betaFeedback

WorksEngineering

Channel Estimation for Stacked Intelligent Metasurfaces with Mutual Coupling

Abstract

Channel estimation for stacked intelligent metasurfaces (SIMs) has relied on the cascade model, which ignores mutual coupling and inter-layer reflection. Under a coupling-consistent multiport model, we show that the pilots are linear in the unknown channel segments and derive least-squares and linear minimum-mean-square-error estimators, an identifiability condition and an exact error decomposition for the reactance training schedule, and a gradient-based schedule design. Cascade-based estimation floors above 00~dB normalized error at practical layer spacings and collapses the sum rate below that of the bare feed, whereas the proposed estimators retain about 80%80\% of the perfect-channel-knowledge sum rate.

Introduction

Stacked intelligent metasurfaces (SIMs) cascade several programmable transmissive layers in front of a compact feed array, so that precoding and combining are performed by the propagating field itself, in the wave domain, before digitization [1], [2]. Since the number of meta-atoms per layer far exceeds the number of radio-frequency chains, acquiring channel state information (CSI) is the bottleneck of every SIM-aided design, and a growing body of work addresses it: subspace-based estimators with optimized training phase shifts [3], minimum mean square error (MMSE) estimation under Rician fading [4], hybrid digital–wave-domain estimation [5], nested-tensor [6] and parity-dependent tensor [7] decompositions, and learning-based refinements surveyed in [2].

Nearly all of these schemes adopt the cascade model: each layer is a diagonal mask of unit-modulus phase shifts, and consecutive layers are connected by a scalar, unilateral Rayleigh–Sommerfeld (RS) propagator [1]. A parallel, circuit-theoretic literature has shown that this abstraction is not the physics of a compact stack. Physically consistent scattering- and impedance-parameter models of SIMs [8], [9] reveal intra-layer mutual coupling, inter-layer reflections, and a load-dependent input impedance seen by the feed. The model of [10] captures all of these in one closed-form impedance kernel sampled at the three-dimensional (3-D) port separations, shows that the cascade is the single-pass limit of one matrix inversion under assumptions that fail in compact stacks, and finds that a precoder designed on the cascade can perform worse than open-circuiting the stack. The only coupling-aware SIM

The only coupling-aware SIM estimator is available in [11], which configures a receive SIM as a single analog projector onto a subspace derived from prior location knowledge, for near-field channel and localization estimation. Training over multiple load states, identifiability without prior knowledge, and the error committed by cascade-based estimators remain open [2], and are addressed here.

Contributions: (i) we formulate uplink SIM channel estimation under a coupling-consistent multiport model and show that, although the loaded stack responds nonlinearly to its reactances, the received pilots are linear in the unknown feed-side and stack-side channel segments, with a known, load-dependent regressor; the model is exact in the radiating near field, so no planar-wavefront assumption is needed; (ii) we derive least-squares (LS) and linear MMSE (LMMSE) estimators over a schedule of reactance states, a necessary and sufficient identifiability condition on the schedule, the exact mean square error (MSE) and its feed/stack decomposition, the exact prediction error at untrained load states, the high-SNR floor of the LMMSE estimator below the pilot counting bound, and a gradient-based schedule design; (iii) we derive the effective-channel normalized MSE (NMSE) of every linear estimator in closed form, including the bias floor of cascade-based estimation and of model errors in general; and (iv) we show that this bias is minor below 0.1λ0.1\lambda of layer spacing, where the stack is nearly inert, but reaches about 100% NMSE or more from λ/6\lambda/6 to λ/2\lambda/2, a range covering the spacings at which a compact SIM delivers its gains [10]; that cascade-based CSI collapses the sum rate of load optimization below that of the bare feed; and that the proposed estimators retain about 80% of the perfect-CSI sum rate and tolerate realistic fabrication errors, 2–3-bit reactances, and multipath.

Notation: boldface lowercase (uppercase) letters are vectors (matrices); (⋅)T(\cdot)^{\mathsf{T}}, (⋅)H(\cdot)^{\mathsf{H}}, (⋅)∗(\cdot)^{*} denote transpose, conjugate transpose, and conjugate; Re⁡{⋅}\operatorname{Re}\{\cdot\} and Im⁡{⋅}\operatorname{Im}\{\cdot\} act entrywise; diag⁡(⋅)\operatorname{diag}(\cdot) maps a vector to a diagonal matrix and a matrix to its diagonal; jj is the imaginary unit with convention e−j2πfte^{-j2\pi ft}; jℓj_{\ell}, yℓy_{\ell}, hℓ=jℓ+jyℓh_{\ell}=j_{\ell}+jy_{\ell} are spherical Bessel and Hankel functions; ∥⋅∥2\lVert\cdot\rVert_{2} and ∥⋅∥F\lVert\cdot\rVert_{\mathrm{F}} are the spectral and Frobenius norms, σmin⁡(⋅)\sigma_{\min}(\cdot) is the smallest singular value, (⋅)+(\cdot)^{+} the pseudoinverse, and μi(⋅)\mu_{i}(\cdot) the ii-th largest eigenvalue; ene_{n} is the nn-th canonical vector; λ\lambda is the wavelength and κ=2π/λ\kappa=2\pi/\lambda.

Coupling-Consistent System Model

We adopt the multiport model of [10], specialized to fixed port positions. A base station (BS) hosts N0N_{0} driven ports on the plane z=0z=0 and an LL-layer SIM whose ll-th sheet carries NlN_{l} identical xx-polarized sub-wavelength elements on the plane z=zlz = z_l, 0<z1<⋯<zL0 < z_1 < \cdots< z_L, with Ns=∑l≥1NlN_{\mathrm{s}} = \sum_{l \ge1} N_l stack ports in total. The mutual impedance between any two ports, on the same sheet or across sheets, is the free-space dyadic Green function projected on the polarization and sampled at their 3-D separation τ\boldsymbol{\tau}; in the point-dipole limit and normalized by the radiation resistance RrR_{\mathrm{r}}, it reads

ϑ(ρ,c)=32[h0(ρ)−h1(ρ)ρ+h2(ρ)c2],(1)\vartheta(\rho,c) = \frac{3}{2}\left[h_0(\rho)-\frac{h_1(\rho)}{\rho}+h_2(\rho)c^2\right], \tag*{(1)}

with ρ=κ∥τ∥\rho= \kappa\lVert\boldsymbol{\tau}\rVert and cc the cosine between the polarization and τ\boldsymbol{\tau}. Finite element profiles smooth (1) at gaps below ∼0.1λ\sim0.1\lambda [10]; nothing below depends on its form beyond the structure of the multiport impedance matrix

Z=(1+ϵ+jζA)IN+Θ,N=N0+Ns,(2)\mathbf{Z} = (1+\epsilon+\mathrm{j}\zeta_{\mathrm{A}})\mathbf{I}_N+\mathbf{\Theta}, \qquad N=N_0+N_{\mathrm{s}}, \tag*{(2)}

where ϵ\epsilon and ζA\zeta_{\mathrm{A}} are the normalized ohmic self-resistance and self-reactance and Θ\mathbf{\Theta} collects (1) off the diagonal. With port currents ι=[ι0T,ιsT]T\boldsymbol{\iota}=[\boldsymbol{\iota}_0^{\mathsf{T}},\boldsymbol{\iota}_{\mathrm{s}}^{\mathsf{T}}]^{\mathsf{T}} and voltages v=Zι\mathbf{v}=\mathbf{Z}\boldsymbol{\iota} partitioned into feed (subscript 0) and stack (subscript s) blocks, and the stack block into sheet blocks Zll′\mathbf{Z}_{ll'}, reciprocity gives Zs0=Z0sT\mathbf{Z}_{\mathrm{s}0}=\mathbf{Z}_{0\mathrm{s}}^{\mathsf{T}} and Zll′=Zl′lT\mathbf{Z}_{ll'}=\mathbf{Z}_{l'l}^{\mathsf{T}}: the backward transfer between sheets equals the forward one, which is precisely what a unilateral propagator discards.

