WorksEngineering
Space-Time Stacked Intelligent Metasurfaces for Doppler-Resilient LEO Broadcasting
Abstract
Stacked intelligent metasurfaces (SIMs) cascade programmable layers in front of a feed. In the SIM designs reported so far, including those proposed for low-Earth-orbit (LEO) satellites, the layer coefficients are space-only controls or are varied randomly in time to support opportunistic scheduling, so that every radiated beam carries the frequency of the feed. This paper introduces space-time SIMs (ST-SIMs), whose layers are modulated in time and use the temporal dimension of the stack deterministically: each spatial mode of the aperture is paired with its own prescribed frequency translation. The application is LEO broadcasting, where a common waveform reaches terminals with different Doppler shifts, and any common Doppler correction leaves residuals of up to half the Doppler span. We derive the harmonic-domain model of the stack, solve the minimum-energy space-frequency synthesis in closed form, and prove that full column rank of the steering matrix is necessary and sufficient for decoupling K terminal-specific translations, a result that extends to groups of terminals. We further propose a phase-only synthesis algorithm with guaranteed stationarity, together with quantization and control-rate laws. Simulations at 20 GHz over 50 random drops show that a four-layer ST-SIM with a single radio-frequency chain removes the Doppler floor of a static SIM for footprint radii above about 45 km and, under the ideal phase-only model and at equal radiated peak energy, approaches ideal digital per-terminal precompensation within 0.5 dB for resolvable terminals, whereas, in the evaluated geometries, a single time-modulated layer cannot synthesize the required modes.
Introduction
-Earth-orbit (LEO) constellations are a cornerstone of fifth- and sixth-generation non-terrestrial networks (NTNs) owing to their global coverage, moderate latency, and standardized integration with terrestrial radio access [1, 2, 3]. Their orbital speed of about km/s is also their most persistent impairment, since the satellite–terminal range changes by kilometers per second and the Doppler shift scales with the carrier frequency. At GHz, the Doppler shift exceeds kHz, i.e., more than ten subcarrier spacings of a 30-kHz orthogonal frequency-division multiplexing (OFDM) numerology, and it drifts at rates of several kHz/s [10, 11, 9]. An uncompensated carrier-frequency offset (CFO) of this size destroys subcarrier orthogonality and generates inter-carrier interference (ICI) [16, 17], and Doppler handling permeates every NTN procedure [12, 13, 15]. Since orbital motion is deterministic, the predominant remedy is precompensation, whereby the transmitter, i.e., the gateway for a transparent and the satellite for a regenerative payload, shifts the transmitted carrier by the Doppler predicted from the ephemeris [9, 13, 14]. Such a correction is exact for a single direction. A broadcast beam, however, serves terminals with different line-of-sight (LoS) directions and radial velocities, and, as shown in § 3, the best common translation leaves a worst-terminal residual equal to half the Doppler span, a floor independent of the transmit power, the noise level, and any digital processing of the common waveform. For a 200-km footprint radius at GHz this span exceeds kHz, i.e., ten subcarrier spacings, and the residual is left to the terminals [15, 12].
Stacked intelligent metasurfaces (SIMs) cascade programmable metasurface layers of meta-atoms each in front of a feed, so that inter-layer diffraction implements a high-dimensional transformation of the incident wave [22, 21, 5]. Each layer applies a diagonal phase profile to the field impinging on it, and the free-space propagation between consecutive layers mixes the outputs of all meta-atoms, so that the cascade realizes a dense linear operator with only phase controls and without any digital multiplication. The stack can thus perform beamforming, precoding, or a discrete Fourier transform in the wave domain at the speed of light, which reduces the number of RF chains, the baseband load, and the power consumption, while its depth and its meta-atoms per layer provide degrees of freedom that a single-layer surface lacks. SIMs have been studied for multiuser beamforming [23], channel estimation [27], physically consistent modeling [24], LEO satellite downlinks [26, 4], wideband MIMO-OFDM [28], unmanned aerial vehicles (UAVs) [7, 6], and doubly-dispersive channels [29, 25]. In all of these works, the layer coefficients are space-only control variables: the stack shapes the spatial or frequency-selective response of a beam, but every beam carries the frequency of the feed, and temporal modulation of the layers has not been exploited as a deterministic signal-processing degree of freedom.
State-of-the-Art
Doppler compensation in NTNs has been studied extensively. The 3GPP NTN study [9] relies on ephemeris-assisted precompensation at the network and on CFO estimation from reference signals at the terminal [13], and dedicated solutions have been proposed for random access [12], narrowband Internet-of-Things (IoT) uplinks [14], multibeam initial synchronization [15], and massive multiple-input multiple-output (MIMO) LEO downlinks with per-user Doppler and delay precompensation in the digital domain [11]. Doppler-robust waveforms [18, 19] relocate the problem to the receiver equalizer, and reconfigurable intelligent surfaces (RISs) have mitigated Doppler in high-mobility links [20] as externally illuminated reflectors rather than as the satellite front end.
In parallel, the electromagnetics literature has established that time-varying surfaces manipulate frequency. Time-modulated arrays were used decades ago to shape sidelobes and generate harmonics [31], and a single-layer time-modulated array has been shown to separate downlink users by pre-shifted carriers with one RF chain [47]. Space–time gradient metasurfaces [32] and spacetime metamaterials [33] showed that joint spatial and temporal variation yields frequency conversion and nonreciprocity, and space–time-coding digital metasurfaces [34] demonstrated programmable control of the propagation direction and harmonic power distribution, independent control of harmonic amplitude and phase [35], joint multi-frequency beam shaping and steering [48], Doppler-effect emulation and frequency conversion [37, 38, 39], space- and frequency-division multiplexing [36], sideband-free frequency conversion in a waveguide-fed antenna [49], and frequency conversion with beam shaping on a 1-bit metasurface antenna [40, 41]. Closest to the present work, a single-layer 1-bit spatiotemporal reflecting metasurface on the propagation path offsets the Doppler shift of one high-mobility link by steering one frequency-translated harmonic toward the user, with a switching sequence obtained by a neural network [45]. Stacks of spatially uniform time-varying metasurfaces realize nonreciprocal isolation and frequency circulation [46], and in the SIM literature temporal variation has so far served only to randomize the channel for opportunistic scheduling [30].
Table 1. Positioning With Respect to Time-Varying Metasurfaces and SIMs
Table 1 positions this paper. In single-layer time-modulated surfaces, the beam frequencies are harmonics that serve as channels [47, 48, 36, 49], or a single translation offsets the Doppler of one link [45], so that terminals with different Doppler shifts retain half their Doppler span (Proposition 2); in multilayer stacks, the time dimension has served randomization [30] or, with uniform layers, nonreciprocity [46]. The deterministic pairing of distinct spatial modes of a stack with Doppler-matched translations, and the conditions under which a finite-depth, finite-resolution, and finitely fast phase-only stack realizes it, have not been investigated.
Contributions
This paper proposes space-time (ST-) SIMs for deterministic space–frequency processing at the LEO satellite (Figure 1). The spatial modes and the temporal translations cannot be designed independently, since every frequency-translated mode must be spatially nulled at the unintended terminals and the resulting field must be synthesized by a finite-depth, finite-resolution, and finitely fast phase-only stack. The main contributions of our work are:
We present the ST-SIM architecture and derive its harmonic-domain model, which shows that stacking convolves the temporal spectra of the layers while mixing their spatial fields, and that desired modes and unwanted sidebands share one passive energy budget.
