{"id":"97c1a760-3fc5-4f95-a992-7819c361d7f7","arxiv_id":"1908.05368","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A dithered one-bit compressed sensing estimator over ReLU generative priors achieves O~(kn log d / epsilon^2) uniform recovery and a benign optimization landscape under a weight distribution condition.","lead":"This paper proves that a simple empirical risk minimization can recover signals from noisy one-bit measurements when the signal lies on a ReLU generative model, with a near-optimal statistical rate. The result matters because it shows generative priors can be used in quantized compressed sensing with few measurements, and it gives a landscape argument supporting local optimization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Computational half rests on WDC with ε_wdc ≤ n^{−32} (Thm 3.8) / n^{−76} (Rem 3.11); the only known WDC-valid weights force widths ≥ n^{Ω(1)} and m = O(k n^{65} log d), far beyond the near-optimal rate, and the paper itself leaves WDC necessity open (after Def. 3.7).","rationale":"I read the paper as making three claims: (i) the ERM estimator (4) achieves uniform recovery at rate O~(kn log d/ε^2) under sub-exponential measurements (Theorem 3.2); (ii) for shallow networks this is near-optimal (Theorem 3.4 and Corollary 3.5); (iii) under 'proper conditions on the network weights,' the empirical risk has no spurious stationary points and its global minimizer lies near x0 (Theorems 3.8 and 3.10). I could not machine-check the proofs. Structurally, the statistical half is sound: Lemma 4.1 correctly uses E[sign(V+τ)|V] = V/λ on {|V| ≤ λ} for the dithering bias; Lemma 4.2's uniform bound follows from a δ-net over G(R^k) ∩ B(R) with log net size Θ(kn log d + k log(R/δ)), the (dk+1)^n piece count, and a VC-dimension bound O(kn log d) for the sign-change counting class; the implicit m-bounds (log(λm), k log m) are resolved in Remark 3.3. The lower bound construction (Proposition 4.4) embeds nonnegative k-group sparse vectors via triangle functions in a 3-layer offset-free ReLU network, and the packing-based minimax argument (Lemma 4.5, Lemma B.4) is standard. The load-bearing weakness is the computational half. Theorems 3.8 and 3.10 import the Hand–Voroninski (2018) landscape machinery through Lemmas C.3–C.4, whose hypothesis is WDC at every layer with ε_wdc ≤ n^{−32} (Theorem 3.8) or ε_wdc ≤ n^{−76} (Remark 3.11). Since the only known WDC construction needs widths polynomial in ε_wdc^{−1}, ε_wdc ≤ n^{−32} forces widths ≥ n^{Ω(1)} and sample counts m ≳ k n^{65} log d; the landscape theorem's own radii δ1 = c n^3 ε_wdc^{1/4}||x0|| and δ2 = c n^{14} ε_wdc^{1/4}||x0|| are not simultaneously 'small' under the stated condition (δ2 can be O(n^6)||x0||), and the −ρ_n x0 basin is never shown escapable. The paper's only gesture toward actual computation is 'The landscape will further inspire us to design efficient algorithms' (§3.3), confirming that no algorithm is analyzed, and it candidly states the gap ('The question whether WDC is necessary... remains open'). Since this is exactly the concern the reader identified, and I found no additional internal inconsistency in the statistical theorem or the lower bound, I do not change the verdict: CONDITIONAL remains appropriate, with the computational tractability claim formally true but practically out of reach of the near-optimal regime.","tokens_in":51242,"tokens_out":46859,"duration_ms":395275,"concrete_test":"Estimate ε_wdc empirically for the paper's own Figure 1 network (W1 ∈ R^{64×2}, W2 ∈ R^{1024×64}, entries i.i.d. N(0,1/d_i)): for each layer, compute the empirical supremum of ||W_{i,+,x}^T W_{i,+,z} − Q_{x,z}||_2 over many random (x,z) pairs from the input space. Theorem 3.8 requires ε_wdc ≤ 2^{−32} ≈ 2.3×10^{−10} for n = 2; if the measured supremum is ≥ 0.1, the computational theorem's hypothesis fails in the very regime the authors demonstrate. Then plug the Hand–Voroninski width scaling (d_i ~ d_{i−1} log d_{i−1} · ε_wdc^{−p}) into ε_wdc = (c1 n^8)^{−4} to compute the minimal width and the implied m = O(kn log d/ε_wdc^2); if the required width exceeds the network output dimension d and m exceeds the statistical theorem's bound at the same recovery error, the computational guarantee is confirmed to live outside the near-optimal regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The statistical half (Theorem 3.2) is the sounder part: it needs no WDC, and the dithering-bias lemma (Lemma 