A reconfigurable intelligent surface is a flat panel of tiny programmable elements that steers radio waves. Ordinary versions only reflect, so they serve only the half of space in front of them. Newer "transmit-and-reflect" surfaces (STAR-RIS) fix this blind spot by splitting each incoming wave into a reflected part and a part that passes through. Today's beam-steering algorithms describe each element by one fixed pair of numbers: a reflection coefficient and a transmission coefficient. That description is only accurate at a single frequency. Real 5G/6G links occupy wide bands, and each element's response drifts across the band. The drift is worst for the transmitted beam, the very feature that removes the blind spot. This proposal borrows a tool from particle physics to handle that drift.
Particle physicists faced the same kind of problem when describing how quarks bind into mesons as a function of energy transfer. They stopped fitting hand-built resonance models, which turned out to be biased, and switched to a short power series in a remapped variable. The coefficients of that series are capped by a conservation law, which gives a fit with a provable error bar. The same mathematics applies to any passive device. Passivity means the element cannot create energy, so reflected plus transmitted power can never exceed incident power. That rule caps the series coefficients in exactly the way the physics conservation law does. The result is a compact description of each element's wideband response, with a guaranteed bound on what was left out. The bound can be fed directly into beamforming and channel estimation.
Core Claim
Model each STAR-RIS element's pair of frequency responses $(R(\omega), T(\omega))$ as a short series in a conformal variable. A conformal map here is a change of variable that squeezes the whole upper half of the complex frequency plane into the unit disk, so the operating band lands in a small region near the center. Passivity then supplies a hard cap on the coefficients. Three things follow:
- Fits that use a handful of coefficients predict held-out frequencies more accurately than single-resonance circuit models.
- The fits come with truncation error bars that actually cover the truth.
- Beamforming that is robust to the bounded leftover gives better worst-subcarrier rates in both half-spaces than designs that assume a frequency-flat response.
The particle-physics literature also supplies a cautionary lesson about degeneracies, which shapes how we design the channel-estimation experiments.
The mechanism has two parts. Causality makes $R$ and $T$ analytic in the upper half frequency plane. Pick a complex anchor $\omega_0 = \omega_c + i\gamma$, with $\omega_c$ the band center and $\gamma$ a width chosen so that the band maps to small $|z|$, and define
$$z(\omega) = \frac{\omega - \omega_0}{\omega - \omega_0^*}.$$
This sends the upper half-plane to the unit disk. Write $R = \sum_n a_n z^n$ and $T = \sum_n b_n z^n$. Passivity says $|R|^2 + |T|^2 \le 1$ on the real frequency axis. Integrating that inequality around the circle gives
$$\sum_{n=0}^{\infty} \left(|a_n|^2 + |b_n|^2\right) \le 1.$$
This is the direct analogue of the unitarity bound $\sum a_n^2 \le 1$ used for meson form factors. Consequently, the error from keeping only terms up to order $N$ is at most
$$\sqrt{1 - \sum_{n \le N} \left(|a_n|^2 + |b_n|^2\right)} \cdot \frac{|z|^{N+1}}{\sqrt{1-|z|^2}}.$$
This follows from the Cauchy–Schwarz inequality on the tail of the series. The feasible coefficient set is a ball, which is convex. That convexity is what makes the bound usable inside an optimizer.
Grounding in Prior Work
The series-expansion side supplies both the method and the motivating failure. CLEO-c analyses of tagged $D \to \pi e\nu$ and $D \to K e\nu$ decays (Cronin-Hennessy et al., Ge et al.) found two problems with single-pole models. When the pole mass floated, the fitted "effective" pole moved away from the physical resonance mass. When it was fixed to the physical mass, the shape was wrong, a bias that more data does not remove. The bounded conformal expansion (Besson et al., Cronin-Hennessy et al.) fixed this with two or three terms. We reuse their construction almost unchanged and apply it to element responses rather than form factors.
The direct counterpart of the single-pole model in metasurfaces is the equivalent-circuit element. The coupled phase-shift model of Liu, Mu, Schober and Poor derives, from circuit impedances and energy conservation, that a passive lossless element must satisfy $\theta^t - \theta^r = \pm\pi/2$ or $\pm 3\pi/2$. That paper constrains our proposal: any expansion we fit must recover this coupling in the lossless, single-frequency limit. Our proposal differs from it in two ways. It treats the coupling as something to test across the band rather than impose. It also allows loss and dispersion, which a single-frequency lossless circuit cannot represent.
