (7) satisfies

qn=en⊗hn+nn.(8)\mathbf{q}_n=\mathbf{e}_n\otimes\mathbf{h}_n+\mathbf{n}_n. \tag*{(8)}

Using vec⁡(hnϵnT)=ϵn⊗hn\operatorname{vec}(\mathbf{h}_n\boldsymbol{\epsilon}_n^T)=\boldsymbol{\epsilon}_n\otimes\mathbf{h}_n, qn\mathbf{q}_n can be reshaped into

Qn≜unvec⁡T×M(qn)=hnϵnT+Nn,n=1,…,N,(9)\mathbf{Q}_n \triangleq\operatorname{unvec}_{T\times M}(\mathbf{q}_n)=\mathbf{h}_n\boldsymbol{\epsilon}_n^T+\mathbf{N}_n,\qquad n=1,\ldots,N, \tag*{(9)}

where Nn\mathbf{N}_n denotes the corresponding reshaped noise matrix. Hence, the structured tensor factorization is reduced to NN separate, algebraically decoupled rank-one matrix approximation problems. This decoupling is algebraic and does not, by itself, imply statistical independence of the noise terms.

Let

Qn=UnΣnVnH(10)\mathbf{Q}_n=\mathbf{U}_n\boldsymbol{\Sigma}_n\mathbf{V}_n^H \tag*{(10)}

be the singular value decomposition of Qn\mathbf{Q}_n, and let (σn,1,un,1,vn,1)(\sigma_{n,1},\mathbf{u}_{n,1},\mathbf{v}_{n,1}) denote its dominant singular triplet. A compatible pair of unnormalized factor estimates is

h~n=σn,1un,1,ϵ~n=σn,1vn,1∗.(11)\widetilde{\mathbf{h}}_n=\sqrt{\sigma_{n,1}}\mathbf{u}_{n,1},\qquad\widetilde{\boldsymbol{\epsilon}}_n=\sqrt{\sigma_{n,1}}\mathbf{v}_{n,1}^{*}. \tag*{(11)}

The complex conjugation in (11) follows from the transpose in hnϵnT\mathbf{h}_n\boldsymbol{\epsilon}_n^T. The SVD factors in (11) give the unweighted Frobenius-norm best rank-one approximation of each reshaped observation, h~nϵ~nT=σn,1un,1vn,1H\widetilde{\mathbf{h}}_n\widetilde{\boldsymbol{\epsilon}}_n^T=\sigma_{n,1}\mathbf{u}_{n,1}\mathbf{v}_{n,1}^H, with squared residual ∑i=2min⁡{T,M}σn,i2\sum_{i=2}^{\min\{T,M\}}\sigma_{n,i}^2. The minimizer is unique when its dominant singular value is simple, although its factor pair retains a reciprocal scalar ambiguity. Such an unweighted least-squares fit corresponds to a maximum-likelihood estimate only under an i.i.d. white Gaussian noise model. The two factors in (11) are subject to a reciprocal scaling ambiguity. This ambiguity is resolved by the undeformed reference configuration in (1), which implies [ϵn]1=1[\boldsymbol{\epsilon}_n]_1=1. Defining

βn≜[ϵ~n]1,(12)\beta_n\triangleq[\widetilde{\boldsymbol{\epsilon}}_n]_1, \tag*{(12)}

the normalized estimates are obtained as

h^n=βnh~n,ϵ^n=ϵ~nβn.(13)\widehat{\mathbf{h}}_n=\beta_n\widetilde{\mathbf{h}}_n,\qquad\widehat{\boldsymbol{\epsilon}}_n=\frac{\widetilde{\boldsymbol{\epsilon}}_n}{\beta_n}. \tag*{(13)}

Under the exact model, βn≠0\beta_n\ne0 because the first entry of each morphing-factor column equals one. With noise, βn≠0\beta_n\ne0 remains a necessary nondegeneracy condition, and a small ∣βn∣|\beta_n| makes this reference-state normalization ill-conditioned.