Each stack port nn is terminated on a tunable reactance xnx_n, i.e., on the load network M=jdiag⁡(x)\mathbf{M}=\mathrm{j}\operatorname{diag}(\mathbf{x}), x=[x1,…,xNs]T∈RNs\mathbf{x}=[x_1,\ldots,x_{N_{\mathrm{s}}}]^{\mathsf{T}}\in\mathbb{R}^{N_{\mathrm{s}}}. Substituting the terminal law vs=−Mιs\mathbf{v}_{\mathrm{s}}=-\mathbf{M}\boldsymbol{\iota}_{\mathrm{s}} into the multiport relation vs=Zs0ι0+Zssιs\mathbf{v}_{\mathrm{s}}=\mathbf{Z}_{\mathrm{s}0}\boldsymbol{\iota}_0+\mathbf{Z}_{\mathrm{ss}}\boldsymbol{\iota}_{\mathrm{s}} gives the induced currents

ιs=−B(x)ι0,B(x)≜(Zss+jdiag⁡(x))−1Zs0,(3)\boldsymbol{\iota}_{\mathrm{s}}=-\mathbf{B}(\mathbf{x})\boldsymbol{\iota}_0,\qquad\mathbf{B}(\mathbf{x})\triangleq\left(\mathbf{Z}_{\mathrm{ss}}+\mathrm{j}\operatorname{diag}(\mathbf{x})\right)^{-1}\mathbf{Z}_{\mathrm{s}0}, \tag*{(3)}

which contain every multiple reflection, intra-layer coupling, and feedback term. Since Re⁡{Zss}⪰ϵI\operatorname{Re}\{\mathbf{Z}_{\mathrm{ss}}\}\succeq\epsilon\mathbf{I} by passivity, (3) is well posed for every x\mathbf{x} when ϵ>0\epsilon>0.

A single-antenna user kk, modeled as an xx-polarized dipole at rk\mathbf{r}_k, is one more port of the same multiport: its couplings to the feed and stack ports are the kernel samples [gk,0]n=ϑ(κ∥rk−r0,n∥,c)[\mathbf{g}_{k,0}]_n=\vartheta(\kappa\lVert\mathbf{r}_k-\mathbf{r}_{0,n}\rVert,c) and [gk,s]n=ϑ(κ∥rk−rs,n∥,c)[\mathbf{g}_{k,\mathrm{s}}]_n=\vartheta(\kappa\lVert\mathbf{r}_k-\mathbf{r}_{\mathrm{s},n}\rVert,c), with r0,n,rs,n\mathbf{r}_{0,n},\mathbf{r}_{\mathrm{s},n} the port positions and cc the polarization cosine of each pair, as in (1). This is a near-field line-of-sight channel; with scatterers, gk,0\mathbf{g}_{k,0} and gk,s\mathbf{g}_{k,\mathrm{s}} become the user’s field at the ports and nothing below changes. As the user’s open-circuit voltage is gk,0Tι0+gk,sTιs\mathbf{g}_{k,0}^{\mathsf{T}}\boldsymbol{\iota}_0+\mathbf{g}_{k,\mathrm{s}}^{\mathsf{T}}\boldsymbol{\iota}_{\mathrm{s}}, (3) yields the effective end-to-end channel, i.e., the user–feed entry of the Schur complement of the loaded multiport with the stack ports eliminated,

g~kT(x)=gk,0T−gk,sTB(x),(4)\widetilde{g}_k^{\mathsf{T}}(\mathbf{x})=\mathbf{g}_{k,0}^{\mathsf{T}}-\mathbf{g}_{k,\mathrm{s}}^{\mathsf{T}}\mathbf{B}(\mathbf{x}), \tag*{(4)}

which, by reciprocity, is also the uplink channel from user kk to the open-circuit port voltages of the BS. Since Zss\mathbf{Z}_{\mathrm{ss}} and Zs0\mathbf{Z}_{\mathrm{s}0} depend only on the known geometry, B(x)\mathbf{B}(\mathbf{x}) is computable at the BS for any load state, and (4) is linear in the only unknowns, gk,0\mathbf{g}_{k,0} and gk,s\mathbf{g}_{k,\mathrm{s}}, with a regressor that the BS steers through x\mathbf{x}.

Remark 1: Once gk,0\mathbf{g}_{k,0} and gk,s\mathbf{g}_{k,\mathrm{s}} are estimated, (4) predicts the effective channel at every load state, which is what a load-optimization loop needs (Corollary 1, Sec. V-C), whereas end-to-end estimates [5] hold only at the training state. The price of physical consistency is the size of the unknown: every sheet radiates toward the user, so the channel is needed at all N0+NsN_0+N_{\mathrm{s}} ports, rather than at the outer layer only as in a cascade with known propagators.

Training Protocol

Users transmit mutually orthogonal pilots over TT slots of a coherence block; after pilot matching, the per-user observation in slot tt, during which the loads are set to the schedule entry xt\mathbf{x}_t, is

yt=g~k(xt)p+nt=(gk,0−BtTgk,s)p+nt,(5)\mathbf{y}_t=\widetilde{\mathbf{g}}_k(\mathbf{x}_t)\sqrt{p}+\mathbf{n}_t=(\mathbf{g}_{k,0}-\mathbf{B}_t^{\mathsf{T}}\mathbf{g}_{k,\mathrm{s}})\sqrt{p}+\mathbf{n}_t, \tag*{(5)}

with Bt≜B(xt)\mathbf{B}_t\triangleq\mathbf{B}(\mathbf{x}_t), pilot power pp (set to 1 henceforth), and nt∼CN(0,σ2IN0)\mathbf{n}_t\sim\mathcal{CN}(\mathbf{0},\sigma^2\mathbf{I}_{N_0}). Stacking the slots and defining θ≜[gk,0T,gk,sT]T∈CN\boldsymbol{\theta}\triangleq[\mathbf{g}_{k,0}^{\mathsf{T}},\mathbf{g}_{k,\mathrm{s}}^{\mathsf{T}}]^{\mathsf{T}}\in\mathbb{C}^{N},

y=Aθ+n,A=[IN0−B1T⋮⋮IN0−BTT]∈CTN0×N.(6)\mathbf{y}=\mathbf{A}\boldsymbol{\theta}+\mathbf{n},\qquad \mathbf{A}= \begin{bmatrix} \mathbf{I}_{N_0} & -\mathbf{B}_1^{\mathsf{T}}\\ \vdots& \vdots\\ \mathbf{I}_{N_0} & -\mathbf{B}_T^{\mathsf{T}} \end{bmatrix}\in\mathbb{C}^{TN_0\times N}. \tag*{(6)}