We develop the space–frequency synthesis theory: we solve the minimum-energy multiuser synthesis in closed form with its complete solution set, geometry-dependent energy bounds, and long-run peak energy; we prove that full column rank of the steering matrix is necessary and sufficient for exactly decoupling terminal-specific translations, a condition that grouped compensation relaxes; we show that any common translation, hence any static SIM or single-translation time-modulated surface, leaves half the Doppler span; and we quantify in closed form the OFDM interference caused by residual Doppler and ephemeris errors.
We propose a realizable design: a phase-only synthesis algorithm with closed-form gain scaling, reusable multilayer gradients, guaranteed monotone decrease to stationary points, and quantization-aware refinement, and we derive explicit quantization and control-rate laws.
We demonstrate the architecture in a link-level study at GHz over random drops, in which we compare the ST-SIM with a single-layer time-modulated aperture, a static SIM with optimal common precompensation, with the same payload combined with per-terminal CFO correction, and with ideal digital per-terminal precompensation, and we map the operating regimes in footprint radius, aperture size, number of terminals, and grouping.
Figure 1. User-dependent Doppler precompensation in the electromagnetic domain with an ST-SIM. A single feed illuminates an -layer stack of time-varying phase-only metasurfaces on a LEO satellite, which transforms the common waveform into spatial modes directed toward the terminals; mode is translated in frequency by so that terminal receives a carrier-Doppler-free signal.
Paper Organization: § 2 presents the system model, § 3 the space–frequency synthesis theory, and § 4 the phase-only synthesis algorithm and its hardware laws. § 5 reports the numerical results, and § 6 concludes the paper.
Notation: Boldface lowercase and uppercase letters denote vectors and matrices. , , , and denote transpose, Hermitian transpose, conjugation, and Moore–Penrose pseudoinverse; and the Euclidean/spectral and Frobenius norms; , the extreme eigenvalues of a Hermitian matrix; the diagonal matrix built from ; the identity; the th canonical vector; the entrywise phase wrap to ; and .
System Model
We consider a 6G multicast–broadcast service in which one LEO beam delivers a common data stream to registered terminals whose positions are known to the network from their registration reports.
LEO Propagation and Doppler
Let denote transmit time and receive time. In an Earth-centered inertial frame, the satellite and terminal have positions and , a wavefront emitted at is received at , and the one-way light time solves the light-time equation [8]. Let be the unit-power complex envelope applied to a common RF feed at carrier , the slowly varying amplitude gain, the LoS steering vector of the radiating aperture, and the aperture field produced by the ST-SIM. After ideal downconversion at , the received baseband signal is
with the retarded transmit time and additive noise. Model (1) does not introduce Doppler as a separate impairment: the carrier Doppler is the time derivative of the delay-induced carrier phase. Appendix § 7 shows that, with the unit line-of-sight (LoS) vector and the relative velocity, the instantaneous carrier Doppler is up to a relative correction below for LEO motion, Over one OFDM block of duration centered at , the carrier phase is linear in up to , the delay varies by at most , and the LoS direction rotates by at most , with the range1. Under these conditions, (1) reduces to the blockwise model
where absorbs the constant phase and path gain. Model (2) is a LoS model; multipath Doppler, phase noise, and oscillator offsets remain for the terminal’s synchronization. The ST-SIM thus performs the coarse synchronization by removing the orbital Doppler of the LoS path, and the terminal performs only the fine synchronization.
Multilayer ST-SIM Transfer Function
The satellite carries an -layer ST-SIM with meta-atoms per layer, illuminated by one RF feed, which radiates the OFDM signal at carrier with bandwidth , through a fixed coupling vector , . The time-varying phase matrix of layer is
and , , is the fixed propagation operator from layer to layer . The aperture field is
Let and let be the transit time across the stack.
Assumption 1 (Frequency-flat inter-layer propagation). , so a common applies to all generated sidebands.
Assumption 2 (Negligible modulation during transit). , so all layers are evaluated at the same control time in (4).
For held control, the modulation spectrum of the layers extends to the update rate and its harmonics rather than to , so both conditions must hold with replaced by the control bandwidth. For a passive stack we assume
with equality for lossless propagation; the are moreover assumed nonsingular, as the Rayleigh–Sommerfeld operators of § 5 are. Equations (3)–(4) constitute the ideal phase-only wave-domain model of the SIM literature [22, 23]: diagonal unit-modulus layers, fixed inter-layer operators, and no multiple reflections, state-dependent or amplitude–phase coupling, or bias-network parasitics. In this model, is the energy of the element excitations, which equals the radiated power only up to the radiation matrix of the aperture; the accuracy of the single-pass cascade depends on the inter-layer spacing and on the element design [24, 5], and the transmission efficiency is a property of the normalized model rather than of a device.
Harmonic-Domain Representation
Suppose first that each layer is periodic with modulation period , so that with diagonal Fourier coefficients .
Lemma (Harmonic convolution across the stack). with and the th aperture harmonic is .
Stacking thus convolves the temporal spectra of the layers while the inter-layer operators mix the associated spatial fields. The reconfiguration capabilities of the meta-atoms enter (6) only through the : a state held over a reconfiguration period weights the harmonics by and replicates them at multiples of , so that only tones with can be synthesized (Proposition 7), and a -bit reconfiguration accuracy diverts energy into the unwanted harmonics (Proposition 6). Lemma (?) also specifies which modal fields the stack can produce. Let the tones to be synthesized, , lie on a grid of spacing (cf. § 4) and set . Each tone is then a harmonic of the stack, with , and the aperture field decomposes as
where collects the unwanted harmonics generated by (6). The number of modes is the number of beams, each of which carries one frequency translation, whether it serves one terminal or several: for one beam per terminal and for one multicast beam per group in the grouped compensation. Synthesis therefore amounts to concentrating the harmonic content (6) on the indices with the prescribed spatial profiles; § 4 performs this in the time domain. For tones that do not share a common grid, the control is quasi-periodic, and (6) holds with the integer indices replaced by the frequencies generated by the layers and index addition replaced by frequency addition.
Lemma 2 (Spectral-energy constraint). Let the inter-layer operators satisfy (5) and let . Then, for every control trajectory , the aperture field (4) satisfies with equality for all if is unitary for every . Moreover, if the control is periodic with period , the aperture harmonics of Lemma (?) satisfy
Lemma 2 shows that temporal modulation redistributes the aperture energy across spectral modes rather than adding power: : the desired modes and the unwanted sidebands share the same energy budget. Moreover, a lossless stack has constant energy , whereas a multi-tone field has time-varying energy (Lemma 3); the difference is carried by or absorbed by loss.
Space–Frequency Synthesis and Doppler Compensation
Minimum-Energy Multiuser Synthesis
Within one block, collect the steering vectors in and let . To cancel the Doppler of terminal , one aperture mode is assigned the temporal factor :
with . Terminal must observe only mode , i.e., , or
where prescribes the aperture gain toward terminal . Among all decoupling matrices we select the one of minimum modal energy,
where is also the long-time average energy of (10) for distinct mode frequencies, because the temporal exponentials are orthogonal in time average.