Collecting the normalized column estimates yields

HCF=[h^1,…,h^N],ECF=[ϵ^1,…,ϵ^N].(14)\mathbf{H}_{\mathrm{CF}}=[\widehat{\mathbf{h}}_1,\ldots,\widehat{\mathbf{h}}_N],\qquad\mathbf{E}_{\mathrm{CF}}=[\widehat{\boldsymbol{\epsilon}}_1,\ldots,\widehat{\boldsymbol{\epsilon}}_N]. \tag*{(14)}

Global Rank-One Projection and Factor Refinement

Since the matrices Qn\mathbf{Q}_n are approximated separately, their assembled estimate HCF\mathbf{H}_{\mathrm{CF}} will generally not satisfy the global physical constraint rank⁡(H)=1\operatorname{rank}(\mathbf{H})=1 in noise. We therefore enforce the common angular-response structure by

Hr1=P1{HCF},(15)\mathbf{H}_{\mathrm{r1}}=\mathcal{P}_1\{\mathbf{H}_{\mathrm{CF}}\}, \tag*{(15)}

where P1{⋅}\mathcal{P}_1\{\cdot\} denotes the best rank-one approximation in the Frobenius-norm sense. Specifically, if (σCF,1,uCF,1,vCF,1)(\sigma_{\mathrm{CF},1},\mathbf{u}_{\mathrm{CF},1},\mathbf{v}_{\mathrm{CF},1}) is a dominant singular triplet of HCF\mathbf{H}_{\mathrm{CF}}, then

Hr1=arg⁡min⁡X∈CT×Nrank⁡(X)≤1∥HCF−X∥F2=σCF,1uCF,1vCF,1H.(16)\mathbf{H}_{\mathrm{r1}}=\underset{\mathbf{X}\in\mathbb{C}^{T\times N}}{\arg\min}_{\operatorname{rank}(\mathbf{X})\le1}\|\mathbf{H}_{\mathrm{CF}}-\mathbf{X}\|_F^2=\sigma_{\mathrm{CF},1}\mathbf{u}_{\mathrm{CF},1}\mathbf{v}_{\mathrm{CF},1}^H. \tag*{(16)}

This projection is the nearest rank-one matrix to the assembled factor. The subsequent update is a single conditional refinement. For fixed Hr1\mathbf{H}_{\mathrm{r1}}, define the mode-3 design matrix

Ar1≜(Hr1⋄IN)T∈CN×NT.(17)\mathbf{A}_{\mathrm{r1}}\triangleq(\mathbf{H}_{\mathrm{r1}}\diamond\mathbf{I}_N)^T\in\mathbb{C}^{N\times NT}. \tag*{(17)}

The columns of Hr1⋄IN\mathbf{H}_{\mathrm{r1}}\diamond\mathbf{I}_N have disjoint support, which gives

Ar1Ar1H=diag⁡{∥hr1,1∥2,…,∥hr1,N∥2}.(18)\mathbf{A}_{\mathrm{r1}}\mathbf{A}_{\mathrm{r1}}^H=\operatorname{diag}\left\{\|\mathbf{h}_{\mathrm{r1},1}\|^2,\ldots,\|\mathbf{h}_{\mathrm{r1},N}\|^2\right\}. \tag*{(18)}

Here, hr1,n\mathbf{h}_{\mathrm{r1},n} denotes the nnth column of Hr1\mathbf{H}_{\mathrm{r1}}. Thus, if every such column is nonzero, Ar1\mathbf{A}_{\mathrm{r1}} has full row rank and the unconstrained LS estimate of the morphing factor is unique:

E~CF=arg⁡min⁡E∈CM×N ∥Y(3)−EAr1∥F2=Y(3)Ar1†.(19)\widetilde{\mathbf{E}}_{\mathrm{CF}}=\underset{\mathbf{E}\in\mathbb{C}^{M\times N}}{\arg\min}\ \|\mathcal{Y}_{(3)}-\mathbf{E}\mathbf{A}_{\mathrm{r1}}\|_F^2=\mathcal{Y}_{(3)}\mathbf{A}_{\mathrm{r1}}^\dagger. \tag*{(19)}

This is ordinary, unweighted conditional least squares (and is conditionally maximum likelihood under i.i.d. white Gaussian noise). Define DCF≜diag⁡{(E~CF(1,:))T}\mathbf{D}_{\mathrm{CF}}\triangleq\operatorname{diag}\{(\widetilde{\mathbf{E}}_{\mathrm{CF}}(1,:))^T\}, the final closed-form factor estimates are then

E^=E~CFDCF−1,H^=Hr1DCF.(20)\widehat{\mathbf{E}}=\widetilde{\mathbf{E}}_{\mathrm{CF}}\mathbf{D}_{\mathrm{CF}}^{-1},\qquad\widehat{\mathbf{H}}=\mathbf{H}_{\mathrm{r1}}\mathbf{D}_{\mathrm{CF}}. \tag*{(20)}

