{"id":"a6c35d02-852f-4d8b-897a-879eb494766d","arxiv_id":"2602.00140","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Mutual information between applied loads and discrete sensors provides a task-agnostic metric for information propagation in elastic solids, tunable by geometry.","lead":"A solid elastic body is modeled as an information channel: inputs are applied loads, outputs are sensor reaction forces, and mutual information quantifies how much load information survives. The measure tracks Saint-Venant decay in a half-space and shows architected pore/slit geometries can enhance or block information flow.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The NMI saturation claim rests on an estimator artifact: h(X) is computed from joint X–Y k-nn statistics, so the denominator shrinks as sensors are added, inflating I/h(X) toward 1.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing defect: the NMI denominator is not an input intrinsic quantity because the LDDP/KSG estimator computes h(X) from joint X–Y statistics. This directly undermines the headline results: NMI≈1 at k=dx, the geometry comparison table, and the optimization objective. My independent reading confirms that the paper's own Appendix C shows h_est(X) decreasing with k while I(X;Y) plateaus, which is exactly the signature of denominator bias. This is not a disagreement with consensus but an internal inconsistency: the claimed bound NMI∈[0,1] is violated by their own Figure 6 for k>dx, and the saturation claim is therefore not established. The proposed concrete test isolates the bias without requiring full reimplementation: because the true mutual information for a rank-3 linear map is unchanged by appending duplicate sensors, any shift in the estimated NMI is attributable to the estimator. I recommend keeping the conditional verdict: the framework and qualitative trends (e.g., Saint-Venant decay, stress-line correlation) remain plausible, but the quantitative NMI values and 'theoretical maximum' statements require a corrected entropy estimator or independent validation before they can be accepted.","tokens_in":28916,"tokens_out":5136,"duration_ms":57755,"concrete_test":"Construct a synthetic linear encoder with dx=3: X∼U([-10,10]^3), Y=A X with A a fixed full-rank 3×3 matrix, so the true NMI is well-defined (for the true input entropy h(X)=3 log 20). Run the paper's estimator on this system with k=3 sensors and compute NMI. Then append three exact duplicate columns to Y, so dy=6 but the map remains rank-3 and the true I(X;Y) is unchanged. If the estimated NMI increases above its k=3 value (or exceeds 1), or if h_est(X) decreases, the reported saturation is an artifact of the joint-space normalization. Additionally, run the same test with k=dx=6 for a geometry from Section 3.2 and compare NMI to the corresponding k=6 value from the paper's pipeline to quantify how much of the geometry effect is denominator drift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim—that greedy selection with k=dx sensors yields NMI≈1, interpreted as the elastic body transmitting the full information content of the load—depends on the denominator h(X) in NMI = I(X;Y)/h(X). But h(X) is not an input-intrinsic entropy. In the LDDP/KSG implementation, the marginal invariant measures are derived from the joint-space k-nn statistics (Eqs. 33–37, Appendix A.3), with m(x) explicitly built from ⟨ρ^{dx+dy}⟩ (Eq. 36). When X and Y are strongly dependent, this violates the factorized-measure assumption m(x,y)=m(x)m(y) (Eq. 34), which the authors admit holds only if X and Y are independent. Consequently, h_est(X) depends on the sensor configuration, the number of sensors k, and the geometry via Y. The paper's own Appendix C (Fig. 6) demonstrates the consequence: for k>dx, I(X;Y) plateaus but h_est(X) continues to decrease, so NMI=I/h_est(X) rises solely because the denominator shrinks. This contradicts the claim NMI∈[0,1] and indicates the k=dx saturation is at least partly an estimator artifact. Moreover, in Sections 3.2 and 3.3, h_est(X) is recomputed under different geometries with different Y distributions; thus the reported NMI differences (e.g., 0.709 vs. 0.197) conflate genuine changes in mutual information with changes in the bias of the denominator. Without a correction or an independent estimator for h(X), the quantitative comparisons of geometries and the 'theoretical maximum' interpretation are not reliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an information-theoretic framework for mechanical information transmission: an elastic body is treated as an encoder that maps an applied traction X (parameterized by Legendre coefficients) to noiseless sensor reaction forces Y, and transmission efficiency is quantified by the normalized mutual information NMI = I(X;Y)/h(X). In an elastic halfspace, the authors connect single-sensor mutual information maps to Saint-Venant's principle, propose a greedy conditional-mutual-information sensor selection algorithm, report that k=dx sensors reach the 'theoretical maximum' NMI=1, and complete the channel with a neural-network decoder evaluated with rate-distortion