4.1) and the uniform empirical-process bound (Lemma 4.2 with Lemmas A.2–A.4) are internally consistent as far as I can verify. The load-bearing assumption is the Weight Distribution Condition (Definition 3.7) for every layer, on which the entire computational half rests. Theorem 3.8 requires c1 n^8 ε_wdc^{1/4} ≤ 1, i.e. ε_wdc ≤ c n^{−32}, and Remark 3.11 requires c n^{19} ε_wdc^{1/4} ≤ 1, i.e. ε_wdc ≤ c n^{−76}. The only cited route to WDC (Hand and Voroninski, 2018) needs each layer width polynomial in ε_wdc^{−1}; ε_wdc ≤ n^{−32} then forces widths ≥ n^{Ω(1)} (exponent ≥ 32 at best), and since each width is at most d, also d ≥ n^{32} k log k. Then Theorem 3.8's own sample bound becomes m ≳ k n^{65} log d, whereas the statistical theorem reaches accuracy ε with m = O(kn log d/ε^2): the two halves operate at incompatible scales, so the 'joint statistical and computational analysis' does not deliver computational tractability near the optimal rate. The landscape conclusion is only that the global minimizer lies in B(x0, δ1) ∪ B(−ρ_n x0, δ2), where δ1 ≤ c n^{−5}||x0|| but δ2 = c5 n^{14} ε_wdc^{1/4}||x0|| can be as large as c n^6||x0|| under the stated condition, so the 'small neighborhood around −ρ_n x0' is not guaranteed small. No concrete algorithm with a convergence proof is given, and the spurious basin at −ρ_n x0 (higher value by Theorem 3.10) is never shown to be escapable by local search. The paper itself concedes the key gap: 'The question whether WDC is necessary for analyzing the computational aspect of the generative network remains open and warrants further studies' (§3.3). The central claim as a joint statistical–computational statement is therefore conditional in a stronger sense than the abstract suggests, although the statistical theorem and the lower bound appear structurally sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies robust one-bit compressed sensing with dithered sub-exponential measurements and an n-layer ReLU generative prior. The authors propose an unconstrained empirical risk minimization (ERM) for recovering the target G(x0), prove a uniform statistical sample-complexity bound of O~(kn log d / ε^2) under a weak sub-exponential measurement assumption, and establish an information-theoretic lower bound showing near-optimality for shallow networks. On the computational side, they analyze the landscape of the ERM under a Weight Distribution Condition (WDC), claiming that with high probability no spurious stationary points exist outside small neighborhoods of x0 and -ρ_n x0, and that the global minimizer lies near x0 under further assumptions. The statistical proof uses a dithering-bias lemma and a VC-dimension-based uniform empirical-process bound; the computational proof uses WDC and concentration lemmas. The paper is carefully written with detailed proofs in the supplement.","tokens_in":51707,"tokens_out":14143,"duration_ms":130218,"significance":"If the statistical theorem is correct, it is a genuine contribution: it provides uniform recovery guarantees for one-bit compressed sensing under sub-exponential measurements with a generative prior, without REC-type assumptions, and matches the information-theoretic lower bound up to logarithmic factors for shallow networks. The lower-bound construction via group-sparse signals is elegant and correctly transfers from linear to quantized observations. The landscape analysis is a useful step toward understanding non-convex ERM with generative priors, and the WDC-based framework connects to prior work. However, the computational half is substantially weaker than the statistical half: its assumptions force enormous layer widths and sample sizes, and it does not provide an actual algorithm. The paper's main value is the near-optimal statistical rate and the clean proof of the uniform bound; the computational claims require significant qualification.","major_comments":[{"comment":"The computational guarantee is not at near-optimal sample complexity. Theorem 3.8 requires c1 n^8 ε_wdc^{1/4} ≤ 1 and m ≥ c2 λ^2 log^2(λm)(kn log(ed) + k log(2R) + k log m + u)/ε_wdc^2. At the boundary of the WDC condition, ε_wdc ~ n^{-32}, so m = Ω(k n^{64} log d / polylog); under the stronger condition of Remark 3.11 (c n^{19} ε_wdc^{1/4} ≤ 1), m = Ω(k n^{152} log d / polylog). These rates are astronomically larger than the O~(kn log d / ε^2) statistical guarantee in Theorem 3.2. Thus the phrases 'joint statistical and computational