Provided every diagonal entry of DCF\mathbf{D}_{\mathrm{CF}} is nonzero, this diagonal rescaling enforces the reference-state constraint without changing the fitted tensor:

E^(H^⋄IN)T=E~CFAr1.(21)\widehat{\mathbf{E}}(\widehat{\mathbf{H}}\diamond\mathbf{I}_N)^T=\widetilde{\mathbf{E}}_{\mathrm{CF}}\mathbf{A}_{\mathrm{r1}}. \tag*{(21)}

In noisy data, small diagonal entries likewise make this normalization ill-conditioned and should be treated as a degenerate estimate. Consequently,

rank⁡(H^)=1,E^(1,:)=1NT.(22)\operatorname{rank}(\widehat{\mathbf{H}})=1,\qquad\widehat{\mathbf{E}}(1,:)=\mathbf{1}_N^T. \tag*{(22)}

The algebraic factor recovery stage is deterministic and noniterative. Its main operations are NN independent dominant singular vector computations, one global rank-one projection, and one structured least-squares update; physical angle extraction is subsequently performed by the manifold-matching step.

Remark 1 (Recoverability and complexity): The proposed estimator does not rely on generic Kruskal-rank uniqueness. Since the known identity factor IN\mathbf{I}_N decouples the mode-1 unfolding into NN rank-one factorizations Qn=hnϵnT\mathbf{Q}_n=\mathbf{h}_n\boldsymbol{\epsilon}_n^T, their scalar ambiguities are removed by the reference condition [ϵn]1=1[\boldsymbol{\epsilon}_n]_1=1. Recovery of α\boldsymbol{\alpha} further requires rank⁡(S)=K\operatorname{rank}(\mathbf{S})=K, i.e., T≥KT\ge K. For algebraic factor recovery, the main cost is computing the leading singular triplet of each of the NN matrices of size T×MT\times M, with complexity O(NTMmin⁡{T,M})\mathcal{O}(NTM\min\{T,M\}). This is followed by one global rank-one projection and one structured least-squares refinement; the stated complexity excludes the subsequent angular-manifold search.

Physical Parameter and Channel Recovery

After estimating the algebraic factors H^\widehat{\mathbf{H}} and E^\widehat{\mathbf{E}}, the physical channel parameters can be recovered by exploiting the rank-one structure H=zaT\mathbf{H}=\mathbf{z}\mathbf{a}^T. Let H^=UH^ΣH^VH^H\widehat{\mathbf{H}}=\mathbf{U}_{\widehat{H}}\boldsymbol{\Sigma}_{\widehat{H}}\mathbf{V}_{\widehat{H}}^H, and denote its dominant singular triplet by (σH^,1,uH^,1,vH^,1)(\sigma_{\widehat{H},1},\mathbf{u}_{\widehat{H},1},\mathbf{v}_{\widehat{H},1}). A compatible pair of estimates is

z~=σH^,1uH^,1,a~=σH^,1vH^,1∗.(23)\widetilde{\mathbf{z}}=\sqrt{\sigma_{\widehat{H},1}}\mathbf{u}_{\widehat{H},1},\qquad\widetilde{\mathbf{a}}=\sqrt{\sigma_{\widehat{H},1}}\mathbf{v}_{\widehat{H},1}^{*}. \tag*{(23)}

Algorithm 1 Closed-Form FIM-PARAFAC Estimator

Require: Y\mathcal{Y}, S\mathbf{S}

Ensure: H^\widehat{\mathbf{H}}, E^\widehat{\mathbf{E}}, α^\widehat{\boldsymbol{\alpha}}, (θ^,ϕ^)(\widehat{\theta},\widehat{\phi})