The training states are reactance vectors, not phase masks: the hardware knob is x\mathbf{x}, to which the coupled stack responds only through B(x)\mathbf{B}(\mathbf{x}). For intuition, consider an isolated atom (coupling neglected): one port of normalized self-impedance a+jζAa+\mathrm{j}\zeta_{\mathrm{A}}, a≜1+ϵa\triangleq1+\epsilon, terminated on one lossless reactance xnx_n. Its induced current per unit incident voltage, ψ(xn)≜−1/(a+j(ζA+xn))\psi(x_n)\triangleq-1/(a+\mathrm{j}(\zeta_{\mathrm{A}}+x_n)), traces a circle through the origin as xnx_n sweeps R\mathbb{R}, locking amplitude and phase, so no lossless one-port load realizes the unit-modulus phase shifts assumed in [3, 4, 5]. We draw training reactances as

xt,n=−ζA+atan⁡ut,n,ut,n∼U(−umax⁡,umax⁡).(7)x_{t,n}=-\zeta_{\mathrm{A}}+a\tan u_{t,n},\qquad u_{t,n}\sim\mathcal{U}(-u_{\max},u_{\max}). \tag*{(7)}

Under (7) the isolated response is −(cos⁡ut,n/a)e−jut,n-(\cos u_{t,n}/a)e^{-\mathrm{j}u_{t,n}}, whose phase is thus sampled uniformly about the element resonance u=0u=0, whereas uniformly drawn reactances would crowd near the open circuit, where atoms are nearly inert. In the coupled stack, (7) is merely a convenient prior over x\mathbf{x}: by Proposition 1 below, it yields identifiability almost surely, as would any absolutely continuous distribution, and Sec. III-C optimizes the schedule.

Estimators, Identifiability, and Schedule Design

Least Squares and Identifiability

When A\mathbf{A} has full column rank, the LS estimate and its error covariance are

θ^LS=(AHA)−1AHy,MSE=σ2tr⁡[(AHA)−1],(8)\widehat{\boldsymbol{\theta}}_{\mathrm{LS}}=(\mathbf{A}^{\mathsf{H}}\mathbf{A})^{-1}\mathbf{A}^{\mathsf{H}}\mathbf{y},\qquad\mathrm{MSE}=\sigma^2\operatorname{tr}\left[(\mathbf{A}^{\mathsf{H}}\mathbf{A})^{-1}\right], \tag*{(8)}

which is unbiased and attains the Cramér–Rao bound (CRB) for deterministic θ\boldsymbol{\theta} [12].

Proposition 1 (Identifiability). A\mathbf{A} in (6) has full column rank N0+NsN_0+N_{\mathrm{s}} if and only if

rank⁡[B2−B1,B3−B1,…,BT−B1]=Ns,(9)\operatorname{rank}\left[\mathbf{B}_2-\mathbf{B}_1,\mathbf{B}_3-\mathbf{B}_1,\ldots,\mathbf{B}_T-\mathbf{B}_1\right]=N_{\mathrm{s}}, \tag*{(9)}

so that (T−1)N0≥Ns(T-1)N_0\geq N_{\mathrm{s}}, i.e., T≥1+⌈Ns/N0⌉T\geq1+\lceil N_{\mathrm{s}}/N_0\rceil is necessary. Moreover, if one schedule of length TT satisfies (9), then almost every schedule drawn from (7), or from any absolutely continuous distribution, does.

Proof: Aθ=0\mathbf{A}\boldsymbol{\theta}=0 iff gk,0=BtTgk,s\mathbf{g}_{k,0}=\mathbf{B}_{t}^{\mathsf{T}}\mathbf{g}_{k,s} for all tt, i.e., iff (Bt−B1)Tgk,s=0(\mathbf{B}_{t}-\mathbf{B}_{1})^{\mathsf{T}}\mathbf{g}_{k,s}=0 for all tt and gk,0=B1Tgk,s\mathbf{g}_{k,0}=\mathbf{B}_{1}^{\mathsf{T}}\mathbf{g}_{k,s}; the null space is trivial iff the stacked differences have rank NsN_s, which bounds TT by counting. For genericity, every entry of B(x)\mathbf{B}(\mathbf{x}) is a rational function of x\mathbf{x} by (3) and Cramer’s rule, so the minors of the matrix in (9) are rational in (x1,…,xT)(x_1,\ldots,x_T), and the rank condition fails only on the common zero set of all top-order minors; these are real-analytic and, once one full-rank schedule exists, not all identically zero, so that set has Lebesgue measure zero. In the setting of Sec. V, such a schedule exists already at the counting bound T=14T=14, for every spacing considered. ■

Proposition 2 (MSE decomposition): Let B‾=1T∑tBt\overline{\mathbf{B}}=\frac{1}{T}\sum_t\mathbf{B}_t, Δt=Bt−B‾\boldsymbol{\Delta}_t=\mathbf{B}_t-\overline{\mathbf{B}}, and S=∑tΔt∗Δt\mathbf{S}=\sum_t\boldsymbol{\Delta}_t^{*}\boldsymbol{\Delta}_t. Under (9), the LS errors of the stack and feed segments are

MSE⁡s=σ2tr⁡(S−1),MSE⁡0=σ2[N0T+tr⁡(B‾S−1B‾∗)].(10)\operatorname{MSE}_{s}=\sigma^{2}\operatorname{tr}(\mathbf{S}^{-1}),\qquad\operatorname{MSE}_{0}=\sigma^{2}\left[\frac{N_{0}}{T}+\operatorname{tr}\left(\overline{\mathbf{B}}\mathbf{S}^{-1}\overline{\mathbf{B}}^{*}\right)\right]. \tag*{(10)}

Proof: AHA=[TI−TB‾−TB‾∗∑tBt∗Bt]\mathbf{A}^{\mathsf{H}}\mathbf{A}=\begin{bmatrix}T\mathbf{I}&-T\overline{\mathbf{B}}\\-T\overline{\mathbf{B}}^{*}&\sum_t\mathbf{B}_t^{*}\mathbf{B}_t\end{bmatrix}; its leading block has Schur complement S\mathbf{S}, and block inversion gives

(AHA)−1=[T−1I+B‾S−1B‾∗B‾S−1S−1B‾∗S−1],(11)(\mathbf{A}^{\mathsf{H}}\mathbf{A})^{-1}= \begin{bmatrix} T^{-1}\mathbf{I}+\overline{\mathbf{B}}\mathbf{S}^{-1}\overline{\mathbf{B}}^{*}&\overline{\mathbf{B}}\mathbf{S}^{-1}\\ \mathbf{S}^{-1}\overline{\mathbf{B}}^{*}&\mathbf{S}^{-1} \end{bmatrix}, \tag*{(11)}

whose diagonal blocks give (10). ■

The stack segment is thus learned only from the variation of the regressor across load states, S≻0\mathbf{S}\succ0 being exactly (9); in the setting of Sec. V at inter-layer spacing d=λ/4d=\lambda/4, it carries about 99% of the MSE, six to seven times the feed error per entry.