The hadronic-uncertainty literature contributes a methodology rather than a mechanism. Ciuchini et al. and Ciuchini et al. compared three ways of treating an unknown hadronic contribution in rare B-meson decays: a model-dependent calculation, a model plus broad error margins, and a fully data-driven polynomial. They did this jointly with the parameters of physical interest. They found that the data-driven polynomial can absorb a shift in the target parameter $C_9$, because that shift is degenerate with the leading polynomial term. In the RIS setting, the same trap appears in channel estimation. A constant offset in an element's response amplitude is degenerate with the gain of the cascaded base-station-to-surface-to-user channel. So an unbounded expansion fitted together with channel parameters will bias channel estimates. We adopt their three-scheme comparison as our experimental design. The passivity cap is our proposed way out of the degeneracy, playing the role that prior bounds on expansion coefficients would play in the flavor fits.
The STAR-RIS literature supplies the application regime and the evaluation harness. Liu et al. define the energy-splitting, time-switching and mode-switching protocols and the electric/magnetic surface-current picture that makes transmission possible. Wu et al. provide the channel-estimation setup that we extend: 80 elements in 20 sub-surfaces, Rician fading, and DFT training patterns. The element-wise sum-rate algorithm of Syed et al. is the optimizer we would make robust. None of these works models frequency dependence of the element response. That gap is what we target.
Research Program
Unit-cell response data. The first stage builds a dataset of element responses. We will run full-wave electromagnetic simulations of two STAR-RIS unit-cell families: a lossless symmetric design and a lossy varactor-tuned design. Each will be swept over its control states and over a band of roughly 10–20% fractional bandwidth around a millimeter-wave and a sub-6 GHz carrier. The file lists no public dataset of STAR-RIS frequency responses, so this data must be generated. We also treat the 2.58–2.63 GHz filtering active prototype of Song et al. as a stress case. Because that design is active, the passivity bound must be replaced by a gain bound $|R|^2 + |T|^2 \le G$, and we will check whether the guarantees degrade gracefully.
Fitting machinery and physics validation. The second stage builds the fitting code: a constrained least-squares or Bayesian fit of $(a_n, b_n)$ inside the passivity ball, with truncation order chosen by held-out error. Before touching wireless data, we validate the code by refitting published CLEO-c form-factor shapes and checking that two-to-three-term bounded fits reproduce the qualitative conclusions of Ge et al.. This is a correctness check, not a contribution.
Robust beamforming. The third stage inserts the fitted expansions into wideband OFDM beamforming, where the band is divided into many narrow subcarriers. The truncation ball defines an uncertainty set per element. The element-wise optimizer of Syed et al. is modified to maximize the worst-case rate over that set, separately for users in the reflection and transmission half-spaces.
Channel estimation. The fourth stage reworks cascaded-channel estimation in the Wu et al. setting. The element response is shared across subcarriers through the expansion coefficients. The estimator runs three ways: with the coefficients fixed from calibration, with them floating but passivity-capped, and with them floating and uncapped. This mirrors the three-scheme design of Ciuchini et al.
Baselines
The naive reference point is the frequency-flat model: each element's response is taken at band center and used on every subcarrier, with the coupled phase constraint of Liu, Mu, Schober and Poor. This is how essentially all current STAR-RIS optimization works. Beating it should be easy at band edges, so the bar is a gain in worst-subcarrier rate that survives once the band is narrowed to a few percent.
The physically motivated competitor is a single-resonance (Lorentzian) equivalent-circuit model whose resonance frequency and quality factor are fitted per control state. This is the exact analogue of the floating-pole model that failed for D decays. It gets the same simulation data, the same number of free parameters as our two-term expansion, and the same held-out frequencies. Beating it means lower held-out prediction error at equal parameter count, plus evidence that its fitted resonance drifts away from the eigenmode frequency extracted directly from the full-wave solver.
The strongest generic competitor is rational approximation with passivity enforcement, often called vector fitting, which electromagnetic engineers routinely use for broadband models. The file contains no paper on it, and we have found no reported result on Lacuna for vector fitting of STAR-RIS elements. We will implement it with the same order budget and the same passivity enforcement. Beating it on point accuracy is not the claim. The claim is better-calibrated error bars and a convex uncertainty set that improves robust beamforming.
A flexible interpolation baseline, a cubic spline through the training frequencies, tests whether structure matters at all. Two ablations isolate the key ingredient. One removes the passivity cap, giving an unbounded expansion. The other removes the conformal map and expands directly in frequency, the way the B-decay fits expand in $q^2$.
For the system-level baselines, the closest reported numbers in the file come from different settings. Wu et al. report that the minimum channel-estimation error under energy splitting is about twice that under time switching, and that the coupling-aware energy-splitting scheme nearly matches the unconstrained bound. We first reproduce that factor of two under a flat response, then measure how much it changes once frequency dependence is introduced. No reported result was found on Lacuna for wideband STAR-RIS sum rate or estimation error with frequency-dependent elements, so our baseline numbers will come from our own reruns at matched compute and tuning budgets.