  1. Form Q=[Y](1)T\mathbf{Q} = [\mathcal{Y}]_{(1)}^{T}.

  2. for n=1,…,Nn = 1,\ldots,N do

  3. Reshape qn\mathbf{q}_n into Qn\mathbf{Q}_n and compute its dominant singular triplet.

  4. Obtain h~n\widetilde{\mathbf{h}}_n and ϵ~n\widetilde{\boldsymbol{\epsilon}}_n from (11).

  5. Normalize the pair using (13).

  6. end for

  7. Form HCF\mathbf{H}_{\mathrm{CF}} and apply Hr1=P1{HCF}\mathbf{H}_{r1} = \mathcal{P}_1\{\mathbf{H}_{\mathrm{CF}}\}.

  8. Refine and normalize E\mathbf{E} using (19)–(20).

  9. Recover α^\widehat{\boldsymbol{\alpha}} using (24).

  10. Estimate (θ^,ϕ^)(\widehat{\theta},\widehat{\phi}) using (26).

A residual scalar ambiguity remains between z\mathbf{z} and a\mathbf{a}. Since the first element of the reference UPA steering vector is [a]1=1[\mathbf{a}]_1 = 1, define c≜[a~]1c \triangleq[\widetilde{\mathbf{a}}]_1. The reference-normalized estimates are a^=a~/c\widehat{\mathbf{a}} = \widetilde{\mathbf{a}}/c and z^=cz~\widehat{\mathbf{z}} = c\widetilde{\mathbf{z}}. The individual user gains are then recovered from the known pilot matrix as

α^=(ST)†z^.(24)\widehat{\boldsymbol{\alpha}} = (\mathbf{S}^{T})^{\dagger}\widehat{\mathbf{z}}. \tag*{(24)}

Unique recovery of the user gains requires rank⁡(S)=K\operatorname{rank}(\mathbf{S}) = K, and T≥KT \ge K. Finally, the common angular parameters are extracted by jointly matching the recovered spatial and morphing factors to the calibrated FIM manifold. Defining

E(θ,ϕ)=[e1(θ,ϕ),…,eM(θ,ϕ)]T,(25)\mathbf{E}(\theta,\phi) = [\mathbf{e}_1(\theta,\phi),\ldots,\mathbf{e}_M(\theta,\phi)]^{T}, \tag*{(25)}

the angular estimate is obtained from

(θ^,ϕ^)=arg⁡min⁡(θ,ϕ)∈Ω  1−∣aUPAH(θ,ϕ)a^∣2∥aUPA(θ,ϕ)∥22∥a^∥22+λE∥E^−E(θ,ϕ)∥F2∥E^∥F2,(26)(\widehat{\theta},\widehat{\phi}) = \underset{(\theta,\phi)\in\Omega}{\arg\min}\; 1 - \frac{\left|\mathbf{a}_{\mathrm{UPA}}^{H}(\theta,\phi)\widehat{\mathbf{a}}\right|^2}{\|\mathbf{a}_{\mathrm{UPA}}(\theta,\phi)\|_2^2\|\widehat{\mathbf{a}}\|_2^2} + \lambda_E\frac{\|\widehat{\mathbf{E}}-\mathbf{E}(\theta,\phi)\|_F^2}{\|\widehat{\mathbf{E}}\|_F^2}, \tag*{(26)}

where Ω\Omega denotes the admissible angular region and λE≥0\lambda_E \ge0 controls the contribution of the morphing factor. The channel of user kk under the mmth morphing configuration is finally reconstructed as

h^k,m=α^kam(θ^,ϕ^),k=1,…,K, m=1,…,M.(27)\widehat{\mathbf{h}}_{k,m} = \widehat{\alpha}_k\mathbf{a}_m(\widehat{\theta},\widehat{\phi}),\quad k = 1,\ldots,K,\ m = 1,\ldots,M. \tag*{(27)}

Simulation Results

We consider an FIM with N=NxNz=16N = N_xN_z = 16 elements, where Nx=Nz=4N_x = N_z = 4 and dx=dz=λ/2d_x = d_z = \lambda/2 with λ=1\lambda= 1. Unless otherwise stated, K=3K = 3, T=16T = 16, and M=6M = 6, and all users share the common angular pair (θ,ϕ)=(35∘,25∘)(\theta,\phi) = (35^\circ,25^\circ). The channel gains follow αk∼CN(0,1/K)\alpha_k \sim\mathcal{CN}(0,1/K), while the nonreference element displacements are independently drawn from U[−λ/4,λ/4]\mathcal{U}[-\lambda/4,\lambda/4]. Deterministic DFT pilots satisfying SSH=TIK\mathbf{S}\mathbf{S}^{H} = T\mathbf{I}_K are employed, and all results are averaged over NMC=2000N_{\mathrm{MC}} = 2000 realizations. For each realization, the noise variance is set as σw2=PX10−SNR/10\sigma_w^2 = P_X10^{-\mathrm{SNR}/10}, where PX=∥X∥F2/(NMT)P_X = \|\mathbf{X}\|_F^2/(NMT). The NMSE is averaged over all Monte Carlo realizations and reported in dB. The joint angular RMSE is