Corollary 1 (Prediction at any load state): The exact-model LS prediction C(x)θ^LS\mathbf{C}(\mathbf{x})\hat{\boldsymbol{\theta}}_{\mathrm{LS}} of the effective channel (4), with C(x)=[IN0,−BT(x)]\mathbf{C}(\mathbf{x})=[\mathbf{I}_{N_0},-\mathbf{B}^{\mathsf{T}}(\mathbf{x})], is unbiased at every state x\mathbf{x}, trained or not, with MSE

σ2[N0T+tr⁡((B(x)−B‾)S−1(B(x)−B‾)∗)].(12)\sigma^{2}\left[\frac{N_{0}}{T}+\operatorname{tr}\left((\mathbf{B}(\mathbf{x})-\overline{\mathbf{B}})\mathbf{S}^{-1}(\mathbf{B}(\mathbf{x})-\overline{\mathbf{B}})^{*}\right)\right]. \tag*{(12)}

Proof: The error covariance is σ2C(x)(AHA)−1CH(x)\sigma^{2}\mathbf{C}(\mathbf{x})(\mathbf{A}^{\mathsf{H}}\mathbf{A})^{-1}\mathbf{C}^{\mathsf{H}}(\mathbf{x}); inserting (11) and collecting terms gives (12). ■

The prediction error grows with the S−1\mathbf{S}^{-1}-weighted distance of B(x)\mathbf{B}(\mathbf{x}) from the training mean, and B=0\mathbf{B}=0 recovers MSE⁡0\operatorname{MSE}_{0}: a schedule should span the load region in which the optimizer of Sec. V-C operates. This makes Remark 1 quantitative.

With KK users and orthogonal pilots, all users are estimated in the same TT load states at KK-fold pilot length, and the per-user problems decouple. Online, both estimators cost one N×TN0N\times TN_0 matrix–vector product per user and coherence block, their matrices being precomputed per geometry and schedule (and noise level, for LMMSE) from TT solves of size NsN_s and one inversion of size NN (LS) or TN0TN_0 (LMMSE). Since (8) attains the CRB among unbiased estimators, the high-SNR floors of the mismatched schemes in Sec. V are due to bias, not to estimator suboptimality; at low SNR, a biased scheme may fall below the bound.

LMMSE with Statistical CSI

Users confined to a known coverage sector induce the prior correlation Cθ=E[θθH]\mathbf{C}_{\theta}=\mathbb{E}[\boldsymbol{\theta}\boldsymbol{\theta}^{\mathsf{H}}], computable offline from (1) by Monte Carlo over the sector. The LMMSE estimate

θ^LMMSE=CθAH(ACθAH+σ2ITN0)−1y.(13)\hat{\boldsymbol{\theta}}_{\mathrm{LMMSE}}=\mathbf{C}_{\theta}\mathbf{A}^{\mathsf{H}}\left(\mathbf{A}\mathbf{C}_{\theta}\mathbf{A}^{\mathsf{H}}+\sigma^{2}\mathbf{I}_{TN_0}\right)^{-1}\mathbf{y}. \tag*{(13)}

Proposition 3 (LMMSE below the counting bound): Let K=ACθ1/2\mathbf{K}=\mathbf{A}\mathbf{C}_{\theta}^{1/2}. The LMMSE error covariance E(σ2)=CθAH(ACθAH+σ2I)−1ACθ\mathbf{E}(\sigma^{2})=\mathbf{C}_{\theta}\mathbf{A}^{\mathsf{H}}(\mathbf{A}\mathbf{C}_{\theta}\mathbf{A}^{\mathsf{H}}+\sigma^{2}\mathbf{I})^{-1}\mathbf{A}\mathbf{C}_{\theta} decreases, in the Loewner order, as σ2↓0\sigma^{2}\downarrow0 to

E∞=Cθ1/2(IN−K+K)Cθ1/2,(14)\mathbf{E}_{\infty}=\mathbf{C}_{\theta}^{1/2}(\mathbf{I}_{N}-\mathbf{K}^{+}\mathbf{K})\mathbf{C}_{\theta}^{1/2}, \tag*{(14)}

which vanishes iff A\mathbf{A} is injective on the range of Cθ\mathbf{C}_{\theta}, so that TN0≥rank⁡CθTN_0\geq\operatorname{rank}\mathbf{C}_{\theta} is necessary; moreover, tr⁡E∞≥∑i>TN0μi(Cθ)\operatorname{tr}\mathbf{E}_{\infty}\geq\sum_{i>TN_0}\mu_i(\mathbf{C}_{\theta}).

Proof: E(σ2)=Cθ1/2[I−KH(KKH+σ2I)−1K]Cθ1/2\mathbf{E}(\sigma^{2})=\mathbf{C}_{\theta}^{1/2}[\mathbf{I}-\mathbf{K}^{\mathsf{H}}(\mathbf{K}\mathbf{K}^{\mathsf{H}}+\sigma^{2}\mathbf{I})^{-1}\mathbf{K}]\mathbf{C}_{\theta}^{1/2}, and the bracket decreases to I−K+K\mathbf{I}-\mathbf{K}^{+}\mathbf{K}, the projector onto the null space of K\mathbf{K}, which annihilates Cθ1/2\mathbf{C}_{\theta}^{1/2} iff rank⁡K=rank⁡Cθ\operatorname{rank}\mathbf{K}=\operatorname{rank}\mathbf{C}_{\theta}. Since tr⁡E∞=tr⁡Cθ−tr⁡(K+KCθ)\operatorname{tr}\mathbf{E}_{\infty}=\operatorname{tr}\mathbf{C}_{\theta}-\operatorname{tr}(\mathbf{K}^{+}\mathbf{K}\mathbf{C}_{\theta}) and K+K\mathbf{K}^{+}\mathbf{K} is a projector of rank at most TN0TN_0, Ky Fan’s maximum principle gives the bound. ■

The overhead is thus governed by the effective rank of the prior, not by NsN_s: in our setting 99% of the prior energy lies in one of N=54N=54 eigendirections, and the normalized floor tr⁡E∞/tr⁡Cθ\operatorname{tr}\mathbf{E}_{\infty}/\operatorname{tr}\mathbf{C}_{\theta} is −9.8-9.8, −25.7-25.7, and −54.2-54.2 dB for T=2T=2, 44, and 88, against T=14T=14 for LS.

Gradient-Based Schedule Design

The MSE in (8) depends on the schedule through Bt\mathbf{B}_t. Differentiating T(Zss+jdiag⁡(x))=I\mathbf{T}(\mathbf{Z}_{\mathrm{ss}}+j\operatorname{diag}(\mathbf{x}))=\mathbf{I} gives ∂T/∂xn=−jTenenTT\partial\mathbf{T}/\partial x_n=-j\mathbf{T}\mathbf{e}_n\mathbf{e}_n^{\mathsf{T}}\mathbf{T}, hence

∂B∂xn=−jTenenTB,T=(Zss+jdiag⁡(x))−1.(15)\frac{\partial\mathbf{B}}{\partial x_n}=-j\mathbf{T}\mathbf{e}_n\mathbf{e}_n^{\mathsf{T}}\mathbf{B},\qquad\mathbf{T}=(\mathbf{Z}_{\mathrm{ss}}+j\operatorname{diag}(\mathbf{x}))^{-1}. \tag*{(15)}

Let G=AHA\mathbf{G}=\mathbf{A}^{\mathsf{H}}\mathbf{A} and f=tr⁡(G−1)f=\operatorname{tr}(\mathbf{G}^{-1}). Since df=−2Re⁡tr⁡(G−2dAdA)df=-2\operatorname{Re}\operatorname{tr}(\mathbf{G}^{-2}d\mathbf{A}d\mathbf{A}), (15) gives the gradient with respect to the reactances of state tt in closed form,