analysis' and 'near-optimal rate' in the abstract and introduction apply only to the statistical half; the landscape theorems hold only in a regime where the sample size has a polynomial-in-n blow-up. This mismatch should be stated explicitly and the computational claims moderated.","section":"Section 3.3, Theorem 3.8 (cf. Theorem 3.2)"},{"comment":"The radius δ2 = c5 n^{14} ε_wdc^{1/4} ||x0|| around -ρ_n x0 is not guaranteed to be small under the stated condition c1 n^8 ε_wdc^{1/4} ≤ 1. At the edge of that condition, ε_wdc^{1/4} ~ n^{-8}, so δ2 is of order n^6 ||x0||. Consequently, the conclusion 'no stationary point outside of small neighborhoods around x0 and -ρ_n x0' is only accurate for the neighborhood of x0; the neighborhood of -ρ_n x0 can be enormous. To make B(-ρ_n x0, δ2) comparable to the O(n^{-5}||x0||) ball used in Theorem 3.10, one needs the separate condition c n^{19} ε_wdc^{1/4} ≤ 1, which is not part of Theorem 3.8 and again forces the n^{152} sample scaling. The statement of Theorem 3.8 should be reframed so that the two neighborhoods are treated asymmetrically, or the condition should be strengthened.","section":"Section 3.3, Theorem 3.8"},{"comment":"The computational half rests entirely on the Weight Distribution Condition for every layer. The only cited construction (Hand and Voroninski 2018) requires layer widths polynomial in ε_wdc^{-1}; with ε_wdc ≤ n^{-32}, this forces widths and hence the ambient dimension d to be at least n^{Ω(1)} k log k. For a fixed or learned ReLU generator there is no such guarantee, and the paper explicitly leaves WDC necessity open. Consequently the landscape results do not apply to the same class of networks covered by the statistical theorem, which requires only Assumption 3.1. This limitation should be elevated from a remark to a clearly stated restriction on the computational contribution, not presented as a companion guarantee for the same recovery problem.","section":"Definition 3.7 and Section 3.3"},{"comment":"No algorithm with a convergence guarantee is provided. The abstract introduces 'an algorithm' and Section 4.3 proves only existence of descent directions and compares function values near x0 and -ρ_n x0; it does not show that any polynomial-time method converges to a global minimizer, nor does it give iteration complexity. Given the paper's claim of a 'joint statistical and computational analysis,' the computational part should either include an explicit algorithm (e.g., a descent or projected scheme with a convergence proof) or be described as landscape analysis only, without implying computational tractability.","section":"Abstract and Sections 1.3, 4.3"}],"minor_comments":[{"comment":"There are several typographical errors: 'Olicz' should be 'Orlicz' in the notation section; Theorem 3.8 contains 'sastisfies' and 'wights'; Figure 1 caption has 'Guassian' instead of 'Gaussian'; Corollary 3.5 has 'quantizd' instead of 'quantized'.","section":"Notation and typos"},{"comment":"The estimator x̂_m is defined as an exact global minimizer of L(x), but the ERM objective is not coercive in general (e.g., directions in which the ReLU network outputs zero produce L(x)=0 along a ray). The paper should state that x̂_m is taken to be an ε-approximate minimizer, or add conditions under which a global minimizer exists.","section":"Equation (4)"},{"comment":"In the proof of Lemma A.4, the chain 'm ≥ √c2 k m ≥ √c2 k log m' appears dimensionally incorrect; the intended inequality is likely '√(c2 k m) ≥ √c2 k log m' or similar. Please correct this step so the lower bound on ηm/(3λ) is transparent.","section":"Supplement, Lemma A.4 proof"},{"comment":"The theorem statements use the condition c1 n^8 ε_wdc^{1/4} ≤ 1, while the proof of Theorem 3.8 invokes 88π n^6 ε_wdc^{1/4} < 1 and Lemma C.3 is stated with √ε. The exponents and constants should be reconciled, and Lemma C.3 should be restated in the notation used by the theorems.","section":"Section 3.3 and Lemma C.3"}],"recommendation":"major_revision","confidential_remarks":"The statistical contribution is solid and likely the main value of the paper: the uniform recovery rate is derived from first principles, the lower bound is clean, and the VC-dimension argument is nontrivial. The computational half, however, is substantially overclaimed: the WDC condition forces sample sizes and widths that are incompatible with the near-optimal rate, and no algorithm is actually provided. I would encourage