Why not just…?
Why not just use a measured lookup table with interpolation? Interpolation between calibration frequencies is accurate where data is dense. It says nothing about the error between or beyond calibration points, and it gives no convex set to optimize against. The spline baseline measures this directly. If splines match our held-out accuracy and robust beamforming built on an empirical spline error gives the same worst-subcarrier rate, the bounded expansion adds nothing, and we would report that.
Why not just use better equivalent-circuit models with more resonances? Multi-resonance circuits will fit better than one resonance. However, the D-meson experience (Cronin-Hennessy et al.) shows that the fitted resonances become effective parameters rather than physical ones, and fitting pole positions is non-convex. We test the single-resonance version head-on with the resonance-drift check. If a two-resonance circuit gives calibrated errors comparable to ours, the circuit route is a legitimate alternative, and our advantage reduces to convexity of the uncertainty set.
Why not just use vector fitting, which already enforces passivity? This is the real threat. On pure point accuracy it may tie or win. Our distinguishing claim is narrower: the coefficient ball yields a tail bound with a known coverage guarantee and a convex constraint for robust optimization, whereas vector fitting yields a point model. The decisive experiment is the robust-beamforming comparison. Both models are given the same order. Vector fitting's error is estimated by bootstrap, our model's error comes from the truncation ball, and the two are compared on worst-subcarrier rate and on coverage of held-out errors. If the bootstrap-wrapped vector fit matches us, we will say so and recommend it.
Proposed Experiments
The decisive checks are four, ordered by how quickly each could kill the idea.
The first is coverage. Across all unit cells and control states, the fraction of held-out frequency points whose true response falls inside the truncation bound must meet or exceed the nominal level. Undercoverage would mean the passivity-derived bound is wrong in practice, for example because mutual coupling between neighboring elements breaks the single-element passivity picture.
The second is the resonance-drift test against the Lorentzian model, which checks whether the D-meson failure pattern recurs in metasurfaces.
The third is a coupling test. In the lossless symmetric family, the fitted expansions should reproduce $\theta^t - \theta^r = \pm\pi/2$ at every frequency. In the lossy family, we should be able to quantify how far the coupling departs from that value across the band.
The fourth is the degeneracy test in channel estimation. Under the uncapped floating scheme, we expect a measurable bias in estimated channel gain that the passivity cap removes. Comparing normalized mean square error and estimated-gain bias across the three schemes, under both energy splitting and time switching, will show whether the capped expansion recovers the factor-of-two gap reported by Wu et al..
The system-level payoff is measured last: worst-subcarrier and average sum rate for reflection-side and transmission-side users as the fractional bandwidth grows. The robust design should widen its lead over the frequency-flat baseline as the band widens, and the transmitted half-space should benefit most.
References
- A Study of the Decays D^0 --> pi^- e^+ nu_e, D^0 --> K^- e^+ nu_e, D^+ --> pi^0 e^+ nu_e, and D^+ --> barK^0 e^+ nu_e
- Study of D0→π-e+νe, D+→π0e+νe, D0→K-e+νe, and D+→K̄0e+νe in tagged decays of the ψ(3770) resonance
- Improved measurements of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>D</mml:mi></mml:math>meson semileptonic decays to<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>π</mml:mi></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>K</mml:mi></mml:math>mesons
- Study of the decays D-0 -> pi(-)e(+)nu(e), D-0 -> K(-)e(+)nu(e), D+ -> pi(0)e(+)nu(e), and D+ -> (K)over-bar(0)e(+)nu(e)
- Simultaneously Transmitting and Reflecting (STAR)-RISs: A Coupled Phase-Shift Model
- B → K∗ ℓ + ℓ − decays at large recoil in the Standard Model: a theoretical reappraisal
- Charming penguins and lepton universality violation in $$\varvec{b \rightarrow s \ell ^+ \ell ^-}$$ decays
- STAR: Simultaneous Transmission and Reflection for 360° Coverage by Intelligent Surfaces
- Channel Estimation for STAR-RIS-Aided Wireless Communication
- Sum-Rate Optimisation of a Multi-User STAR-RIS-Aided System with Low Complexity
- Design and Prototyping of Filtering Active STAR-RIS with Adjustable Power Splitting
- Study of D-0 -> pi(-)e(+)ve, D+ -> pi(0)e(+)ve, D-0 -> K(-)e(+)ve, and D+ -> (K)over-bar(0)e(+)ve in tagged decays of the psi(3770) resonance