∇xtf=2Im⁡{diag⁡(TtVtBtT)},(16)\nabla_{\mathbf{x}_t}f=2\operatorname{Im}\{\operatorname{diag}(\mathbf{T}_t\mathbf{V}_t\mathbf{B}_t^{\mathsf{T}})\}, \tag*{(16)}

where Tt=T(xt)\mathbf{T}_t=\mathbf{T}(\mathbf{x}_t), Vt\mathbf{V}_t is the block of G−2AH\mathbf{G}^{-2}\mathbf{A}^{\mathsf{H}} with stack rows and slot-tt columns, and the symmetry Tt=TtT\mathbf{T}_t=\mathbf{T}_t^{\mathsf{T}}, inherited from reciprocity, has been used. All NsN_s partials of a state thus cost one Ns×NsN_s\times N_s inversion. We minimize ff in the variables ut,nu_{t,n} of (7), box-constrained to [−umax⁡,umax⁡][-u_{\max},u_{\max}], by projected gradient descent with Armijo backtracking from a random draw of (7). The problem is non-convex. Being rational in x\mathbf{x}, but independent random initializations converge to A-optimality values within 0.2 dB of each other, in about 6 s for T=16T=16, Ns=50N_s=50. By Proposition 2, what matters is how evenly the regressor variation covers all NsN_s directions, not its size: at d=λ/4d=\lambda/4 the optimizer lowers the condition number of S\mathbf{S} from 1.2⋅1041.2\cdot10^{4} to 50, while tr⁡S\operatorname{tr}\mathbf{S} even drops from 28.3 to 25.3. The design is performed offline once per geometry, and every optimized state is realizable by construction.

The Cost of the Cascade Model

Suppose the estimator relies on the cascade model, as in prior SIM estimation work. In the multiport language, the cascade keeps in (3) only the forward nearest-neighbor blocks: sheet ll is excited by sheet l−1l-1 alone through Zl,l−1\mathbf{Z}_{l,l-1} (sheet 1 by the feed through Zl0\mathbf{Z}_{l0}), while the backward blocks Zl,l+1\mathbf{Z}_{l,l+1} and the direct blocks Zl0\mathbf{Z}_{l0}, l≥2l\geq2, are discarded. The currents ιl\iota_l of sheet ll then obey ιl=ΨlZl,l−1ιl−1\iota_l = \Psi_l Z_{l,l-1}\iota_{l-1}, Ψl=−(Zll+ȷdiag⁡(x(l)))−1\Psi_l = -(Z_{ll} + \jmath\operatorname{diag}(x^{(l)}))^{-1}, with x(l)x^{(l)} the reactances of sheet ll, yielding a mismatched regressor B˘t\breve{\mathbf{B}}_t and matrix A˘\breve{\mathbf{A}}. The phase-mask cascade additionally neglects intra-layer coupling and replaces Ψl\Psi_l by diag⁡(eȷarg⁡ψ(xn(l)))\operatorname{diag}(e^{\jmath\arg\psi(x_n^{(l)})}), with ψ\psi the isolated response of § 2-A: this is the model under which the training designs of [3], [4], [6], [7] optimize. Both variants keep the exact kernel blocks as propagators and, unlike the cascade of Remark 1, estimate the same θ∈CN\theta\in\mathbb{C}^{N}, so they are credited with the correct polarized transfer and with the user’s coupling to every port: their errors are purely structural. The mismatched LS estimate under the true observations (6) is

θ˘=(A˘HA˘)−1A˘H(Aθ+n),(17)\breve{\theta} = (\breve{\mathbf{A}}^{\mathrm{H}}\breve{\mathbf{A}})^{-1}\breve{\mathbf{A}}^{\mathrm{H}}(\mathbf{A}\theta+ \mathbf{n}), \tag*{(17)}

with bias β=A˘+(A−A˘)θ\beta= \breve{\mathbf{A}}^{+}(\mathbf{A} - \breve{\mathbf{A}})\theta, A˘+≔(A˘HA˘)−1A˘H\breve{\mathbf{A}}^{+} \coloneqq(\breve{\mathbf{A}}^{\mathrm{H}}\breve{\mathbf{A}})^{-1}\breve{\mathbf{A}}^{\mathrm{H}}, independent of the noise: both θ˘\breve{\theta} and the effective channel it reconstructs through its own B˘(x)\breve{\mathbf{B}}(x) carry an NMSE floor at every SNR. The same expression governs any model error, e.g., fabrication tolerances or misestimated ε\varepsilon and ζA\zeta_A, and, for full-column-rank A˘\breve{\mathbf{A}}, bounds the parameter floor as

∥β∥∥θ∥≤∥A−A˘∥2σmin⁡(A˘),(18)\frac{\|\beta\|}{\|\theta\|} \le\frac{\|\mathbf{A} - \breve{\mathbf{A}}\|_2}{\sigma_{\min}(\breve{\mathbf{A}})}, \tag*{(18)}

which favors well-conditioned schedules such as those of Sec. III-C; what sets the cascade apart from such errors is the size of A−A˘\mathbf{A} - \breve{\mathbf{A}}. Both discarded families contribute to it: in the setting of Sec. V at d=λ/4d = \lambda/4, restoring only the direct block Z20\mathbf{Z}_{20} lowers the noiseless floor from +2.7+2.7 to −2.4-2.4 dB, and restoring only the backward block Z12\mathbf{Z}_{12} to +0.7+0.7 dB, so neither truncation alone is benign.

Effective-channel NMSE in closed form: Every estimator above is linear, θ^=Wy\hat{\theta} = \mathbf{W}\mathbf{y}, and predicts the effective channel at a load state xvx_v through its own map C^v=[IN0,−B^T(xv)]\hat{\mathbf{C}}_v = [\mathbf{I}_{N_0},-\hat{\mathbf{B}}^{\mathrm{T}}(x_v)], whereas the truth is Cvθ\mathbf{C}_v\theta with Cv=C(xv)\mathbf{C}_v = \mathbf{C}(x_v). Averaging over the noise and over users with correlation Cθ\mathbf{C}_{\theta}, the NMSE over a set of states {xv}\{x_v\} is

NMSE⁡=∑v[tr⁡(DvCθDvH)+σ2∥C^vW∥F2]∑vtr⁡(CvCθCvH),(19)\operatorname{NMSE} = \frac{\sum_v \left[\operatorname{tr}(\mathbf{D}_v\mathbf{C}_{\theta}\mathbf{D}_v^{\mathrm{H}}) + \sigma^2\|\hat{\mathbf{C}}_v\mathbf{W}\|_{\mathrm{F}}^2\right]}{\sum_v \operatorname{tr}(\mathbf{C}_v\mathbf{C}_{\theta}\mathbf{C}_v^{\mathrm{H}})}, \tag*{(19)}

with Dv=C^vWA−Cv\mathbf{D}_v = \hat{\mathbf{C}}_v\mathbf{W}\mathbf{A} - \mathbf{C}_v. For the exact LS, W=A+\mathbf{W} = \mathbf{A}^{+} and Dv=0\mathbf{D}_v = 0, so (19) decays as σ2\sigma^2, consistently with Corollary 1; for the LMMSE, W=CθAH(ACθAH+σ2I)−1\mathbf{W} = \mathbf{C}_{\theta}\mathbf{A}^{\mathrm{H}}(\mathbf{A}\mathbf{C}_{\theta}\mathbf{A}^{\mathrm{H}} + \sigma^2\mathbf{I})^{-1}; for the cascades, W=A˘+\mathbf{W} = \breve{\mathbf{A}}^{+} and B^=B˘\hat{\mathbf{B}} = \breve{\mathbf{B}}, and the first term is the noise-independent floor of the effective channel itself.