the editor to request a revision that clearly separates the statistical near-optimality from the much weaker computational statement, strengthens the discussion of the WDC regime, and fixes the noted inconsistencies. The paper is within scope for a statistics/information-theory journal, but it is not yet ready as a 'joint statistical and computational' contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is the dithered one-bit compressed sensing paper with ReLU generative priors. The statistical part is the real contribution. Theorem 3.2 gives uniform recovery at m = O~(kn log d / eps^2) under sub-exponential measurements, with no REC/RIP condition and no dependence on the weight distribution condition. The proof structure is believable: dithering smooths the sign function, the bias term is controlled by a sub-exponential tail bound, and the empirical process term is handled by counting linear pieces and a VC dimension bound rather than generic chaining. I did not machine-check every line, but the steps I read are internally consistent. The lower bound for depth-3 networks via k-group sparse vectors is a clean transfer from sparse recovery, and the citation pattern is not a problem; the one self-citation is for context, not as a black box.\n\nThe soft spot is the computational half, and it is not minor. The global landscape theorems hold only under WDC with eps_wdc ≤ n^{-32} (Theorem 3.8) or n^{-76} (Remark 3.11). The only cited route to WDC, Hand–Voroninski, requires layer widths polynomial in eps_wdc^{-1}, which forces width, and hence d, to be at least n^{Omega(1)} k log k. Then the sample bound in the same theorem becomes m ≳ k n^{65} log d. That is nowhere near the near-optimal statistical rate; the two halves operate at incompatible scales. So the paper's claimed joint statistical-computational analysis overstates what is delivered. Further, Theorem 3.8 only rules out stationary points outside B(x0, delta1) union B(-rho_n x0, delta2), and under the stated condition delta2 can be as large as c n^6 ||x0||. The 'small neighborhood' around the negative multiple is not guaranteed small. Theorem 3.10 shows L is lower near x0 than near -rho_n x0, but that does not rule out other spurious points, and no concrete algorithm with a convergence proof is given. The authors do concede that WDC necessity is open, which is honest but confirms the gap.\n\nBottom line: this deserves a serious referee, primarily for the statistical contribution. A good referee should push the authors to separate the near-optimal statistical result from the landscape claims, state clearly that the computational guarantees are at a much weaker scale, and either provide an algorithm with convergence guarantees or explicitly say none is provided. I would not cite the computational half as is, but I would cite the statistical theorem.","headline":"A sound near-optimal statistical result for dithered one-bit recovery with ReLU priors, wrapped in a computational claim whose WDC-dependent constants put it at a far weaker scale.","tokens_in":52310,"tokens_out":2577,"would_cite":true,"duration_ms":28695,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62C20","62J07","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Dithered one-bit measurements with a ReLU generative prior recover every bounded target uniformly at rate $\\widetilde{O}(kn\\log d/\\varepsilon^2)$; under a weight-distribution condition the ERM landscape has no spurious stationary points…","keywords":["one-bit compressed sensing","generative priors","ReLU networks","empirical risk minimization","dithered quantization","sub-exponential measurements","non-convex landscape","sample complexity"],"falsifier":"Choose a two-layer ReLU network whose weight matrix is designed to violate the Weight Distribution Condition (e.g., a rank-deficient or near-zero row), generate dithered one-bit measurements of a fixed $G(x_0)$, and grid-search the representation space for a point far from both $x_0$ and $-\\rho_n x_0$ with no negative directional derivative of the ERM; such a point would refute the landscape theorem, while checking whether ERM solutions still meet the $\\varepsilon$ error bound would test the statistical theorem independently.","tokens_in":51026,"feed_emoji":"📡","tokens_out":10191,"duration_ms":96008,"temperature":0.7,"pith_summary":"Robust one-bit compressed sensing is usually studied under sparsity in a known basis; this paper replaces that assumption