{"id":"d1c1c9af-2e82-45ab-a75d-f8f5ae792750","arxiv_id":"2507.02168","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A closed-form plus gradient-descent technique estimates multiport-network parameters for multi-bit RISs in unknown environments; experiments show mutual-coupling-aware models are far more accurate but not much better for end-to-end optimization.","lead":"A hybrid estimator learns the parameters of a physics-based multiport-network model for programmable wireless surfaces directly from measurements, needing no knowledge of the surface or room. Tested on an 8-element 6-bit RIS, it yields accurate channel predictions, yet a simpler model without mutual coupling achieves nearly the same end-to-end performance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Step 2's SVD initialization is only exact when S_SS is diagonal; under mutual coupling the fixed S_RS/S_ST directions are not the ones needed, so the general MNT-estimation claim is not established.","rationale":"The reader's verdict accepts the paper on the strength of the experimental validation, and the reader's weakest assumption concerned the lumped-port linearity of the hardware. My reading agrees that the experiment is well executed and the unseen-configuration accuracy is credible, but I find a more internal problem. The closed-form SVD step, which is half of the 'hybrid' claim, is derived from the cascaded (zero mutual coupling) form of the model, even though the algorithm's purpose is to estimate mutual coupling. The Woodbury-identity calculation above shows the exact rank-one factor uses embedded vectors involving D^{-1}, not the bare columns of S_RS and S_ST. Because Step 3 only adjusts scalar weights on those fixed directions, the parameterization is not the full MNT model. The paper does not prove that this constrained family still contains a matching parameter set for arbitrary reciprocal S_SS, and the single experimental configuration cannot test that. The proposed synthetic sweep would settle the question cleanly: if the hybrid method's accuracy tracks the full gradient-descent fit across coupling strengths, the concern is resolved; if it degrades with coupling, the general claim should be qualified. I therefore recommend CONDITIONAL rather than a clean ACCEPT, while noting that the concern is about the generality and derivation of the algorithm, not about the honesty or value of the experimental results.","tokens_in":8179,"tokens_out":25472,"duration_ms":306568,"concrete_test":"Build a synthetic noiseless MNT dataset with a known reciprocal S-matrix (NS=8, NR=NT=4, realistic 64-state load reflection coefficients) and sweep the RMS off-diagonal magnitude of S_SS from -40 dB to -5 dB. For each level, run the paper's Step 1-2-3 pipeline with n=3000 and evaluate zeta on 3000 unseen configurations; also run a same-cost full gradient-descent fit without the SVD initialization. If the hybrid zeta falls below the full-fit zeta by more than about 10 dB as coupling increases, Step 2's fixed directions are the limiting factor and the central claim holds only for weak mutual coupling.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing point is Section IV.A, Step 2. From the paper's own equation (1), H = S_RT + S_RS(Phi^{-1} - S_SS)^{-1} S_ST, a change of only c_i gives Delta_i = H(c_i') - H(c_i) = -delta/(1 + delta (D^{-1})_{ii}) (S_RS D^{-1} e_i)(e_i^T D^{-1} S_ST), where D = Phi^{-1} - S_SS. This is exactly rank-one, but the left and right vectors are S_RS D^{-1} e_i and e_i^T D^{-1} S_ST, not S_RS_i and S_ST_i. The two sets are colinear only when S_SS is diagonal, i.e., when inter-element mutual coupling is absent. The paper's statement that 'inspection of (1) reveals' colinearity with S_tilde_RS_i and S_tilde_ST_i is therefore not correct for the MC-aware model. Step 3 then keeps the SVD-determined directions fixed, adjusting only the per-column scalars a_i, b_i, S_tilde_SS, and s_tilde; it never re-estimates the directions. Consequently, the fitted model is a constrained version of the full MNT model, and the paper does not show that a matching set of parameters can be obtained when off-diagonal S_SS is significant. The experimental validation demonstrates the method for one coupling level, but the general central claim is not supported by the derivation. Since the reported MNT-over-CASC gain is 17 dB, S_SS appears to be non-negligible in the validation, making this gap relevant rather than academic.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid two-stage procedure for estimating multiport-network-theory (MNT) parameters of a wireless channel that is parametrized by a multi-bit reconfigurable intelligent surface (RIS), without knowing the RIS hardware or the propagation environment. In Step 2, single-element control changes are used with a rank-one SVD to initialize the coupling vectors between antennas and RIS elements; in Step 3, gradient descent refines the remaining scaling factors, the mutual-coupling matrix, and the load states. The method is experimentally tested in a reverberation chamber with an eight-element, 6-bit RIS prototype, and is benchmarked against a mutual-coupling-unaware cascaded model and a linear regression model. The reported MNT model accuracy reaches 48.5 dB on 3000 unseen configurations, while the cascaded model reaches 31.2 dB; nevertheless, for the four considered end-to-end performance metrics, the cascaded model achieves nearly the same performance as the MNT model.","tokens_in":8573,"tokens_out":11499,"duration_ms":131657,"significance":"If the central claim held in full generality, the paper would fill a real gap: it would provide a practical recipe for estimating physics-consistent MNT parameters from end-to-end measurements in unknown rich-scattering environments, including for multi-bit RIS elements. The experimental validation on unseen configurations is a genuine strength and goes beyond many purely numerical studies. The paper is also careful to identify parameter ambiguities and to test the optimization consequences with directly measured channels, which is methodologically sound. The main significance is therefore contingent on whether the SVD-initialized model family really spans the full MNT model; as argued below, that point is not established and needs revision.","major_comments":[{"comment":"The claim that a single-element change produces a rank-one matrix whose singular vectors are colinear with the i-th column of S_RS and the i-th row of S_ST is not correct for the general MNT model. With D = Phi^{-1} - S_SS, a change delta in the i-th load gives, by the Woodbury identity, Delta_i = -[delta/(1+delta (D^{-1})_{ii})] (S_RS D^{-1} e_i)(e_i^T D^{-1} S_ST). These vectors are colinear with S_RS e_i and e_i^T S_ST only when D^{-1} is diagonal, which in general requires S_SS to be diagonal. Step 3 keeps the SVD-derived directions fixed and optimizes only the scalar factors a_i and b_i, together with S_SS and the load estimates; it never re-estimates the directions. The fitted model is therefore a restricted subclass of the full MNT model when off-diagonal mutual coupling is present. Since the reported accuracy gap between MNT and CASC is about 17 dB (Section V.B), the off-diagonal part of S_SS appears to be material in the validation, so this is not merely a hypothetical limitation. The manuscript should either restrict the claim to the case of diagonal S_SS (or to weak mutual coupling), allow the directions to be refined in Step 3, or provide a proof that the SVD-initialized family can still represent the MNT map for arbitrary reciprocal S_SS. As written, the derivation in Step 2 does not support the general claim that a matching set of MNT parameters is estimated.","section":"Section IV.A, Step 2"},{"comment":"The conclusion states that the paper 'developed and experimentally validated the first hybrid MNT parameter estimation technique', and the Remark in Section I asserts that the technique applies directly to beyond-diagonal RIS and stacked intelligent metasurfaces. Because of the Step 2 issue above, the experimental results validate only the SVD-initialized constrained model on one particular eight-element, 6-bit prototype at one operating frequency. The extension to other hardware classes, especially those with strong mutual coupling, is an assertion rather than a demonstrated consequence. A concrete way to strengthen the claim would be to repeat the estimation with the RIS elements at the two spacings reported (lambda0/2 and lambda0/4) and to show whether the same accuracy is achieved when the SVD directions are intentionally perturbed, or to derive the conditions under which the constrained family remains exact. Without such evidence, the general 'MNT parameter estimation' claim is broader than what the paper establishes.","section":"Section IV.A and Section VII"}],"minor_comments":[{"comment":"The notation tilde S_RS_i and tilde S_ST_i is confusing: one is a column of the left factor and the other is a row of the right factor. Please clarify the orientation explicitly, for example by writing (S_RS)_i and (S_ST)_{i,:}.","section":"Section IV.A, Step 2"},{"comment":"The SNR (65.9 dB) and stability (53.2 dB) metrics are described only in words. Please state whether these are averages over ports, over the two VNA cascades, or over a specific set of configurations, and report the spread if available.","section":"Section V.A"},{"comment":"The metric zeta is said to be 'floored at 0 dB' in the caption of Fig. 2a, but the flooring is not stated in the main text near Equation (3). Please add one sentence explaining that negative values are clipped to 0 dB.","section":"Section V.B, Equation (3)"},{"comment":"The optimization results are reported for a single representative setup and without variability information. Since the parameter estimation is repeated with five random seeds and the coordinate-descent initialization depends on the measured configurations, reporting the spread of the KPI values across seeds or across a few independent runs would make the small MNT-vs-CASC differences in Table I more convincing.","section":"Table I"},{"comment":"The claim that MC-unaware cascaded models 'can yield very good experimental performances' is based on one hardware configuration and one set of KPIs. The sentence should be tempered to reflect the limited experimental scope, for example by saying 'in the tested setup'.","section":"Section VI.C"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong experimental core and the unseen-configuration validation is commendable. The main issue is the Step 2 derivation: the SVD directions are not the physically correct MNT directions when off-diagonal mutual coupling is present, and Step 3 does not correct them. This makes the general 'MNT parameter estimation' claim larger than what is proven. The paper could be made publishable by re-scoping the claim to the constrained model, by adding a direction-refinement step, or by proving exact representability for the restricted family. In its current form, I cannot recommend acceptance without the authors addressing this point."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The bottom line: this is a genuinely useful experimental paper. It validates a hybrid closed-form and gradient-descent estimator for multiport-network parameters on an 8-element, 6-bit RIS in a reverberation chamber, and checks the estimated model against 3000 unseen configurations. The 48.5 dB accuracy on unseen data is the strongest part. The comparison against the MC-unaware cascaded model, with 17 dB lower accuracy but nearly the same end-to-end optimization performance, is an interesting negative result for the field.\n\nWhat is new: the SVD initialization plus gradient descent, and the first application to 6-bit element-wise programming. The experiments are careful: direct VNA measurements, multiple seeds, variation of NA and n, and the ambiguity discussion is honest.\n\nThe soft spots. The derivation in Step 2 is not correct as written. From eq. (1), Δ_i is rank-one, but its left and right vectors are S_RS D^{-1} e_i and e_i^T D^{-1} S_ST, not S_RS_i and S_ST_i, unless S_SS is diagonal. The paper claims \"inspection of (1) reveals\" colinearity with the columns of S_RS and S_ST, which is wrong under mutual coupling. The algorithm fixes those SVD-derived directions and only adjusts per-column scalars, S_SS, and s, so the fitted model is a constrained version of the full MNT model. The general claim that a matching set of MNT parameters is always obtainable is not established. The experiment here suggests the constraint did little harm, but the paper should say this explicitly and discuss when it would break.\n\nThe end-to-end conclusion (MC-unaware models are \"almost as good\") rests on one chamber, eight elements, and one frequency. That is enough to raise doubts, not to settle them. The paper phrases it carefully, but the conclusion is a hypothesis, not a proven fact.\n\nThere is no code or data, which weakens reproducibility. The KPI table has no error bars, though the seed-dependence is reported as minimal.\n\nWho this is for: RIS researchers interested in experimental parameter extraction and the practical value of MC-aware models. It deserves a serious referee. I recommend acceptance after the authors correct the Step 2 derivation, state the constraint on the model family, and add a sentence on the scope of the MC-unawareness claim.","headline":"Solid experimental study of MNT parameter estimation for multi-bit RIS, but the SVD step is mis-justified under mutual coupling and the MC-unawareness conclusion is setup-specific.","tokens_in":9094,"tokens_out":5878,"would_cite":true,"duration_ms":60427,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For an unknown multi-bit RIS in an unknown rich-scattering environment, a hybrid closed-form-plus-gradient estimate of the multiport-network parameters predicts unseen channel matrices to 48.5 dB accuracy.","keywords":["reconfigurable intelligent surface","multiport-network theory","mutual coupling","parameter estimation","virtual VNA","reverberation chamber","RIS optimization","MIMO"],"falsifier":"Directly measure the full scattering matrix $\\mathbf{S}$ and load states $\\mathbf{s}$ of the modular testbed with a 16-port VNA, compute the channel predictions from those physical parameters, and compare them with the predictions of the hybrid estimate on the same unseen configurations; a systematic discrepancy exceeding the 65.9 dB measurement SNR would show that the estimated matching set does not represent the physics, while agreement would confirm that the rank-one-plus-gradient decomposition captures the true mechanism.","tokens_in":7937,"feed_emoji":"📡","tokens_out":11026,"duration_ms":120731,"temperature":0.7,"pith_summary":"Physics-consistent models of RIS-parametrized wireless channels treat the environment as a multiport network whose tunable elements are variable loads, but using these models in practice requires knowing both the RIS design and the radio environment. This paper shows how to estimate a matching set of multiport-network parameters directly from end-to-end channel measurements in an unknown rich-scattering environment, using one reference measurement, $N_S$ single-element toggles, and $n$ extra random configurations. The estimation combines closed-form SVD steps with gradient descent and handles multi-bit (here 6-bit) RIS elements whose load states are unknown; the resulting model reaches 48.5 dB accuracy against 3000 unseen measured configurations. When the estimated model is used to optimize RIS configurations, it clearly outperforms random tuning on four communications KPIs, although a much simpler mutual-coupling-unaware cascaded model achieves nearly the same optimized performance despite being about 17 dB less accurate. The practical consequence is that physics-consistent channel models for reconfigurable intelligent surfaces can be obtained without design knowledge or environment simulation, and that ignoring mutual coupling may be acceptable for some performance-oriented tasks.","feed_headline":"Unknown 6-bit RIS gets mapped to 48.5-dB accuracy","feed_subtitle":"The hybrid closed-form plus gradient estimate needs no RIS design knowledge and predicts 3000 unseen channel matrices at 48.5 dB.","key_machinery":"The load-bearing object is the multiport-network expression $$\\mathbf{H} = \\mathbf{S}_{RT} + \\mathbf{S}_{RS}\\bigl(\\mathbf{\\Phi}^{-1} - \\mathbf{S}_{SS}\\bigr)^{-1}\\mathbf{S}_{ST},$$ with $\\mathbf{\\Phi}=\\mathrm{diag}(\\mathbf{c})$ the diagonal matrix of tunable load reflection coefficients and $\\mathbf{S}$ the scattering matrix of the static part of the environment. Its role is to compress all structural scattering and environmental scattering into a finite set of parameters whose number grows with the number of antennas and RIS elements, not with the complexity of the RIS design or environment. The estimation engine is a three-step hybridization: one reference measurement sets $\\tilde{\\mathbf{S}}_{RT}$; toggling the $i$th RIS element yields a rank-one channel difference whose dominant left and right singular vectors capture, up to scale, the $i$th columns of $\\tilde{\\mathbf{S}}_{RS}$ and $\\tilde{\\mathbf{S}}_{ST}$; and gradient descent then solves for the scales $\\mathbf{a}, \\mathbf{b}$, the symmetric coupling matrix $\\tilde{\\mathbf{S}}_{SS}$, and the load reflection coefficients $\\tilde{\\mathbf{s}}$. The load vector $\\tilde{\\mathbf{s}}$ is learned rather than assumed, which is what allows multi-bit elements—more than two states per element cannot be handled by arbitrarily fixing $\\tilde{\\mathbf{s}}$. Forward evaluations during optimization are accelerated with the Woodbury identity.","core_discovery":"The paper's central claim is that a matching set of multiport-network-theory (MNT) parameters—an $N \\times N$ scattering matrix $\\mathbf{S}$ for the static parts of the RIS-parametrized environment together with the set of accessible load reflection coefficients $\\mathbf{s}$—can be estimated from ordinary end-to-end channel measurements without knowing the RIS element design or the propagation environment. Because only the mapping from control word $\\mathbf{w}$ to channel $\\mathbf{H}$ matters for optimization, the estimate need not be physically unambiguous; any parameter set that lets (1) reproduce the measured channels is sufficient. The estimation proceeds by (i) taking one reference channel measurement, (ii) for each RIS element, measuring the channel after toggling only that element and using the rank-one SVD structure $\\Delta_i = \\mathbf{U}_i \\Sigma_i \\mathbf{V}_i^\\dagger$ to fix the coupling vectors $\\tilde{\\mathbf{S}}_{RS,i}$ and $\\tilde{\\mathbf{S}}_{ST,i}$ up to scalar factors, and (iii) running gradient descent over those scalars, the symmetric mutual-coupling matrix $\\tilde{\\mathbf{S}}_{SS}$, and the load states $\\tilde{\\mathbf{s}}$ on $n$ additional random configurations. The author experimentally validates the approach at 2.45 GHz with eight 6-bit-programmable RIS elements of unknown design in a reverberation chamber, reporting an average accuracy of 48.5 dB for the MNT model against 3000 unseen measured configurations, versus 31.2 dB for a mutual-coupling-unaware cascaded model and below 20 dB for linear regression. The same estimated models are then used with coordinate descent to optimize four key performance indicators, and the optimized configurations are verified experimentally; the MNT and cascaded models yield nearly equal end-to-end gains, whereas linear regression performs clearly worse.","pith_inferences":["Because the paper's experiments use a moderate-coupling eight-element array, the near-equivalence of MNT and CASC may not hold in stronger-coupling regimes; a testable extension is to repeat the KPI comparison with denser arrays, beyond-diagonal RIS, or stacked intelligent metasurfaces, where mutual coupling plays a larger role.","The rotation of $\\tilde{\\mathbf{s}}$ across random seeds visible in the paper's Fig. 2 suggests that the learned load states are not physical measurements; any use of the estimated parameters should treat them as an equivalent model, not as hardware diagnostics.","Since Step 2's rank-one structure does not depend on the environment's complexity, the same estimator could in principle calibrate RIS-aided channels in outdoor or indoor line-of-sight settings, but that generalization goes beyond the paper's reverberation-chamber validation.","The large model-accuracy gap versus the small end-to-end performance gap implies that for system-level optimization a cheap cascaded model may suffice, whereas inverse problems or sensing tasks that rely on precise channel predictions may be where the accurate MNT model earns its extra calibration cost."],"forward_implications":["An off-the-shelf multi-bit RIS in an unknown environment can be characterized without any design knowledge: $1+N_S+n$ channel measurements yield a predictive model that maps any control word to the channel matrix.","Parameter ambiguities do not have to be resolved for optimization; any matching set that reproduces the measured configurations is usable, which is what makes the estimation feasible in practice.","Model accuracy saturates when $n$ reaches the order of the number of unknowns, and extra receive/transmit antennas help only up to a point—quantified gains from $N_A=2$ to 4 are large (about 15 dB for MNT) while $N_A=6$ and 8 add little.","The MC-unaware cascaded model, despite a 17 dB lower accuracy, delivers nearly the same optimized SISO and MIMO metrics as the MC-aware MNT model in the tested rich-scattering environment, while linear regression is clearly inferior.","The MNT model reaches 48.5 dB accuracy on unseen configurations, indicating that unobserved physical degrees of freedom of the 6-bit elements do not materially degrade generalization in this setup."],"supporting_citations":[{"why":"Supplies the multiport-network-theory channel model (Eq. 1) that the whole estimation targets.","marker":"[1]"},{"why":"The prior gradient-descent-only identification of an ambiguous MNT parameter set that this work extends and makes hybrid.","marker":"[12]"},{"why":"Shows that single-element perturbations produce rank-one channel differences; the SVD step borrows this robust retrieval of the coupling vectors.","marker":"[14]"},{"why":"Provides the closed-form Virtual VNA idea that motivates estimating some parameters without gradient descent.","marker":"[8]"},{"why":"Extends Virtual VNA estimation to non-reciprocal systems and informs the ambiguity discussion.","marker":"[9]"},{"why":"Evidence that gradient-descent estimates of the mutual-coupling block are more noise-robust than closed-form ones, justifying Step 3.","marker":"[10]"},{"why":"Gives the scattering-parameter MNT representation and the diagonal form used for beyond-diagonal RIS, supporting the generality remark.","marker":"[13]"},{"why":"Woodbury identity used to speed up forward evaluations of the MNT model during coordinate-descent optimization.","marker":"[15]"}],"fun_headline_variants":["Blind MNT estimation maps 6-bit RIS to 48.5 dB","Unknown RIS? Estimate its multiport model to 48.5 dB","8-element 6-bit RIS: blind model hits 48.5 dB","No design, no known environment: RIS modeled to 48.5 dB","Multiport-network params learned blindly for 6-bit RIS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the static parts of the RIS-parametrized environment, including the commercial antennas and phase-shifter hardware, act as a linear, passive, reciprocal multiport network with lumped ports at 2.45 GHz, and that each RIS element's tunable state is fully captured by a single scalar complex reflection coefficient; if higher-order effects such as frequency dispersion, radiation-dependent loading, or phase-shifter nonlinearity break this lumped-port description, the estimated parameters may fail to generalize to configurations outside the training set.","fun_headline_variants_meta":{"raw":{"variants":["Blind MNT estimation maps 6-bit RIS to 48.5 dB","Unknown RIS? Estimate its multiport model to 48.5 dB","8-element 6-bit RIS: blind model hits 48.5 dB","No design, no known environment: RIS modeled to 48.5 dB","Multiport-network params learned blindly for 6-bit RIS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3476,"prompt_tokens":1086,"completion_tokens":2390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":702,"completion_tokens_details":{"reasoning_tokens":2292}},"tokens_in":702,"tokens_out":2390,"duration_ms":18774,"temperature":1.0,"reasoning_tokens":2292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:36:22.108030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly measure the full scattering matrix $\\mathbf{S}$ and load states $\\mathbf{s}$ of the modular testbed with a 16-port VNA, compute the channel predictions from those physical parameters, and compare them with the predictions of the hybrid estimate on the same unseen configurations; a systematic discrepancy exceeding the 65.9 dB measurement SNR would show that the estimated matching set does not represent the physics, while agreement would confirm that the rank-one-plus-gradient decomposition captures the true mechanism.","supporting_citations":[{"cited_title":"Experimentally realized physical-model-based frugal wave control in metasurface-programmable complex media,","cited_arxiv_id":null,"evidence_quote":"The prior gradient-descent-only identification of an ambiguous MNT parameter set that this work extends and makes hybrid."},{"cited_title":"Virtual VNA 2.0: Ambiguity-free scattering matrix estimation by terminating not-directly-accessible ports with tunable and coupled loads,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-form Virtual VNA idea that motivates estimating some parameters without gradient descent."}],"review_version":1}