Numerical Results

We follow the scenario of [10]: N0=4N_0 = 4 feed ports on a central λ/2\lambda/2 square at z=0z = 0; L=2L = 2 sheets of 5×55 \times5 atoms on a λ/2\lambda/2 grid (Ns=50N_s = 50), the first at standoff z1=λ/2z_1 = \lambda/2 and the second dd beyond it; ε=0.05\varepsilon= 0.05, ζA=0.58\zeta_A = 0.58. Users are dropped uniformly in a ±4λ\pm4\lambda lateral box, 8λ8\lambda–14λ14\lambda beyond the last sheet, within the Fraunhofer distance (≥16λ\ge16\lambda) of the stack; θ\theta is normalized to ∥θ∥2=N\|\theta\|^2 = N, and SNR =1/σ2= 1/\sigma^2. Training uses T=16T = 16 (counting bound T=14T = 14) and umax⁡=1.35u_{\max} = 1.35; the LMMSE prior is estimated from 3000 kernel drops. The NMSE is that of the effective channel at 10 random validation load states: in Fig. 1, lines show (19) with 20000-drop statistics (LMMSE weights from the 3000-drop prior), markers show

Figure 1. Effective-channel NMSE; lines: analysis (19), markers: Monte Carlo. “Exact” denotes estimators built on the coupling-consistent model (3)–(6); T=16T = 16 unless marked; (b) uses the random schedule.

Monte Carlo averages over 60 drops and 4 (a) or 2 (b) noise realizations; the two agree within 0.5 dB.

Estimation Accuracy

In Fig. 1(a), the exact-model LS estimator is consistent, reaching −14.7-14.7 dB at 30 dB SNR, and the optimized schedule gains 8 dB at equal overhead. The LMMSE estimator reaches −27.9-27.9 dB, and −24.9-24.9 dB with T=8T = 8, below the counting bound, as Proposition 3 predicts; a wider prior sector containing the true one (±8λ\pm8\lambda, 5λ5\lambda–20λ20\lambda) still gives −25.1-25.1 dB (not shown). Both cascades saturate at the floors predicted by (19), +2.5+2.5 and +4.1+4.1 dB: their estimate is farther from the truth than the zero vector. In Fig. 1(b), the cascade error is minor below 0.1λ0.1\lambda, where the reactive kernel, growing as ∥η∥−3\|\eta\|^{-3}, makes a resonant stack choke its own currents [10] and leaves the direct link dominant (indicative: finite element profiles soften the kernel there). It crosses 0 dB at d≈0.16λd \approx0.16\lambda, peaks at +2.8+2.8 dB near 0.29λ0.29\lambda, and stays near 0 dB at λ/2\lambda/2, 12–18 dB above the exact LS, whose error varies by only 5 dB.

Multipath leaves this picture unchanged: with Q=10Q = 10 point scatterers re-radiating as dipoles (uniform in ±6λ\pm6\lambda laterally, 3λ3\lambda–16λ16\lambda beyond the last sheet, random complex gains, scattered energy 10 times the line-of-sight energy in every drop), the 99%-energy rank of the prior grows from 9 to 19, and at 30 dB SNR the optimized-schedule LS and the LMMSE estimator, with the prior recomputed, reach −23.1-23.1 and −24.6-24.6 dB, the cascades +2.2+2.2 and +4.4+4.4 dB (40 users, 2 noise realizations each).

Robustness to Model Errors and Quantized Reactances

In Figure 2(b), the true geometry is perturbed by i.i.d. Gaussian jitter of standard deviation σp\sigma_p per coordinate on every port, or the true ζA\zeta_A is off by ±10%\pm10\%, while the estimator keeps the nominal model and optimized schedule (8 draws, 15 users, 2 noise realizations, paired across cases). As (17) shows, such errors matter only once the noise falls below the bias they cause: at 20 dB SNR the loss is at most 0.7 dB for σp≤0.005λ\sigma_p \le0.005\lambda (about @page 4/5 starts_mid=1 ends_mid=1 The currents ιl\iota_l of sheet ll then obey ιl=ΨlZl,l−1ιl−1\iota_l = \Psi_l Z_{l,l-1}\iota_{l-1}, Ψl=−(Zll+ȷdiag⁡(x(l)))−1\Psi_l = -(Z_{ll} + \jmath\operatorname{diag}(x^{(l)}))^{-1}, with x(l)x^{(l)} the reactances of sheet ll, yielding a mismatched regressor B˘t\breve{\mathbf{B}}_t and matrix A˘\breve{\mathbf{A}}. The phase-mask cascade additionally neglects intra-layer coupling and replaces Ψl\Psi_l by diag⁡(eȷarg⁡ψ(xn(l)))\operatorname{diag}(e^{\jmath\arg\psi(x_n^{(l)})}), with ψ\psi the isolated response of § 2-A: this is the model under which the training designs of [3], [4], [6], [7] optimize. Both variants keep the exact kernel blocks as propagators and, unlike the cascade of Remark 1, estimate the same θ∈CN\theta\in\mathbb{C}^{N}, so they are credited with the correct polarized transfer and with the user’s coupling to every port: their errors are purely structural. The mismatched LS estimate under the true observations (6) is

Sum rate with estimated CSI and robustness to model errors

Figure 2. Results at d=λ/4d = \lambda/4. (a) Uplink sum rate, K=3K = 3 users, no training overhead deducted; reactances and MMSE combiners are designed on each scheme’s predicted channel and scored on the true one; the bare feed uses perfect CSI. (b) NMSE of the optimized-schedule LS estimator when the true stack deviates from its nominal model.

θ˘=(A˘HA˘)−1A˘H(Aθ+n),(20)\breve{\theta} = (\breve{\mathbf{A}}^{\mathrm{H}}\breve{\mathbf{A}})^{-1}\breve{\mathbf{A}}^{\mathrm{H}}(\mathbf{A}\theta+ \mathbf{n}), \tag*{(20)}

with bias β=A˘+(A−A˘)θ\beta= \breve{\mathbf{A}}^{+}(\mathbf{A} - \breve{\mathbf{A}})\theta, A˘+≔(A˘HA˘)−1A˘H\breve{\mathbf{A}}^{+} \coloneqq(\breve{\mathbf{A}}^{\mathrm{H}}\breve{\mathbf{A}})^{-1}\breve{\mathbf{A}}^{\mathrm{H}}, independent of the noise: both θ˘\breve{\theta} and the effective channel it reconstructs through its own B˘(x)\breve{\mathbf{B}}(x) carry an NMSE floor at every SNR. The same expression governs any model error, e.g., fabrication tolerances or misestimated ε\varepsilon and ζA\zeta_A, and, for full-column-rank A˘\breve{\mathbf{A}}, bounds the parameter floor as

∥β∥∥θ∥≤∥A−A˘∥2σmin⁡(A˘),(21)\frac{\|\beta\|}{\|\theta\|} \le\frac{\|\mathbf{A} - \breve{\mathbf{A}}\|_2}{\sigma_{\min}(\breve{\mathbf{A}})}, \tag*{(21)}

which favors well-conditioned schedules such as those of Sec. III-C; what sets the cascade apart from such errors is the size of A−A˘\mathbf{A} - \breve{\mathbf{A}}. Both discarded families contribute to it: in the setting of Sec. V at d=λ/4d = \lambda/4, restoring only the direct block Z20\mathbf{Z}_{20} lowers the noiseless floor from +2.7+2.7 to −2.4-2.4 dB, and restoring only the backward block Z12\mathbf{Z}_{12} to +0.7+0.7 dB, so neither truncation alone is benign.