Proposition 1 (Solution set, minimum-energy solution, and geometry bounds). Let and . Then (12) is feasible, every feasible matrix has the form with arbitrary and the orthogonal projector onto , and is the unique minimum-Frobenius-norm solution, with ,
Equation (15) gives terminal geometry a direct physical meaning: when two steering vectors become nearly collinear, and the energy needed to separate their frequency translations grows without bound. Spatial separability and Doppler compensation are therefore fundamentally coupled. Passivity (Lemma 2) constrains the instantaneous rather than the average energy, which motivates the following definition.
Lemma 3 (Long-run peak energy of the modal field). For any and rationally independent , the trajectory is dense in the phase torus, so that satisfies , with the long-time average equal to .
The phases of do not change , since the trajectory sweeps every phase combination.
Perfect Blockwise Multiuser Compensation
For arbitrary temporal frequencies let with . Substituting into (2) and absorbing fixed delay phases into the coefficients yields the space–frequency coupling identity
where . The identity factorizes the two design dimensions: is the spatial leakage of temporal mode into terminal , and is the residual temporal rotation of that mode.
Theorem 1 (Perfect blockwise multiuser Doppler compensation). Let for all , , and suppose the ST-SIM realizes with from (13). Then every terminal has zero residual Doppler and zero inter-mode leakage: Conversely, exact decoupling with a nonsingular diagonal requires ; if the are distinct, the rank condition is also necessary for (18) to hold for all .
Theorem 1 identifies the sufficient aperture field and the rank condition for its existence; whether a finite-depth phase-only stack synthesizes it is addressed in § 4.
Remark 1 (Grouped compensation). The rank condition concerns mutually decoupled translations. If the terminals are partitioned into groups sharing one tone and one multicast mode each, decoupling requires with the group-indicator matrix , which is consistent whenever and possibly otherwise (e.g., for collinear terminals in one group), with minimum-norm solution , i.e., under full rank. By Proposition 2 per group, each terminal keeps at most half the Doppler span of its group; is the common translation and is Theorem 1 (§ 5.5). Terminals should thus be grouped contiguously in Doppler, i.e., by along-track position: a group of Doppler span keeps a residual with interference-limited SIR (Proposition (?)), whereas every additional group costs modal energy (Proposition 1), a trade-off evaluated in § 5.5.
Irreducible Residual of a Common Frequency Translation
A conventional payload with one RF chain and one mixer, a static SIM, or a time-modulated surface that applies one translation to all beams [47, 45], imposes a single translation on the common waveform, with worst-terminal residual .
Proposition 2 (Optimal common precompensation). With and , the minimax translation and residual are and the least-squares translation is . Both minimize a residual rather than the SINR; the SINR-optimal translation , a one-dimensional search over , is evaluated in § 5.3.
Whenever the Doppler span is nonzero, no common translation makes every residual vanish; Theorem 1 removes this floor by pairing a different translation with each spatial mode, and the comparison is architectural rather than algorithmic. Since the payload contains a frequency synthesizer, the numerical results apply to the feed and let the ST-SIM synthesize only the differential tones ; all results hold for arbitrary tone frequencies, and this hybrid mode reduces the largest synthesized tone from to , by a factor of in the baseline geometry and more for clustered terminals.
OFDM Consequences and Residual Errors
Consider subcarriers with spacing and useful duration , and let terminal have residual frequency error with normalized value . After cyclic-prefix removal, the DFT-domain relation is with
The CFO matrix is unitary, so , and under Theorem 1 , , and the carrier-Doppler-induced ICI vanishes for every terminal. For imperfect compensation, the fraction of desired-mode power leaving the intended subcarrier is
Proposition (Local ICI law). As , .
Since the common-compensation residual scales linearly with and the ICI fraction quadratically with the residual normalized by , differential Doppler becomes critical as grows (§ 5.3). Two residuals remain after carrier-Doppler compensation. First, the envelope in (1) is time-scaled by , , which leaves the band-averaged ICI fraction , i.e., and dB at and kHz; this sampling-clock-offset effect is common to all carrier-Doppler schemes and is removed by the timing tracking of an OFDM receiver. Second, ephemeris and terminal-position errors leave a residual tone. With the first-order model , suppose the controller uses and to predict , with the range vector, and let .
Proposition (Residual Doppler under state errors). If , then and, to first order, .
If the ST-SIM applies , the residual CFO is and, by Proposition (?), the ICI fraction is , which exposes the critical factor of high-frequency LEO OFDM links. Only the transverse component of a position error matters, scaled by .
Phase-Only ST-SIM Synthesis
The ideal field of Theorem 1 need not belong to the set of fields realizable by a finite-depth phase-only stack. This section formulates the physical synthesis problem and derives an efficient first-order method with explicit guarantees.
Trajectory Synthesis Problem
Sample a predictable control horizon at transmit times , , where is the control update rate, and let be the desired field generated by (10); since has full column rank and , for every . Let collect all layer phases at slot and . A passive stack cannot reach an arbitrary absolute amplitude (Lemma 2), so a common complex scaling , which absorbs the transmission gain of the stack, is optimized jointly with the phases. The finite-resolution design problem is
where is the -bit phase alphabet, are trajectory weights, and
penalizes switching between consecutive slots. The penalty (24) is -periodic and smooth, unlike a penalty on wrapped phase differences, so it does not affect the differentiability assumed below; its gradient is . A hard slew limit, if required by the drivers, can be imposed afterwards by projecting each increment onto . Problem (23) is nonconvex because of the cascaded phase products and the discrete alphabet. We study its continuous relaxation, then quantify the effect of quantization.
Optimal Scaling and Reusable Multilayer Gradient
For one slot (index suppressed), define the field incident on layer and the transfer from the output of layer to the aperture, with ,
so that for every .
Lemma (Optimal scaling and phase derivatives). For fixed , with and , the minimizer of over is and has gradient . For meta-atom of layer , , and for ,
All derivatives of a slot follow from one forward sweep for and one backward sweep for , i.e., matrix–vector products of size , without differentiating the full cascade for every meta-atom.
Algorithm and Convergence
Algorithm alg:synthesis summarizes the proposed synthesis. Step 1 (lines 2–8) performs wrapped gradient descent on the continuous relaxation, where wrapping only changes the coordinate representation of a point on the phase torus; the trial step is the Barzilai–Borwein estimate of the inverse local curvature [43], safeguarded to and then reduced by Armijo backtracking until sufficient decrease is achieved. Step 2 (line 9) projects onto , and Step 3 (lines 10–12) performs quantization-aware coordinate refinement, in which each meta-atom is assigned the best of its states given all others, using and so that each trial costs . With fixed, the coordinate cost of collects all terms of that depend on it,
where , , and the switching terms are omitted at and . For the slots decouple and are updated in parallel; for the even and the odd slots are updated in alternate half-sweeps, so that every update sees fixed neighbors.