with a fixed $n$-layer ReLU generative prior $G:\\mathbb{R}^k\\to\\mathbb{R}^d$, so the target is $\\theta_0=G(x_0)$. It proposes to recover $G(x_0)$ by an unconstrained ERM over the low-dimensional representation $x$, and proves that $m=\\widetilde{O}(kn\\log d/\\varepsilon^2)$ dithered sub-exponential measurements suffice for every bounded target simultaneously, up to error $\\varepsilon$. For shallow generators, an information-theoretic lower bound shows this rate is optimal up to logarithmic factors. The same loss, under a condition on the layer weights, has no bad stationary points away from the true representation and its negative multiple, making the non-convex recovery problem computationally tractable.","feed_headline":"ReLU generative priors hit near-optimal one-bit recovery rates","feed_subtitle":"One ERM solution recovers any bounded target from O~(kn log d / ε²) dithered measurements, with no known sparsity basis.","key_machinery":"The load-bearing objects are dithering and piecewise linearity. Uniform dithering makes $\\mathbb{E}_\\tau[\\mathrm{sign}(V+\\tau)]=V/\\lambda$ when $|V|\\le\\lambda$, which linearizes the quantized observation and yields the quadratic bias term in the ERM. The ReLU network's piecewise linearity cuts $\\mathbb{R}^k$ into at most $(dk+1)^n\\le(2d)^{kn}$ linear pieces; counting these pieces converts a two-parameter empirical-process supremum into a covering argument over $2k$-dimensional subspaces, and a VC-dimension bound of order $kn\\log d$ controls the fraction of indices whose signs change under the covering. For the landscape analysis, the named condition is the Weight Distribution Condition (WDC): each layer matrix, after the ReLU nonlinearity, approximately preserves the angle between any two inputs. WDC plus an angle-recursion map $h_{x,x_0}$ forces the gradient to align with either $x_0$ or $-\\rho_n x_0$, where $\\rho_n\\in(0,1]$ is a depth-dependent scalar tending to $1$.","core_discovery":"On its own terms, the central discovery is that the sign-quantization difficulty disappears when dithering noise $\\tau\\sim\\mathrm{Unif}[-\\lambda,\\lambda]$ is introduced: conditionally, the expected signed measurement becomes proportional to $\\langle a,G(x_0)\\rangle+\\xi$, so the population version of the loss $L(x)=\\|G(x)\\|_2^2-\\frac{2\\lambda}{m}\\sum_{i=1}^m y_i\\langle a_i,G(x)\\rangle$ has a quadratic bias term in the output space. The remaining obstacle is uniform concentration of the empirical process over pairs $(x,x_0)$, and the paper shows that the piecewise-linear structure of ReLU networks bounds the relevant complexity by the number of linear pieces, $(2d)^{kn}$, leading to the stated rate. The landscape theorems assert that whenever every layer satisfies the Weight Distribution Condition, any point outside small neighborhoods of $x_0$ and $-\\rho_n x_0$ has a descent direction, and the global minimum lies in the $x_0$-neighborhood rather than the negative-multiple one.","pith_inferences":["Because the statistical argument uses only piecewise linearity, the same rate should carry over to leaky-ReLU, max-pooling, or other piecewise-linear generators; the paper says the extension is plausible but does not prove it.","The linear dependence on depth $n$ enters through the $(2d)^{kn}$ linear-piece count; for realistic trained networks with far fewer effective regions, the bound may be loose, and measuring the effective linear-piece count would be a direct test.","Whether the Weight Distribution Condition can be enforced or certified for learned weights is open; if training can maintain approximate angle preservation, the landscape theorem becomes usable beyond random expansive Gaussian networks.","The theorem requires $\\lambda$ to grow logarithmically in $1/\\varepsilon$, so extreme accuracy demands a wide dithering range; hardware with limited dynamic range will face a trade-off between precision and the validity of the bound."],"forward_implications":["A fixed ReLU generative model can serve as a drop-in prior for one-bit sensing: sample complexity scales with latent dimension $k$ and depth $n$, with only a logarithmic dependence on ambient dimension $d$.","Recovery is uniform: the same measurement count guarantees simultaneous recovery of every target $G(x_0)$ with $\\|G(x_0)\\|_2\\le R$, so the target may be chosen adversarially.","Under the Weight Distribution Condition, gradient-type local search on the ERM will not be