Effective-channel NMSE in closed form: Every estimator above is linear, θ^=Wy\hat{\theta} = \mathbf{W}\mathbf{y}, and predicts the effective channel at a load state xvx_v through its own map C^v=[IN0,−B^T(xv)]\hat{\mathbf{C}}_v = [\mathbf{I}_{N_0},-\hat{\mathbf{B}}^{\mathrm{T}}(x_v)], whereas the truth is Cvθ\mathbf{C}_v\theta with Cv=C(xv)\mathbf{C}_v = \mathbf{C}(x_v). Averaging over the noise and over users with correlation Cθ\mathbf{C}_{\theta}, the NMSE over a set of states {xv}\{x_v\} is

NMSE⁡=∑v[tr⁡(DvCθDvH)+σ2∥C^vW∥F2]∑vtr⁡(CvCθCvH),(22)\operatorname{NMSE} = \frac{\sum_v \left[\operatorname{tr}(\mathbf{D}_v\mathbf{C}_{\theta}\mathbf{D}_v^{\mathrm{H}}) + \sigma^2\|\hat{\mathbf{C}}_v\mathbf{W}\|_{\mathrm{F}}^2\right]}{\sum_v \operatorname{tr}(\mathbf{C}_v\mathbf{C}_{\theta}\mathbf{C}_v^{\mathrm{H}})}, \tag*{(22)}

with Dv=C^vWA−Cv\mathbf{D}_v = \hat{\mathbf{C}}_v\mathbf{W}\mathbf{A} - \mathbf{C}_v. For the exact LS, W=A+\mathbf{W} = \mathbf{A}^{+} and Dv=0\mathbf{D}_v = 0, so (19) decays as σ2\sigma^2, consistently with Corollary 1; for the LMMSE, W=CθAH(ACθAH+σ2I)−1\mathbf{W} = \mathbf{C}_{\theta}\mathbf{A}^{\mathrm{H}}(\mathbf{A}\mathbf{C}_{\theta}\mathbf{A}^{\mathrm{H}} + \sigma^2\mathbf{I})^{-1}; for the cascades, W=A˘+\mathbf{W} = \breve{\mathbf{A}}^{+} and B^=B˘\hat{\mathbf{B}} = \breve{\mathbf{B}}, and the first term is the noise-independent floor of the effective channel itself.

Numerical Results

We follow the scenario of [10]: N0=4N_0 = 4 feed ports on a central λ/2\lambda/2 square at z=0z = 0; L=2L = 2 sheets of 5×55 \times5 atoms on a λ/2\lambda/2 grid (Ns=50N_s = 50), the first at standoff z1=λ/2z_1 = \lambda/2 and the second dd beyond it; ε=0.05\varepsilon= 0.05, ζA=0.58\zeta_A = 0.58. Users are dropped uniformly in a ±4λ\pm4\lambda lateral box, 8λ8\lambda–14λ14\lambda beyond the last sheet, within the Fraunhofer distance (≥16λ\ge16\lambda) of the stack; θ\theta is normalized to ∥θ∥2=N\|\theta\|^2 = N, and SNR =1/σ2= 1/\sigma^2. Training uses T=16T = 16 (counting bound T=14T = 14) and umax⁡=1.35u_{\max} = 1.35; the LMMSE prior is estimated from 3000 kernel drops. The NMSE is that of the effective channel at 10 random validation load states: in Fig. 1, lines show (19) with 20000-drop statistics (LMMSE weights from the 3000-drop prior), markers show

Figure 1 (PDF p. 4). Effective-channel NMSE; lines: analysis (19), markers: Monte Carlo. “Exact” denotes estimators built on the coupling-consistent model (3)–(6); T=16T = 16 unless marked; (b) uses the random schedule.

Monte Carlo averages over 60 drops and 4 (a) or 2 (b) noise realizations; the two agree within 0.5 dB.

Estimation Accuracy

In Fig. 1(a), the exact-model LS estimator is consistent, reaching −14.7-14.7 dB at 30 dB SNR, and the optimized schedule gains 8 dB at equal overhead. The LMMSE estimator reaches −27.9-27.9 dB, and −24.9-24.9 dB with T=8T = 8, below the counting bound, as Proposition 3 predicts; a wider prior sector containing the true one (±8λ\pm8\lambda, 5λ5\lambda–20λ20\lambda) still gives −25.1-25.1 dB (not shown). Both cascades saturate at the floors predicted by (19), +2.5+2.5 and +4.1+4.1 dB: their estimate is farther from the truth than the zero vector. In Fig. 1(b), the cascade error is minor below 0.1λ0.1\lambda, where the reactive kernel, growing as ∥η∥−3\|\eta\|^{-3}, makes a resonant stack choke its own currents [10] and leaves the direct link dominant (indicative: finite element profiles soften the kernel there). It crosses 0 dB at d≈0.16λd \approx0.16\lambda, peaks at +2.8+2.8 dB near 0.29λ0.29\lambda, and stays near 0 dB at λ/2\lambda/2, 12–18 dB above the exact LS, whose error varies by only 5 dB.

Multipath leaves this picture unchanged: with Q=10Q = 10 point scatterers re-radiating as dipoles (uniform in ±6λ\pm6\lambda laterally, 3λ3\lambda–16λ16\lambda beyond the last sheet, random complex gains, scattered energy 10 times the line-of-sight energy in every drop), the 99%-energy rank of the prior grows from 9 to 19, and at 30 dB SNR the optimized-schedule LS and the LMMSE estimator, with the prior recomputed, reach −23.1-23.1 and −24.6-24.6 dB, the cascades +2.2+2.2 and +4.4+4.4 dB (40 users, 2 noise realizations each).

Robustness to Model Errors and Quantized Reactances

In Figure 2(b), the true geometry is perturbed by i.i.d. Gaussian jitter of standard deviation σp\sigma_p per coordinate on every port, or the true ζA\zeta_A is off by ±10%\pm10\%, while the estimator keeps the nominal model and optimized schedule (8 draws, 15 users, 2 noise realizations, paired across cases). As (17) shows, such errors matter only once the noise falls below the bias they cause: at 20 dB SNR the loss is at most 0.7 dB for σp≤0.005λ\sigma_p \le0.005\lambda (about 54 μ\mum at 28 GHz, a typical printed-circuit tolerance) and for the ζA\zeta_A error; at 30 dB, the nominal −22.7-22.7 dB becomes −19.0-19.0, −14.7-14.7, and −9.3-9.3 dB for σp=0.005λ\sigma_p = 0.005\lambda, 0.01λ0.01\lambda, and 0.02λ0.02\lambda, still about 12 dB below the cascade floor. With bb-bit reactances (2b2^b levels of uu equally spaced on [−umax⁡,umax⁡][-u_{\max}, u_{\max}]), projecting the optimized schedule and running two sweeps of coordinate descent keeps full rank and costs 0.6, 1.4, and 10.7 dB at 30 dB SNR for b=3b = 3, 2, and 1.