\caption{Phase-only ST-SIM trajectory synthesis}
\label{alg:synthesis}
\begin{algorithmic}[1]
\REQUIRE $\{\mathbf{g}_r^\star\}_{r=0}^{R-1}$, $\{\mathbf{W}_l\}$, $\mathbf{f}$, $B_\phi$, $0<c<1$, $0<\rho<1$, $I_{\max}$, $n_{\rm ref}$
\STATE Initialize $\Theta^{(0)}$ uniformly at random on $[0,2\pi)^{LN\times R}$
\FOR{$i=0,1,\ldots,I_{\max}-1$}
\STATE Forward sweeps: $\{\mathbf{u}_l[r]\}$, $\mathbf{g}_r$, and $\alpha^\star$ by \eqref{eq:alphastar}
\STATE Backward sweeps: $\mathbf{\Gamma}_i\triangleq\nabla\tilde{\mathcal J}(\Theta^{(i)})$ by \eqref{eq:objectivegradient}
\STATE Trial step $\mu\gets\langle\mathbf{S}_i,\mathbf{S}_i\rangle/\langle\mathbf{S}_i,\mathbf{\Gamma}_i-\mathbf{\Gamma}_{i-1}\rangle$, $\mathbf{S}_i=\mathcal W_{2\pi}(\Theta^{(i)}-\Theta^{(i-1)})$, clipped to $[\mu_{\min},\mu_{\max}]$
\STATE \textbf{while} $\tilde{\mathcal J}(\mathcal W_{2\pi}(\Theta^{(i)}-\mu\mathbf{\Gamma}_i))>\tilde{\mathcal J}(\Theta^{(i)})-c\mu\left\lVert \mathbf{\Gamma}_i\right\rVert_F^2$ \textbf{do} $\mu\gets\rho\mu$
\STATE $\mu_i\gets\mu$;\ $\Theta^{(i+1)}\gets\mathcal W_{2\pi}(\Theta^{(i)}-\mu_i\mathbf{\Gamma}_i)$
\ENDFOR
\STATE Quantize: $\widehat\Theta\gets\arg\min_{\Theta\in\mathcal Q_{B_\phi}^{LN\times R}}\left\lVert \mathcal W_{2\pi}(\Theta-\Theta^{(I_{\max})})\right\rVert_F$
\FOR{$n_{\rm ref}$ rounds, every slot $r$, layer $l$, meta-atom $n$}
\STATE $\widehat\theta_{l,n}[r]\gets\arg\min_{\phi\in\mathcal Q_{B_\phi}}\mathcal J_{l,n,r}(\phi)$ in \eqref{eq:coordcost} \hfill (even $r$, then odd $r$)
\ENDFOR
\ENSURE $\widehat\Theta$, $\alpha^\star$
\end{algorithmic}
Proposition 5 (Stationarity of Armijo descent). Let with and the smooth penalty (24), which is continuously differentiable with Lipschitz gradient on the compact phase torus (the data term is a smooth function of unit-modulus entries and since and the are nonsingular), and let be generated by Step 1 of Algorithm alg:synthesis with trial steps in , . Then , , and every accumulation point is first-order stationary; in particular whenever the accepted step sizes are bounded away from zero along nonstationary subsequences. Every update of Step 3 is nonincreasing in for any .
The trial step is well defined for every curvature estimate: when , the Barzilai–Borwein quotient is replaced by , which the backtracking then reduces until sufficient decrease holds, so the guarantee of the proposition is unaffected. In the baseline geometry of § 5, after iterations is , , , and for , , , and , and the single layer meets the tolerance after iterations: its stall at dB is a stationary point. The proposition guarantees monotone decrease and convergence to first-order stationary points, without global optimality, for the continuous relaxation of ; the discrete alphabet is handled by Steps 2–3. Each iteration and backtracking trial cost .
Finite-Resolution Phase Quantization
Let denote the nearest-neighbor -bit quantization of and the corresponding stack matrix, and define , so that .
Proposition 6 (Multilayer quantization scaling law). With , where the last inequality holds for a passive stack.
The bound ignores cancellation among layers and is loose for coarse quantization, where it exceeds the norm of the field itself ( for with ); it is to be read as a scaling law: for fixed depth the field error decays as , i.e., the synthesis error , which is a squared norm, as or dB per bit, and for fixed resolution its worst-case accumulation is at most linear in (Figure 2b).
Control Rate, Switching Harmonics, and Control Bandwidth
If the desired field contains temporal frequencies in and the control is interpolated ideally, sampling theory requires . Practical digital-coding metasurfaces, however, hold each state constant over , which generates switching harmonics.
Proposition 7 (Control-rate law for held control). Let the aperture field be held constant over each control slot, for , and let the ideal field toward terminal be the tone . Then the held field toward terminal contains the tone at with amplitude and images at , , with amplitudes . The energy fraction diverted into the images is so that, when the images fall inside the signal band, the self-interference ratio of terminal is capped at .
Equation (30) has the same rectangular-window origin as Proposition (?) and yields an explicit hardware requirement: an in-band SIR of 30 dB for a synthesized tone of kHz requires MHz, and when exceeds the signal bandwidth the images fall out of band, where they are an emission rather than an interference problem (Figure 2c).
The binding hardware constraint is not the MHz-rate switching of single elements, which is established [34, 36, 40], but the aggregate control bandwidth: streaming every state requires bit/s, i.e., Gbit/s for , , and MHz. If the tones are rounded to a grid , the target is periodic with period , and each meta-atom can replay a local table of states, provided that one full period is synthesized offline, since a nonconvex synthesis does not return a periodic trajectory by itself; for Hz and MHz, a period comprises slots ( phases for ), which is tractable because (23) couples the slots only through the switching penalty and the gain : synthesized in parallel blocks of slots with fixed and overlapping boundaries ( s per block on one core), the period costs about core-minutes per pass geometry. The tables are refreshed when a rounded tone has drifted by , which reduces the streaming rate by , i.e., with the drift of at most kHz/s of § 5, by for Hz and for Hz (residual below ).
Numerical Results
In this section, we extensively investigate the performance of the proposed ST-SIM architecture.
Setup, Benchmarks, and Metrics
Table 2 lists the baseline parameters. At , the four baseline terminals at ground offsets , , , and km (along-, cross-track) see Doppler shifts of , , , and kHz. Geometry follows § 2 with static terminals in the free-space LoS channel of (1), with steering vectors and Doppler shifts evaluated at the actual directions and slant ranges; the path gain is absorbed into and taken equal for all terminals, since the slant range grows from km at the beam center to km at km from it, i.e., by at most dB of free-space loss; and follow the Rayleigh–Sommerfeld model of [22, 42] with . The payload broadcasts one common waveform, and the control trajectory is synthesized over the OFDM symbol including its cyclic prefix ( samples) and held over samples, so switching harmonics are included; the ST-SIM synthesizes the differential tones . Algorithm alg:synthesis uses , , a gradient tolerance of , , and . Figures show medians over random drops unless stated.
| Parameter | Value | Parameter | Value |
| Carrier | 20 GHz | Altitude | 600 km |
| Terminals | 4 | Footprint radius | 200 km |
| Layers | 4 | Meta-atoms/layer | |
| Element spacing | Layer / feed spacing | / | |
| Subcarriers | 1024 | Spacing / CP | 30 kHz / 72 samples |
| Control rate | 7.68 MHz | Slots per symbol | 274 |
| Phase bits | 1–3, | Armijo | |
| Random drops | 50 | Feed SNR | 25 dB |
Table 2. Baseline Simulation Parameters
(a) Convergence of Algorithm~alg:synthesis
(b) Synthesis error vs.\ phase resolution
(c) In-band SIR and out-of-band power vs.\ control rate
Figure 2. Realizability (, , , unless stated). (a) Objective of Step 1 of Algorithm alg:synthesis for (three initializations each), fixed-step, Adam, and layer-wise descent for . (b) Synthesis error with quantization-aware refinement () and projection only (), with the scaling law of Proposition 6. (c) In-band SIR (left) and out-of-band power (right) versus , with the law of Proposition 7.