theory. They then use NMI to compare architected geometries (pores and slits) and to drive Bayesian optimization of elliptical voids, reporting normalized mutual information values from about 0.197 to 0.709 (Table 1). The paper includes analytical Flamant-based stress fields, FEniCSx finite-element simulations, a mesh refinement study, and an open-source repository with pytest tests.","tokens_in":29292,"tokens_out":4800,"duration_ms":54339,"significance":"If the normalization were sound, this would be a useful task-agnostic benchmark for mechanical intelligence, with appealing connections to classical mechanics (Saint-Venant's effect, principal stress lines) and to design optimization. Strengths include the clean noiseless-channel identity I(X;Y)=h(Y), the reproducible implementation with publicly available code and tests, the analytical halfspace derivation, mesh refinement checks, and the authors' explicit acknowledgement of the restrictive condition behind their entropy estimator. However, the central normalization relies on a joint-space invariant measure and is therefore configuration-dependent; the quantitative claims and geometry comparisons need substantial rework before they can be taken at face value.","major_comments":[{"comment":"The normalization h(X) in NMI=I(X;Y)/h(X) is not an input-intrinsic entropy. The marginal invariant measure m(x) is constructed from joint-space k-nn statistics, specifically from ⟨ρ^{dx+dy}⟩ in Eq. (36), so h_est(X) changes with the sensor set Y, the number of sensors, and the domain geometry. The paper itself states in Appendix C (Fig. 6, right) that for k>dx, I(X;Y) plateaus while h(X) continues to decrease, so NMI=I/h(X) rises because the denominator shrinks. Moreover, the factorization m(x,y)=m(x)m(y) (Eq. 34) is admitted in Appendix A.3 to hold only when X and Y are independent, whereas the mechanical encoder is deterministic and strongly dependent. Thus the claim that k=dx sensors reach the theoretical maximum I(X;Y)/h(X)=1 (Sections 3.1.1, 3.2, 3.3) is at least partly an estimator artifact. Please recompute with a separate estimator of h(X) using only X samples, or use an analyti","section":"§2.2, Eq. (4); Appendix A.3, Eqs. (33)–(37); Appendix C, Fig. 6"},{"comment":"The cross-geometry comparisons are expressed as NMI=I/h_est(X), where h_est(X) is recomputed for each geometry using joint-space k-nn statistics that depend on Y. The reported spread (e.g., 0.709 vs. 0.197 in Table 1) therefore conflates genuine changes in mutual information with changes in the bias of the denominator. A fair quantitative comparison should report raw I(X;Y) alongside a separately estimated h(X), or use a normalization that is independent of the sensor/output distribution. Without this, the architected-material tuning claim and the Bayesian optimization objectives in Section 3.3 are not quantitatively supported as stated.","section":"§3.2, Fig. 4; Table 1; §3.3"},{"comment":"The saturation of raw I(X;Y) for k≥dx (Appendix C, left panel) is a useful and defensible numerical observation: additional greedily selected sensors do not increase the mutual information between load and sensor readings. However, this saturation does not by itself establish I(X;Y)=h(X); equality requires h(X) to be the true differential entropy of the load, which is exactly the quantity affected by the invariant-measure issue described above. Please either reframe the k=dx result as saturation of the raw mutual information, or provide an independent estimate of h(X) before claiming that the 'full information content' of the load has been transmitted.","section":"§3.1.1 and Appendix C"}],"minor_comments":[{"comment":"The Figure 2 caption says dx=3 and dx=6, while the text says dx=3 and dx=5 (and Section 3.1.2 uses dx=4). Please make the values consistent.","section":"Figure 2 caption vs. §3.1.1"},{"comment":"The symbol k denotes the number of sensors in Algorithm 1 and the main text, but denotes the KSG nearest-neighbor count in Appendix A.3. This is confusing because Section 3.1.1 compares k with dx while Appendix A.3 sets k=5. Please use distinct symbols (e.g., n_sensors and k_NN).","section":"Notation conflict"},{"comment":"Typo: 'in tehe ˆe2 direction' should be 'in the ˆe2 direction'.","section":"Appendix B.3"},{"comment":"'rate-disortion' should be 'rate-distortion'; 'principle stress lines' should be 'principal stress lines' in several places.","section":"§2.2.2 and elsewhere"},{"comment":"The noise-variance estimate O(1/N) is described as formally derived for the KSG estimator with the L2 distance metric, while the implementation uses L∞. Please clarify whether this mismatch affects the Bayesian optimization noise model.","section":"Appendix H"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the framework is promising, with strong reproducibility practices. However, the central normalization issue is load-bearing: the claimed achievement of a 'theoretical maximum' and the cross-geometry NMI comparisons are compromised by the configuration-dependent invariant measure. This is fixable in a revision by reporting raw mutual information alongside a properly estimated input entropy (or an analytic bound), but it is not merely a presentation problem. I would be willing to re-review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely new application: treating an elastic body as a channel from load coefficients to sensor forces and using KSG mutual information to measure how much of the load survives to the sensors. Second, the central quantitative claim—NMI near 1 with k=dx sensors, and geometry tuning NMI between about 0.2 and 0.7—sits on an entropy estimator whose normalization is not input-intrinsic. The paper's own Appendix A.3 admits the invariant measure factorizes only under independence, and Appendix C shows the NMI rise for k>dx is driven by h(X) shrinking. That problem does not disappear at k=dx: the architected-material comparisons use dx=6, k=6, but changing geometry changes the joint-space statistics and therefore the bias in the denominator. So I would treat the numerical NMI values as provisional.\n\nWhat is actually good: the Saint-Venant half-space validation is clean, and greedy sensor selection saturating the raw mutual information I(X;Y) is a real, intuitive result. The principal-stress-line visualizations give a nice qualitative bridge from information to load paths. They ship code with pytest tests, and the mesh refinement study is honest. The paper also states its own limitations clearly, which earns respect.\n\nWeak spots, in proportion: the normalization issue is load-bearing. Without an independent estimator for h(X) or at least error bars and a sensitivity check, the fine differences in Table 1 (0.709 vs 0.704, 0.213 vs 0.197) are not meaningful, and even the larger spread between geometries should be re-examined. The optimization results do not beat hand-designed structures; the authors know this and say it, so that is minor. The rate-distortion detour adds little, and the near-zero reconstruction error is discussed but still slightly odd.\n\nWho should read it: people in mechanical intelligence and metamaterials who want a benchmark metric, and information-theory folks who want to see their tools used in a new domain. It deserves a serious referee—the idea is worth pursuing and the issues are correctable—but I would want the normalization fixed, or at least very clearly caveated and backed by a second estimator, before trusting the numbers. Send it to review, with a request for substantial revision on the entropy estimation.","headline":"A useful framework for quantifying load-to-sensor information in elastic solids, but the NMI normalization is built from joint X–Y statistics, so the quantitative numbers are provisional until the entropy estimator is corrected.","tokens_in":29841,"tokens_out":4488,"would_cite":false,"duration_ms":50769,"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":"An elastic solid can be treated as an information encoder: the applied load is the input, the reaction forces at discrete sensors are the output, and normalized mutual information (NMI) measures how much of the load's information reaches th","keywords":["information theory","mutual information","mechanical intelligence","elastic solids","sensor placement","architected materials","rate-distortion theory"],"falsifier":"Recompute the halfspace NMI with a different estimator (e.g., a histogram or kernel method) on a known ground-truth load distribution and dense sensor sampling, and check whether I(X;Y) stays flat and NMI stays at 1 when the number of sensors exceeds the number of load coefficients; if h(X) drifts with sensor count or NMI drops, the saturation claim fails.","tokens_in":28703,"feed_emoji":"🧱","tokens_out":10595,"duration_ms":98904,"temperature":0.7,"pith_summary":"The paper proposes that an elastic solid, together with discrete force sensors, forms an information channel: the applied load is the input signal and the sensor readings are the output. Because the sensors are assumed noiseless, the mutual information between load and sensors equals the entropy of the sensor readings, and normalizing by the entropy of the load defines a quantity in [0,1] called normalized mutual information (NMI). The central claims are that greedily placed sensors—one per load coefficient—recover essentially all of the load's information (NMI≈1), that information decays with depth in a way consistent with the classical effect that statically equivalent loads become indistinguishable far from the load, and that architected geometry can tune NMI from roughly 0.2 to 0.7, letting structures either transmit or block information. If correct, this gives a task-agnostic, comparable metric for mechanical information processing and a design target for structures that process information through their own deformation.","feed_headline":"Solids can transmit or block mechanical information","feed_subtitle":"Normalized mutual information gauges how much of a load's signal reaches sensors; geometry tunes it from 0.2 to 0.7.","key_machinery":"The central object is the mechanical information channel and the normalized mutual information NMI=I(X;Y)/h(X). The input is a random load built from coefficients of an orthogonal polynomial basis; the encoder is the elastic body; the output is reaction forces at discrete sensors. With noiseless sensors, I(X;Y)=h(Y), so NMI lies in [0,1] and measures how close the body comes to transmitting the full load information. The greedy sensor-selection algorithm adds sensors one at a time by maximizing conditional mutual information, and the rule that the number of sensors equals the number of load coefficients emerges as the point where NMI saturates. Principal-stress lines act as the geometric cor","core_discovery":"On the paper's own terms, the discovery is that information propagation in an elastic solid has a precise information-theoretic signature. Treating the applied traction (parameterized by polynomial coefficients) as a random source X and the reaction forces at k discrete sensors as the output Y, the mechanical encoder satisfies I(X;Y)=h(Y) because the sensors are noiseless; the theoretical maximum is h(X). The paper defines NMI=I(X;Y)/h(X) and shows numerically that sequentially selecting sensors by conditional mutual information reaches NMI≈1 when k equals the number of load coefficients, so the body transmits the full information content of the load. It further shows that in a halfspace the","pith_inferences":["Editorial inference: the same NMI-style measure could be adapted to quantify mechanical memory by treating time-delayed states as the output, turning the framework from a sensing metric into a general mechanical-information metric.","Editorial inference: because the entropy estimator's normalization is built from joint load-sensor statistics, absolute NMI values may shift with sensor count and geometry; comparisons are safest within a fixed setup unless the normalization is independently validated.","Editorial inference: the load-path correlation suggests a concrete design rule—route principal-stress lines through sensors for high transmission—which could be tested by deliberately placing a void barrier between load and sensors and measuring the NMI drop.","Editorial inference: the fact that optimized structures do not beat simple slits suggests the design parameterization or optimizer, not the physics, is the current bottleneck; more expressive design spaces might approach the theoretical limits more closely."],"forward_implications":["If the framework is right, the information content of a mechanical load can be measured rather than assumed, and any elastic body can be benchmarked by a single number NMI.","Greedy sensor selection with as many sensors as load coefficients is sufficient to capture the full load information in the tested systems, so sensor placement can be guided by conditional mutual information.","Architected geometry can act as a tunable information filter: pore-like patterns block information (NMI≈0.2) while slit-like patterns transmit it (NMI≈0.7), giving a mechanical design route for information control.","The correlation between principal-stress lines and NMI suggests that load-path visualizations can guide designs for information-transmitting or information-blocking structures.","Rate-distortion results indicate that a mechanical encoder plus a learned decoder approaches the theoretical lower bound when sensors are chosen greedily, meaning the body behaves as a near-optimal lossy compressor of load information."],"fun_headline_variants":["Solids transmit or block info—geometry is the dial","Load signals in solids: geometry tunes information flow","How solids encode loads: geometry controls sensor info","Solids as info channels: geometry sets transmission","Mechanical intelligence: geometry gates load information"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole normalized measure rests on the assumption that the load's entropy can be estimated in a way that does not depend on the sensor setup; the estimator used here builds its normalization from joint load-sensor statistics, so this assumption is only exact when the load and sensor readings are independent, and it becomes questionable exactly in the high-correlation regime where the paper claims the theoretical maximum.","fun_headline_variants_meta":{"raw":{"variants":["Solids transmit or block info—geometry is the dial","Load signals in solids: geometry tunes information flow","How solids encode loads: geometry controls sensor info","Solids as info channels: geometry sets transmission","Mechanical intelligence: geometry gates load information"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1146,"prompt_tokens":687,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":387}},"tokens_in":431,"tokens_out":459,"duration_ms":6488,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:05:19.931421+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the halfspace NMI with a different estimator (e.g., a histogram or kernel method) on a known ground-truth load distribution and dense sensor sampling, and check whether I(X;Y) stays flat and NMI stays at 1 when the number of sensors exceeds the number of load coefficients; if h(X) drifts with sensor count or NMI drops, the saturation claim fails.","supporting_citations":[],"review_version":1}