Sum Rate with Estimated CSI

To test Corollary 1 in use, K=3K = 3 users are estimated as above (data SNR equal to pilot SNR), and each scheme maximizes F(x)=log⁡2det⁡(IK+σ−2H^H^H)F(x) = \log_2 \det(I_K + \sigma^{-2}\hat{H}\hat{H}^{H}) of its own prediction H^(x)=G^0−BT(x)G^s\hat{H}(x) = \hat{G}_0 - B^T(x)\hat{G}_s of the effective channels, where G^0∈CN0×K\hat{G}_0 \in\mathbb{C}^{N_0 \times K} and G^s∈CNs×K\hat{G}_s \in\mathbb{C}^{N_s \times K} collect the estimated segments, by projected gradient ascent over the box of (7) from three random starts. Since dF=2σ2ln⁡2Re⁡tr⁡(MH^HdH^)dF = \frac{2}{\sigma^2 \ln2}\operatorname{Re}\operatorname{tr}(M\hat{H}^{H}d\hat{H}) with M=(IK+σ−2H^H^H)−1M = (I_K + \sigma^{-2}\hat{H}\hat{H}^{H})^{-1} and dH^=−dBTG^sd\hat{H} = -dB^T\hat{G}_s, (15) gives, for the exact model,

∇xF=−2σ2ln⁡2Im⁡{diag⁡(TG^sMH^HBT)}(23)\nabla_x F = -\frac{2}{\sigma^2 \ln2}\operatorname{Im}\left\{\operatorname{diag}\left(T\hat{G}_s M\hat{H}^{H}B^T\right)\right\} \tag*{(23)}

and the cascade models are differentiated likewise. The BS then applies linear MMSE combining designed on H^(x)\hat{H}(x), and the sum rate is scored on the true channel. Figure 2(a) averages 40 drops, deducts no training overhead, and grants the bare feed (stack open-circuited, T=0T = 0) perfect CSI. At 20 dB SNR, optimizing the reactances with perfect CSI raises the sum rate from 15.4 to 23.9 b/s/Hz; the exact-model LMMSE and optimized-schedule LS estimates retain 19.4 and 19.0 b/s/Hz, about 80% of the perfect-CSI rate and half of the stack’s gain, whereas cascade-based CSI never exceeds 2.4 b/s/Hz, 4.5 to 10 times below the bare feed. Charging the KT=48KT = 48 pilot slots through the pre-log factor 1−KT/τc1 - KT/\tau_c (1−K/τc1 - K/\tau_c for the bare feed), the LMMSE-estimated stack wins once the coherence block exceeds τc≈210\tau_c \approx210–310310 symbols across the SNR range (e.g., 17.5 vs. 15.3 b/s/Hz at 20 dB for τc=500\tau_c = 500), i.e., for low-mobility users; T=8T = 8 does not lower this threshold.

Conclusion

This letter formulated channel estimation for SIMs under a mutual-coupling-consistent multiport model, in which the pilots are linear in the unknown channel segments. The LS and LMMSE estimators come with an identifiability condition, an exact error decomposition and prediction error at untrained load states, the LMMSE floor below the counting bound, and a closed-form gradient that optimizes the schedule over realizable reactances. Cascade-based estimation incurs a bias floor, predicted in closed form, of about 100% NMSE or more from λ/6\lambda/6 to λ/2\lambda/2 and collapses the sum rate of load optimization below that of the bare feed, whereas the proposed estimators retain about 80% of the perfect-CSI rate, paying off once the coherence block exceeds a few hundred symbols, and tolerate position errors up to 0.005λ0.005\lambda, 2–3-bit reactances, and multipath. Since the linear model (6), the estimators (8) and (13), the rank condition (9), and Proposition 2 require only that Zss+MZ_{ss}+M be invertible and symmetric, they hold for any lossless reciprocal beyond-diagonal load network M [13]; redesigning (7) and (15)–(16) for such networks, and wideband kernels, are left for future work.

References

  1. [1]J. An et al., “Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,” IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2380–2396, 2023.arxiv.org/abs/2305.08079
  2. [2]C. K. Sheemar, W. U. Khan, S. Solanki, G. C. Alexandropoulos, and S. Chatzinotas, “A survey on stacked intelligent metasurfaces: Fundamentals, recent advances, and challenges,” arXiv preprint arXiv:2603.05633, 2026.arxiv.org/abs/2603.05633
  3. [3]X. Yao, J. An, L. Gan, M. Di Renzo, and C. Yuen, “Channel estimation for stacked intelligent metasurface-assisted wireless networks,” IEEE Wireless Commun. Lett., vol. 13, 2024.DOI
  4. [4]A. Papazafeiropoulos, P. Kourtessis, D. I. Kaklamani, and I. S. Venieris, “Channel estimation for stacked intelligent metasurfaces in Rician fading channels,” arXiv preprint arXiv:2502.12692, 2025.arxiv.org/abs/2502.12692
  5. [5]Q.-U.-A. Nadeem, J. An, and A. Chaaban, “Hybrid digital-wave domain channel estimator for stacked intelligent metasurface enabled multi-user MISO systems,” in Proc. IEEE Wireless Elec. Commun. Netw. Conf. (WCNC), 2024.arxiv.org/abs/2309.16204
  6. [6]N. Ginige, A. S. de Sena, N. H. Mahmood, M. Di Renzo, N. Rajatheva, and M. Latva-Aho, “Nested tensor-based channel estimation for stacked intelligent metasurface-assisted wireless networks,” IEEE Commun. Lett., vol. 29, no. 11, pp. 2731–2735, 2025.
  7. [7]A. L. F. de Almeida and G. C. Alexandropoulos, “TenSIM: Tensor-based channel estimation for MIMO systems with stacked intelligent metasurfaces,” arXiv preprint arXiv:2606.00917, 2026.arxiv.org/abs/2606.00917
  8. [8]M. Nerini and B. Clerckx, “Physically consistent modeling of stacked intelligent metasurfaces implemented with beyond diagonal RIS,” IEEE Commun. Lett., vol. 28, no. 7, pp. 1693–1697, 2024.arxiv.org/abs/2402.12602
  9. [9]A. Abrardo, G. Bartoli, and A. Toccafondi, “A novel comprehensive multiport network model for stacked intelligent metasurfaces (SIM) characterization and optimization,” IEEE Trans. Commun., vol. 73, no. 11, pp. 11 559–11 573, 2025.arxiv.org/abs/2501.02597
  10. [10]G. Iacovelli, C. K. Sheemar, and S. Chatzinotas, “Stacked fluid metasurfaces: Mutual coupling-aware modeling and optimization,” arXiv preprint arXiv:2609.31591, 2026.arxiv.org/abs/2609.31591
  11. [11]A. Abrardo and G. Bartoli, “SIM-aided near-field channel and localization estimation with dimensionality reduction: A multiport network theory approach,” arXiv preprint arXiv:2605.23779, 2026.arxiv.org/abs/2605.23779
  12. [12]S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. Englewood Cliffs, NJ, USA: Prentice-Hall, 1993.
  13. [13]H. Li, S. Shen, and B. Clerckx, “Beyond diagonal reconfigurable intelligent surfaces: From transmitting and reflecting modes to single-, group-, and fully-connected architectures,” IEEE Trans. Wireless Commun., vol. 22, no. 4, pp. 2311–2324, 2023.

Paper details

vlnc:W_f5v4u4qe
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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