Three benchmarks are considered. Benchmark 1 is a space-only SIM of the same geometry [23, 26] synthesizing the max–min multicast beam , computed by multi-start weighted power iteration, which is not certified to be globally optimal but is never worse than the sum of the decoupled modes, with its feed shifted by ; it represents every single-translation front end (Proposition 2). Benchmark 2, ideal digital per-terminal precompensation, applies one digital translation per terminal and a lossless zero-forcing precoder, i.e., radiates exactly; it requires RF chains and a lossless multibeam network and upper-bounds every exact realization of Theorem 1. Benchmark 3 is the payload of Benchmark 1 with every terminal correcting its residual CFO, including the error of Moose’s two-symbol estimator, [17], and, where stated, two pilot symbols per 14-symbol slot. The no-compensation reference is the space-only SIM without frequency correction. Finally, a single-layer time-modulated aperture is also investigated, , with the same feed, phase resolution, control rate, and synthesis objective as the ST-SIM (the one-layer case of Algorithm alg:synthesis), isolating the contribution of stacking.
The performance metrics are obtained as follows. The aperture field is evaluated on a four-times oversampled grid, Doppler-shifted, band-limited to by an ideal receiver filter, and sampled; for terminal , the response over the DFT window gives the ICI matrix , whose diagonal equals ; hence , and is the average interference for i.i.d. data, which includes switching images, quantization noise, and inter-mode leakage. Since the receiver filter acts on the product of the aperture response and the OFDM waveform, this cyclic model was verified by a full waveform simulation (random CP-OFDM symbols through the time-varying aperture, Doppler shift, receiver filter, decimation, and FFT): over the drops of Figure 3b, the per-terminal SIR of the ST-SIM differs by dB in median ( dB at the 95th percentile), the cyclic model being slightly pessimistic, and the worst-terminal SINR by less than dB; over the control rates of Figure 2c, the average SIR of the full waveform differs from the cyclic model by to dB with continuous and three-bit phases, except by dB with continuous phases at MHz, where the SIR exceeds dB. The SINR is with , i.e., the SNR of a lossless full-aperture beam including the array gain ( dB per meta-atom at dB for ), and, since all terminals decode the same broadcast stream, the figures report the worst-terminal SINR and the broadcast spectral efficiency . All schemes are compared at equal long-run radiated peak energy, defined by for every ideal field , with the long-run peak energy (16), and by for every realized field, i.e., the gain of (26) fits the realized field to the ideal target and is reduced if the energy exceeds one within the simulated window, as required by passivity (Lemma 2); the equal feed power figures of merit additionally scale the stack-based schemes by their simulated transmission efficiency .
Realizability: Convergence, Quantization, and Control Rate
In Figure 2a, the objective of Algorithm alg:synthesis decreases monotonically (Proposition 5). The single layer stalls at dB for every initialization: a phase-only layer cannot change the amplitude of any element excitation, whereas the multi-tone target (10) requires element amplitudes that vary in time, so that inter-layer mixing is necessary rather than merely helpful. For it reaches to dB after iterations and the three initializations of each depth end within dB; stalls near dB, lacking the degrees of freedom to shape four frequency-translated modes. Fixed-step and layer-wise descent reach only and dB; Adam descends faster ( and dB at learning rates and ) but is not monotone ( and ascending steps out of ) and must be tuned per problem; since the quantization floor of dB lies more than dB above the stationary points of either method, the monotone, tuning-free descent of Algorithm alg:synthesis is preferable for an autonomous payload.
Phase resolution does not change this picture (Figure 2b): the single layer stays between and dB from three bits on, whereas two layers reach dB with continuous phases. In Figure 2b, the refinement of Step 3 lowers the four-layer error after projection from , , and dB to , , and dB for , , and bits, which is worth more than one bit, and all refined curves follow the -dB-per-bit slope of Proposition 6 ( and dB per bit from two to four bits for ). Beyond depth barely helps (, , and dB at three bits for , , and ), whereas drops from to , , and , so the simulated diffraction loss limits the useful depth. In Figure 2c, the in-band SIR with continuous phases follows Proposition 7 within dB up to MHz and exceeds it from MHz on ( dB on average, dB for the terminal with the largest tone), where the receiver filter rejects part of the images, whereas two and three bits saturate at and dB, the quantization floor. The gain at high rates is not free: the out-of-band fraction decreases from dBc at MHz to dBc at MHz for continuous phases, but rises again for quantized phases ( dBc for two bits at MHz), whose quantization noise spreads over a bandwidth proportional to ; generated in the radiating aperture, beyond any payload filter, these images must be controlled through , , and the control waveform.
(a) Radiated power vs.\ direction and frequency
(b) Worst-terminal SINR vs.\ footprint radius
Figure 3. Link-level performance (, , , dB, -km footprint, drops, equal long-run radiated peak energy, unless stated). (a) Radiated power of the three-bit stack versus along-track direction and frequency offset for four terminals on the ground track (crosses: terminal directions and tones). (b), (c) Median worst-terminal SINR versus footprint radius and versus aperture size (, , ; drops). (d) Median broadcast spectral efficiency versus . (e) Grouped ST-SIM for : spectral efficiency with and without terminal-side correction of the in-group residual, and frequency hypotheses. (f) Worst-terminal SINR at dB versus terminal-position error, with the prediction of Propositions (?) and (?).
What Time Modulation Adds to a SIM
Figure 3a shows the capability that distinguishes the ST-SIM from a static SIM: each terminal direction receives its own tone in a beam nulled toward the other terminals, i.e., in (17); at each terminal, the other tones are at least dB down, so the -dB in-band SIR of the three-bit stack is set by quantization and switching noise, not by inter-terminal leakage. Along the pass, Benchmark 1 attains an average SIR of only to dB, since the common-translation residual reaches kHz () near zenith, whereas the ST-SIM keeps – dB (continuous) and – dB (three bits).