RMSEang=12NMC∑i=1NMC[(Δθ(i))2+(Δϕ(i))2].(28)\mathrm{RMSE}_{\mathrm{ang}} = \sqrt{\frac{1}{2N_{\mathrm{MC}}}\sum_{i=1}^{N_{\mathrm{MC}}}\left[(\Delta\theta^{(i)})^2+(\Delta\phi^{(i)})^2\right]}. \tag*{(28)}

where Δϕ(i)\Delta\phi^{(i)} is wrapped to [−180∘,180∘)[-180^\circ,180^\circ). For a fair comparison, the equal-overhead rigid-panel LS baseline uses Trigid=MTT_{\mathrm{rigid}} = MT pilots under the undeformed configuration, with Srigid=[S,…,S]\mathbf{S}_{\mathrm{rigid}} = [\mathbf{S},\ldots,\mathbf{S}].

Estimation and Angular-Recovery Performance

Figure 2(a) compares the normalized mean squared error (NMSE) of the equivalent rank-one matrix H\mathbf{H}. The proposed structured estimator improves as SNR increases, whereas the morphing-averaged singular value decomposition (SVD) exhibits an error floor because it does not account for configuration-dependent phase responses. The oracle LS curve serves as a lower bound. At equal pilot-symbol overhead, the rigid-panel LS attains lower H-NMSE because all MTMT pilots observe the same fixed spatial response and thus provide greater noise averaging. The proposed method instead distributes those pilots across multiple morphing configurations and estimates the morphing-dependent factor jointly. This tradeoff highlights the distinct benefit of morphing thanks to improved angular identifiability. Figure 2(b) demonstrates the value of morphing diversity for physical-angle recovery. The proposed estimator’s angular root mean square error (RMSE) decreases with SNR, whereas the rigid-panel LS retains an error floor of approximately 35∘35^\circ. For the undeformed xx-zz-plane array, aUPA(θ,ϕ)=aUPA(θ,−ϕ)\mathbf{a}_{\mathrm{UPA}}(\theta,\phi) = \mathbf{a}_{\mathrm{UPA}}(\theta,-\phi), so the sign of ϕ\phi is unidentifiable. With ϕ=25∘\phi= 25^\circ, the indistinguishable solution −25∘-25^\circ incurs a 50∘50^\circ azimuth error and, under the joint RMSE definition in (28), yields approximately 35.36∘35.36^\circ. Out-of-plane morphing adds a phase dependence on sin⁡θsin⁡ϕ\sin\theta\sin\phi, breaking this ambiguity through joint spatial–morphing manifold matching.

Performance versus SNR plots for NMSE of H and joint angular RMSE

Figure 2. Performance versus SNR for N=16N = 16, K=3K = 3, T=16T = 16, and M=6M = 6: (a) NMSE of H\mathbf{H}; (b) joint angular RMSE.

Impact of Morphing Diversity and Model Mismatch

Figure 3a shows that the NMSE of α\alpha decreases as MM grows, since additional calibrated configurations provide more observations for structured factor recovery. The gains diminish at larger MM. The closed-form and projected alternating estimators perform nearly identically under the matched common angular-response model, indicating that the noniterative factor-recovery stage captures most of the available benefit. Because each configuration uses TT pilot symbols, increasing MM also increases the total training overhead in this experiment.

Figure 3. Impact of morphing diversity and angular spread: (a) NMSE of α\alpha versus MM at SNR = 15 dB; (b) noise-free common-angular-response model mismatch versus angular-spread half-width.

Figure 3b assesses the common angular-response model under nonzero angular spread. It plots the normalized, noise-free residual between the true multiuser observation tensor and its best structured common-angle PARAFAC approximation. At zero spread, all users share the same spatial and morphing responses, so the residual approaches numerical precision. As the spread grows, user-dependent responses cannot be represented by the single rank-one factor H=zaTH = za^{T}, and the irreducible model mismatch increases. This experiment therefore indicates the regime in which the proposed formulation is appropriate: users with a shared dominant direction or a sufficiently compact angular cluster.

Conclusion

This letter introduced a structure-aware PARAFAC estimator for multiuser uplink channel estimation across calibrated FIM morphing configurations under a common-angular-response model. The estimator exploits the known identity factor, the global rank-one pilot-weighted channel structure, and the undeformed reference state to recover the algebraic factors deterministically without alternating updates; calibrated-manifold matching then yields the common angles, user gains, and configuration-dependent channels. Simulations demonstrate accurate factor recovery, show that out-of-plane

morphing resolves the azimuth-sign ambiguity of an undeformed planar aperture, and quantify the model mismatch that arises with nonzero angular spread.