trapped at spurious stationary points; the only basins are near the true representation and its negative multiple.","For shallow generators the sample rate is optimal up to $\\log(1/\\varepsilon)$ and $k$ factors, matching classical sparse one-bit sensing without requiring a known sparsity basis."],"supporting_citations":[{"why":"established dithered one-bit compressed sensing with sub-Gaussian measurements and a convex recovery algorithm; this paper extends the setting to sub-exponential measurements and a generative prior.","marker":"Dirksen and Mendelson (2018a)"},{"why":"introduced generative-model compressed sensing via empirical risk minimization over the range of $G$, the linear-model framework this ERM adapts to one-bit noisy observations.","marker":"Bora et al. (2017)"},{"why":"supplies the Weight Distribution Condition and the angle-preservation lemmas used to analyze the non-convex landscape.","marker":"Hand and Voroninski (2018)"},{"why":"provides the two lemmas comparing function values around $x_0$ and $-\\rho_n x_0$ used in the proof of Theorem 3.10.","marker":"Huang et al. (2018)"},{"why":"gives the group-sparse construction and lower-bound perspective that the shallow-network information-theoretic lower bound builds on.","marker":"Liu and Scarlett (2019)"},{"why":"supplies the minimax packing-number lower bound used to prove the sample-complexity lower bound.","marker":"Plan et al. (2016)"},{"why":"bounds the number of regions produced by hyperplane arrangements, used to count linear pieces of the ReLU network.","marker":"Winder (1966)"},{"why":"is the sparse one-bit recovery baseline with uniform guarantees whose rate the generative result is compared against.","marker":"Plan and Vershynin (2013)"}],"fun_headline_variants":["ReLU nets achieve near-optimal robust one-bit recovery","One ERM, ReLU priors: robust one-bit recovery at optimal rate","Dithering makes one-bit recovery easy for ReLU generative nets","Global landscape: ReLU nets avoid bad stationary points in one-bit recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every layer's weight matrix satisfies the Weight Distribution Condition with a very small error (roughly $\\varepsilon_{wdc}\\lesssim n^{-32}$ in the theorem's stated form); for arbitrary or learned ReLU weights there is no guarantee, so a network violating the condition could still have spurious stationary points even though the uniform statistical recovery bound remains valid.","fun_headline_variants_meta":{"raw":{"variants":["ReLU nets achieve near-optimal robust one-bit recovery","One ERM, ReLU priors: robust one-bit recovery at optimal rate","Dithering makes one-bit recovery easy for ReLU generative nets","Global landscape: ReLU nets avoid bad stationary points in one-bit recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1563,"prompt_tokens":1118,"completion_tokens":445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":734,"tokens_out":445,"duration_ms":4676,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:16:41.407110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a two-layer ReLU network whose weight matrix is designed to violate the Weight Distribution Condition (e.g., a rank-deficient or near-zero row), generate dithered one-bit measurements of a fixed $G(x_0)$, and grid-search the representation space for a point far from both $x_0$ and $-\\rho_n x_0$ with no negative directional derivative of the ERM; such a point would refute the landscape theorem, while checking whether ERM solutions still meet the $\\varepsilon$ error bound would test the statistical theorem independently.","supporting_citations":[{"cited_title":"Information-Theoretic Lower Bounds for Compressive Sensing with Generative Models","cited_arxiv_id":"1908.10744","evidence_quote":"gives the group-sparse construction and lower-bound perspective that the shallow-network information-theoretic lower bound builds on."},{"cited_title":", Vershynin, R","cited_arxiv_id":null,"evidence_quote":"supplies the minimax packing-number lower bound used to prove the sample-complexity lower bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"bounds the number of regions produced by hyperplane arrangements, used to count linear pieces of the ReLU network."},{"cited_title":"and Vershynin, R","cited_arxiv_id":null,"evidence_quote":"is the sparse one-bit recovery baseline with uniform guarantees whose rate the generative result is compared against."}],"review_version":1}