Figure 3b shows two regimes. In small footprints, the Doppler span is small, and the terminals are not spatially resolvable: Benchmark 1 performs well ( dB at km), whereas the ST-SIM and Benchmark 2, which must separate the terminals to give them different frequencies, pay the penalty of Proposition 1 and Lemma 3 ( dB at km for the ideal field; the approximate realizations of the stack lie – dB below it, trading leakage against energy). The single-layer aperture, whose interference floor lies near dB in the baseline geometry, stays and dB below the four-layer stack at and km, the gain of stacking at equal radiated peak energy, phase resolution, and control rate. The SINR-optimal translation leaves Benchmark 1 unchanged up to km and raises it by , , and dB at , , and km, where it exploits the zeros of the ICI law, still more than dB below the ST-SIM. The curves cross at about km; beyond it, Benchmark 1 collapses, while the continuous-phase ST-SIM stays within dB of Benchmark 2 in median from km on ( dB at km and dB at km, – dB less with three bits). The long-run peak-to-average ratio of the ideal field, paid by every scheme, is dB in median at km and dB at km. Benchmark 3 attains – dB at every radius, i.e., dB more than the ST-SIM at km, for two reasons inherent to per-terminal translation at the payload, digital or electromagnetic: its static field has no peak-to-average penalty and delivers a common gain to all terminals instead of separating them; it presumes, however, that every terminal has acquired and tracks a residual of several subcarrier spacings. At equal feed power, where every stack-based field is divided by the fitted gain of (26) and Benchmark 2 is unchanged, the stack-based schemes lose about dB to the lossless Benchmark 2 (), the simulated diffraction loss of the normalized cascade model: in the baseline geometry, Benchmark 2 attains dB, Benchmark 3 dB, and the ST-SIM and dB with continuous and three-bit phases, the latter reflecting a smaller fitted gain rather than a benefit of quantization (medians over the drops at km: , , and dB for Benchmark 3, the ST-SIM, and Benchmark 2). The lossless single layer () reaches dB at equal feed power in the baseline geometry, and the four-layer stack overtakes it from dB on, where the interference floor of the single layer dominates. The optimal depth at equal feed power is smaller: in the baseline geometry, the equal-feed worst-terminal SINR at dB is dB for , dB for , dB for , dB for , and dB for , whereas to coincide within dB at equal radiated energy; two layers, whose -dB synthesis error lies below the noise at these SNRs, overtake the single layer from dB and are the best equal-feed choice in the normalized model, where each additional layer costs about dB of diffraction loss. The near-digital performance thus holds at equal radiated peak energy only, and the single-layer comparison is conditional on the SNR, on the geometry, and on the diffraction loss of the model. The baseline dB is optimistic for current high-power amplifiers and Ka-band link budgets [9]; at dB, the median worst-terminal SINR at km is dB for the ST-SIM and Benchmark 2 and dB for Benchmark 3, and the crossover with Benchmark 1 moves to about km, which calls for the larger apertures of § 5.4.
Aperture Size and Number of Terminals
Fig. fig:aperture sweeps the aperture from to meta-atoms at km. The static Benchmark 1 is limited by the Doppler floor at dB and does not improve with , and Benchmark 3 stays at – dB, whereas the ST-SIM gains dB from to , because a larger aperture resolves the terminals with less modal energy (Proposition 1); with the continuous-phase stack tracks Benchmark 2 within dB for every , three bits cost – dB, falls – dB short, and brings no further gain.
The same mechanism governs Fig. fig:sumrate: the broadcast spectral efficiency of the continuous-phase ST-SIM tracks Benchmark 2 within bit/s/Hz for every with a single RF chain, three bits cost at most bit/s/Hz, and Benchmark 1 collapses for , and the single-layer aperture loses bit/s/Hz to the four-layer stack at . Benchmark 3 without pilot overhead reaches bit/s/Hz at and exceeds the ST-SIM ( bit/s/Hz) for , since a static beam does not pay the separation cost that grows with ; with two pilot symbols per slot it drops to bit/s/Hz, and the ST-SIM leads unless CFO tracking costs less than , , and of the resources for , , and . Time-division beam hopping serves each terminal only during its dwell: bit/s/Hz at .
Grouped Compensation and Robustness
Remark 1 is evaluated with contiguous partitions of the Doppler-ordered terminals, each group receiving the midpoint of its Doppler range. In Figure 3b, adaptive grouping selects for every drop the partition that maximizes the spectral efficiency predicted from the ideal grouped field, which the stack then synthesizes without terminal-side correction. It follows Benchmark 1 in small footprints ( in all drops up to km and in at km) and the per-terminal ST-SIM in large ones ( in of the drops at km), and exceeds both in between: at km, where it mostly selects or , its median SINR of dB lies dB above the per-terminal ST-SIM, and at and km the gain is still and dB, so the mode switch at the crossover becomes a smooth transition. Fig. fig:grouping sweeps for terminals in a -km footprint with the partition that minimizes the largest in-group residual. Without terminal-side correction, the spectral efficiency grows from bit/s/Hz at (Benchmark 1) to bit/s/Hz at , since in-group residuals of several subcarrier spacings are fatal; with correction of the in-group residual, it decreases from bit/s/Hz at (Benchmark 3), because separating groups costs energy, while the frequency search of the terminal shrinks from to and hypotheses on a grid for and at and bit/s/Hz. For , reduces the search from to hypotheses at a loss in spectral efficiency. Grouping thus turns the choice between payload- and terminal-side correction into a continuum.
The SINR loss of the ST-SIM is about dB for terminal-position errors up to km, as for Benchmark 2 and as predicted by Propositions (?) and (?) (Fig. fig:robust), and dB for satellite-velocity errors up to m/s. The initial acquisition of an unregistered terminal is outside the scope of the ST-SIM, since no dedicated mode exists before its position is known (§ 2); for the reacquisition of a registered terminal and for tracking during a session, the ST-SIM delivers a carrier-Doppler-free signal, so that a terminal needs a single frequency hypothesis, whereas a common-translation payload leaves () hypotheses in median (95th percentile) at km and () at km; the oscillator offset adds the same range to all schemes.
Conclusion
This paper introduced space–time SIMs, which pair each spatial mode of a multilayer stack with its own frequency translation, and applied them to Doppler-resilient LEO broadcasting with a single RF chain. For the ideal aperture field, stacking convolves the temporal spectra of the layers (Lemma (?)), the minimum-energy space–frequency synthesis has a closed-form solution governed by the terminal geometry (Proposition 1), full column rank of the steering matrix is necessary and sufficient for decoupling terminal-specific translations (Theorem 1), grouping relaxes this condition (Remark 1), and any common translation, i.e., any static SIM or single-translation time-modulated surface, leaves half the Doppler span (Proposition 2). Algorithm alg:synthesis decreases the synthesis objective monotonically to stationary points, and explicit laws quantify quantization, control rate, and ephemeris errors. Under the ideal phase-only wave-domain model, a four-layer stack reached synthesis errors of dB with continuous and dB with three-bit phases, removed the Doppler floor of a static SIM above a footprint radius of about km, matched ideal digital per-terminal precompensation within dB and bit/s/Hz at equal radiated peak energy, and kept an in-band SIR of dB with three bits at a -MHz control rate, whereas a single time-modulated layer stalled at a synthesis error of dB, which identifies inter-layer mixing as the enabling element of the architecture.
The results also delimit the regime of the architecture. Separating terminals costs modal energy that grows with their number and shrinks with the aperture ( dB from to meta-atoms per layer). A static beam with terminal-side CFO correction attained a higher SINR and, with dedicated pilots, a higher broadcast spectral efficiency for three or more terminals; the ST-SIM led for two terminals when CFO tracking costs more than of the resources, and always delivered a carrier-Doppler-free signal that removes the frequency search of registered terminals, with grouped compensation providing a continuum between the two approaches. At equal feed power, each layer of the normalized cascade model costs about dB of diffraction loss, which favors two or three layers.
Light-Time Doppler Model
Let , where solves [8], and let be the unit vector from to . Since and , the light-time equation has a unique solution, is strictly increasing, and the retarded transmit time is well defined.
Proposition 8 (Light-time LEO Doppler). The instantaneous carrier Doppler frequency of (1) is so that .
For LEO motion the relative correction is at most , so the first-order form is used throughout.
References
- [1]Kodheli, O., Lagunas, E., Maturo, N., Sharma, S. K., Shankar, B., Montoya, J. F. M., Duncan, J. C. M., Spano, D., Chatzinotas, S., Kisseleff, S., Querol, J., Lei, L., Vu, T. X., & Goussetis, G. (2021). Satellite Communications in the New Space Era: A Survey and Future Challenges. IEEE Commun. Surveys Tuts., 23(1), 70–109. https://doi.org/10.1109/COMST.2020.3028247
- [2]Kumar, S., Sheemar, C. K., Querol, J., Yilmaz, T., Chatzinotas, S., Hammouda, M., Heyn, T., Schlichter, T., Marques, P., Pereira, L., et al. (2023). 5G NTN LEO based demonstrator using OpenAirInterface5G. IET Conference Proceedings CP873, 69–75.
- [3]Sheemar, C. K., Iacovelli, G., Khan, W. U., Solanki, S., & Chatzinotas, S. (2026). Beam Tracking for SIM-Assisted LEO Satellites: From Orbital Prediction to Wave-Domain Mobility Management. Engineering Archive. https://doi.org/10.31224/8514
- [4]Sheemar, C. K., Iacovelli, G., Solanki, S., Khan, W. U., & Chatzinotas, S. (2026). One-Shot Beam Tracking for Stacked Intelligent Metasurfaces-Assisted LEO Broadcasting. https://arxiv.org/abs/2610.05520
- [5]Iacovelli, G., Sheemar, C. K., & Chatzinotas, S. (2026). Stacked Fluid Metasurfaces: Mutual-Coupling-Aware Modeling and Optimization. https://arxiv.org/abs/2609.31591
- [6]Sheemar, C. K., Iacovelli, G., Solanki, S., Khan, W. U., & Chatzinotas, S. (2026). Stacked Intelligent Metasurfaces Assisted UAV Communications. https://arxiv.org/abs/2607.24374
- [7]Sheemar, C. K., Iacovelli, G., Solanki, S., Khan, W. U., Alexandropoulos, G. C., & Chatzinotas, S. (2026). Energy Efficient Multi-User Beamforming and 3D Position Optimization for SIM-Assisted UAVs. https://arxiv.org/abs/2608.13765
- [8]Montenbruck, O., & Gill, E. (2000). Satellite Orbits: Models, Methods, and Applications. Springer. https://doi.org/10.1007/978-3-642-58351-3
- [9]3GPP. (2023). Solutions for NR to Support Non-Terrestrial Networks (NTN) (Release 16). In 3rd Generation Partnership Project (3GPP) (Numbers TR 38.821 V16.2.0).
- [10]Ali, I., Al-Dhahir, N., & Hershey, J. E. (1998). Doppler Characterization for LEO Satellites. IEEE Trans. Commun., 46(3), 309–313. https://doi.org/10.1109/26.662636
- [11]You, L., Li, K.-X., Wang, J., Gao, X., Xia, X.-G., & Ottersten, B. (2020). Massive MIMO Transmission for LEO Satellite Communications. IEEE J. Sel. Areas Commun., 38(8), 1851–1865. https://doi.org/10.1109/JSAC.2020.3000803
- [12]Kodheli, O., Astro, A., Querol, J., Gholamian, M., Kumar, S., Maturo, N., & Chatzinotas, S. (2021). Random Access Procedure Over Non-Terrestrial Networks: From Theory to Practice. IEEE Access, 9, 109130–109143. https://doi.org/10.1109/ACCESS.2021.3101291
- [13]Lin, X., Lin, Z., Löwenmark, S. E., Rune, J., & Karlsson, R. S. (2021). Doppler Shift Estimation in 5G New Radio Non-Terrestrial Networks. Proc. IEEE Global Commun. Conf. (GLOBECOM), 1–6. https://doi.org/10.1109/GLOBECOM46510.2021.9685184
- [14]Kim, M.-G., & Jo, H.-S. (2022). Performance Analysis of NB-IoT Uplink in Low Earth Orbit Non-Terrestrial Networks. Sensors, 22(18), 7097. https://doi.org/10.3390/s22187097
- [15]Meshram, A. K., Kumar, S., Querol, J., Andrenacci, S., & Chatzinotas, S. (2025). Reduced Complexity Initial Synchronization for 5G NR Multibeam LEO-Based Non-Terrestrial Networks. IEEE Open J. Commun. Soc., 6, 1528–1551. https://doi.org/10.1109/OJCOMS.2025.3543625
- [16]Pollet, T., Bladel, M. V., & Moeneclaey, M. (1995). BER Sensitivity of OFDM Systems to Carrier Frequency Offset and Wiener Phase Noise. IEEE Trans. Commun., 43(2/3/4), 191–193. https://doi.org/10.1109/26.380034
- [17]Moose, P. H. (1994). A Technique for Orthogonal Frequency Division Multiplexing Frequency Offset Correction. IEEE Trans. Commun., 42(10), 2908–2914. https://doi.org/10.1109/26.328961
- [18]Hadani, R., Rakib, S., Tsatsanis, M., Monk, A., Goldsmith, A. J., Molisch, A. F., & Calderbank, R. (2017). Orthogonal Time Frequency Space Modulation. Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 1–6. https://doi.org/10.1109/WCNC.2017.7925924
- [19]Bemani, A., Ksairi, N., & Kountouris, M. (2023). Affine Frequency Division Multiplexing for Next Generation Wireless Communications. IEEE Trans. Wireless Commun., 22(11), 8214–8229. https://doi.org/10.1109/TWC.2023.3260906
- [20]Souza, P. H. C. de ., & Mendes, L. L. (2024). Doppler Shift Compensation Assisted by RIS Systems in High-Mobility Communications. IEEE Commun. Lett., 28(11), 2658–2662. https://doi.org/10.1109/LCOMM.2024.3466562
- [21]An, J., Renzo, M. D., Debbah, M., & Yuen, C. (2023). Stacked Intelligent Metasurfaces for Multiuser Beamforming in the Wave Domain. Proc. IEEE Int. Conf. Commun. (ICC), 2834–2839. https://doi.org/10.1109/ICC45041.2023.10279173
- [22]An, J., Xu, C., Ng, D. W. K., Alexandropoulos, G. C., Huang, C., Yuen, C., & Hanzo, L. (2023). Stacked Intelligent Metasurfaces for Efficient Holographic MIMO Communications in 6G. IEEE J. Sel. Areas Commun., 41(8), 2380–2396. https://doi.org/10.1109/JSAC.2023.3288261
- [23]An, J., Renzo, M. D., Debbah, M., Poor, H. V., & Yuen, C. (2025). Stacked Intelligent Metasurfaces for Multiuser Downlink Beamforming in the Wave Domain. IEEE Trans. Wireless Commun., 24(7), 5525–5538. https://doi.org/10.1109/TWC.2025.3547779
- [24]Nerini, M., & Clerckx, B. (2024). Physically Consistent Modeling of Stacked Intelligent Metasurfaces Implemented With Beyond Diagonal RIS. IEEE Commun. Lett., 28(7), 1693–1697.
- [25]Sheemar, C. K., Khan, W. U., Solanki, S., Alexandropoulos, G. C., & Chatzinotas, S. (2026). A Survey on Stacked Intelligent Metasurfaces: Fundamentals, Recent Advances, and Challenges. https://arxiv.org/abs/2603.05633
- [26]Lin, S., An, J., Gan, L., Debbah, M., & Yuen, C. (2024). Stacked Intelligent Metasurface Enabled LEO Satellite Communications Relying on Statistical CSI. IEEE Wireless Commun. Lett., 13(5), 1295–1299. https://doi.org/10.1109/LWC.2024.3368238
- [27]Yao, X., An, J., Gan, L., Renzo, M. D., & Yuen, C. (2024). Channel Estimation for Stacked Intelligent Metasurface-Assisted Wireless Networks. IEEE Wireless Commun. Lett., 13(5), 1349–1353. https://doi.org/10.1109/LWC.2024.3369874
- [28]Li, Z., An, J., & Yuen, C. (2026). Stacked Intelligent Metasurface-Enhanced MIMO OFDM Wideband Communication Systems. IEEE Trans. Wireless Commun., 25, 9608–9622. https://doi.org/10.1109/TWC.2025.3636393
- [29]Ranasinghe, K. R. R., Sandoval, I. A. M., Rou, H. S., Abreu, G. T. F. de ., & Alexandropoulos, G. C. (2026). Doubly-Dispersive MIMO Channels With Stacked Intelligent Metasurfaces: Modeling, Parametrization, and Receiver Design. IEEE Trans. Wireless Commun., 25, 3801–3817. https://doi.org/10.1109/TWC.2025.3605839
- [30]Darsena, D., Iudice, I., Galdi, V., & Verde, F. (2026). Randomized Space-Time Stacked Intelligent Metasurfaces for Massive Multiuser Downlink Connectivity. IEEE Trans. Commun., 74, 8059–8074. https://doi.org/10.1109/TCOMM.2026.3686695
- [31]Kummer, W. H., Villeneuve, A. T., Fong, T. S., & Terrio, F. G. (1963). Ultra-Low Sidelobes from Time-Modulated Arrays. IEEE Trans. Antennas Propag., 11(6), 633–639. https://doi.org/10.1109/TAP.1963.1138102
- [32]Hadad, Y., Sounas, D. L., & Alù, A. (2015). Space-Time Gradient Metasurfaces. Phys. Rev. B, 92(10), 100304(R). https://doi.org/10.1103/PhysRevB.92.100304
- [33]Caloz, C., & Deck-Léger, Z.-L. (2020). Spacetime Metamaterials---Part I: General Concepts. IEEE Trans. Antennas Propag., 68(3), 1569–1582. https://doi.org/10.1109/TAP.2019.2944225
- [34]Zhang, L., Chen, X. Q., Liu, S., Zhang, Q., Zhao, J., Dai, J. Y., Bai, G. D., Wan, X., Cheng, Q., Castaldi, G., Galdi, V., & Cui, T. J. (2018). Space-Time-Coding Digital Metasurfaces. Nat. Commun., 9, 4334. https://doi.org/10.1038/s41467-018-06802-0
- [35]Dai, J. Y., Zhao, J., Cheng, Q., & Cui, T. J. (2018). Independent Control of Harmonic Amplitudes and Phases via a Time-Domain Digital Coding Metasurface. Light Sci. Appl., 7, 90. https://doi.org/10.1038/s41377-018-0092-z
- [36]Zhang, L., Chen, M. Z., Tang, W., Dai, J. Y., Miao, L., Zhou, X. Y., Jin, S., Cheng, Q., & Cui, T. J. (2021). A Wireless Communication Scheme Based on Space- and Frequency-Division Multiplexing Using Digital Metasurfaces. Nat. Electron., 4(3), 218–227. https://doi.org/10.1038/s41928-021-00554-4
- [37]Ramaccia, D., Sounas, D. L., Alù, A., Toscano, A., & Bilotti, F. (2020). Phase-Induced Frequency Conversion and Doppler Effect With Time-Modulated Metasurfaces. IEEE Trans. Antennas Propag., 68(3), 1607–1617. https://doi.org/10.1109/TAP.2019.2952469
- [38]Taravati, S., & Eleftheriades, G. V. (2021). Pure and Linear Frequency-Conversion Temporal Metasurface. Phys. Rev. Appl., 15(6), 064011. https://doi.org/10.1103/PhysRevApplied.15.064011
- [39]Taravati, S., & Eleftheriades, G. V. (2022). Microwave Space-Time-Modulated Metasurfaces. ACS Photon., 9(2), 305–318. https://doi.org/10.1021/acsphotonics.1c01041
- [40]Sun, Y., Wu, G.-B., & Chan, C. H. (2025). A Versatile Shaped-Beam Space-Time-Coding Metasurface Antenna Based on 1-bit Phase Control. IEEE Trans. Antennas Propag., 73(11), 8876–8887. https://doi.org/10.1109/TAP.2025.3590874
- [41]Gholami, M., Khajavi, S., Neshat, M., Tewes, S., & Sezgin, A. (2026). Realization of Reconfigurable Intelligent Surfaces With Space--Time-Coded Metasurfaces. IEEE Trans. Antennas Propag., 74(5), 4281–4291. https://doi.org/10.1109/TAP.2026.3660044
- [42]Lin, X., Rivenson, Y., Yardimci, N. T., Veli, M., Luo, Y., Jarrahi, M., & Ozcan, A. (2018). All-Optical Machine Learning Using Diffractive Deep Neural Networks. Science, 361(6406), 1004–1008. https://doi.org/10.1126/science.aat8084
- [43]Barzilai, J., & Borwein, J. M. (1988). Two-Point Step Size Gradient Methods. IMA J. Numer. Anal., 8(1), 141–148. https://doi.org/10.1093/imanum/8.1.141
- [45]Zhu, X., Qian, C., Zhang, J., Jia, Y., Xu, Y., Zhao, M., Zhao, M., Qu, F., & Chen, H. (2024). On-demand Doppler-offset beamforming with intelligent spatiotemporal metasurfaces. Nanophotonics, 13(8), 1351–1360. https://doi.org/10.1515/nanoph-2023-0569
- [46]Movahediqomi, M., Tretyakov, S., Asadchy, V., & Wang, X. (2026). Stacked Time-Varying Metasurfaces. Adv. Opt. Mater., 14(4), e02831. https://doi.org/10.1002/adom.202502831
- [47]He, C., Liang, X., Zhou, B., Geng, J., & Jin, R. (2015). Space-division multiple access based on time-modulated array. IEEE Antennas Wireless Propag. Lett., 14, 610–613. https://doi.org/10.1109/LAWP.2014.2373431
- [48]Castaldi, G., Zhang, L., Moccia, M., Hathaway, A. Y., Tang, W. X., Cui, T. J., & Galdi, V. (2021). Joint multi-frequency beam shaping and steering via space--time-coding digital metasurfaces. Adv. Funct. Mater., 31(6), 2007620. https://doi.org/10.1002/adfm.202007620
- [49]Wu, G.-B., Dai, J. Y., Cheng, Q., Cui, T. J., & Chan, C. H. (2022). Sideband-free space--time-coding metasurface antennas. Nat. Electron., 5(11), 808–819. https://doi.org/10.1038/s41928-022-00857-0