References

  1. [1]J. An, C. Yuen, M. Di Renzo, M. Debbah, H. V. Poor, and L. Hanzo, “Flexible intelligent metasurfaces for downlink multiuser MISO communications,” IEEE Trans. Wireless Commun., vol. 24, no. 4, pp. 2940–2955, 2025.DOI
  2. [2]J. An, Z. Han, D. Niyato, M. Debbah, C. Yuen, and L. Hanzo, “Flexible intelligent metasurfaces for enhancing MIMO communications,” IEEE Trans. Commun., vol. 73, no. 9, pp. 7349–7365, 2025.arxiv.org/abs/2502.16478
  3. [3]J. Jiang, J. An, L. Gan, N. Al-Dhahir, A. Nallanathan, and Z. Han, “Enhanced channel estimation for flexible intelligent metasurface-aided communication systems,” IEEE Trans. Veh. Technol., vol. 75, no. 6, pp. 11 698–11 703, 2026.DOI
  4. [4]J. Xiao, J. Wang, Q. Cui, Y. Yang, X. Li, D. Niyato, and C. Yuen, “Channel estimation for flexible intelligent metasurfaces: From model-based approaches to neural operators,” IEEE Trans. Wireless Commun., vol. 25, no. 10,684–10 701, 2026.DOI
  5. [5]R. Zhang, L. Cheng, W. Zhang, X. Guan, Y. Cai, W. Wu, and R. Zhang, “Channel estimation for movable-antenna MIMO systems via tensor decomposition,” IEEE Wireless Commun. Lett., vol. 13, no. 11, pp. 3089–3093, 2024.DOI
  6. [6]V. L. Romano, A. L. de Almeida, and D. C. Araújo, “Channel estimation for flexible intelligent metasurface aided MIMO communications,” arXiv:2605.29227, 2026.arxiv.org/abs/2605.29227
  7. [7]X. Han, X. Zhao, J. Ying, and F. Gao, “Tensor-based information monitoring receiver in UAV-aided MIMO communication systems,” IEEE Wireless Commun. Lett., vol. 11, no. 1, pp. 155–159, 2022.DOI
  8. [9]G. T. de Araújo, A. L. F. de Almeida and R. Boyer, “Channel Estimation for Intelligent Reflecting Surface Assisted MIMO Systems: A Tensor Modeling Approach,” IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 3, pp. 789–802, 2021.DOI
  9. [10]K. Ardah, S. Geheshkho, A. L. F. de Almeida, and M. Haardt, “TRICE: A channel estimation framework for RIS-Aided millimeter-wave MIMO systems,” IEEE Signal Process. Lett., vol. 28, pp. 513–517, 2021.DOI
  10. [11]Y. Cao, C. Xing, N. Wei, S. Gong, N. Zhao, and D. Niyato, “Tensor-based joint channel estimation and activity detection for reconfigurable intelligent surface-assisted massive connectivity,” IEEE Trans. Commun., vol. 73, no. 11, pp. 11 870–11 885, 2025.DOI
  11. [12]L. Wei, C. Huang, G. C. Alexandropoulos, C. Yuen, Z. Zhang, and M. Debbah, “Channel estimation for RIS-empowered multi-user MISO wireless communications,” IEEE Trans. Commun., vol. 69, no. 6, pp. 4144–4157, 2021.DOI
  12. [13]X. Zheng, P. Wang, J. Fang, and H. Li, “Compressed channel estimation for IRS-assisted millimeter wave OFDM systems: A low-rank tensor decomposition-based approach,” IEEE Wireless Commun. Lett., vol. 11, no. 6, pp. 1258–1262, 2022.arxiv.org/abs/2203.16164
  13. [14]X. Han, P. Li, J. Ying, F. Gao, and C. Yuen, “Tensor-based joint channel and symbol estimation with subspace-based parameter extraction for multi-RIS uplink MIMO systems,” IEEE Trans. Commun., vol. 74, pp. 7271–7284, 2026.DOI
  14. [15]A. L. F. de Almeida, B. Sokal, H. Li, and B. Clerckx, “Channel estimation for beyond diagonal RIS via tensor decomposition,” IEEE Trans. Signal Process., vol. 73, pp. 4744–4779, 2025.

Paper details

vlnc:W_tchrrw9j
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents