{"id":"bdf21417-2ded-4fad-ab19-985301562d45","arxiv_id":"2508.20650","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Self-composing neural operators that reuse one backbone block and are trained with a depth-increasing curriculum sharply reduce prediction error on Darcy flow and ultrasound tomography benchmarks while using far fewer parameters.","lead":"The paper presents neural operators that build depth by repeatedly applying one shared block, mirroring iterative numerical solvers, and trains them with a curriculum that increases depth. If sound, it offers a parameter-efficient way to model high-frequency PDEs such as ultrasound computed tomography, with substantially lower errors than Fourier neural operator baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.3 is stated with 'for any n there exist L, P, G', so it bounds the best error in the class of n-fold compositions, not a fixed backbone's accuracy as n grows. The compressed Lemma 2.2 and the paper's own caveats keep the central depth-accuracy claim conditional.","rationale":"The reader's weakest assumption and my concern overlap: Theorem 2.3 is the central mechanism for the depth-accuracy claim. My reading sharpens it: even granting the external theorem, Theorem 2.3 as written is an existence result for each n rather than a statement about a fixed backbone, and the proof's rescaling step is compressed. This matters because the paper interprets the theorem as 'larger n, lower error' and Train-and-Unroll relies on that intuition, while the empirical sections provide only indirect support. I found no reason to distrust the results or to reject the paper outright; the correct response remains conditional acceptance with a request to fix or strengthen the theoretical claim and to release the missing artifacts. Hence the verdict is unchanged.","tokens_in":20306,"tokens_out":18434,"duration_ms":186001,"concrete_test":"Check the exact statement and quantifier order of Theorem 1.3 in Zhang, Lu, and Zhao (2023). If it asserts a single fixed-size network g with error <= C_F n^{-1/k} for all n, then strengthen Theorem 2.3 to 'there exists G such that for all n ...' and re-verify Lemma 2.2 by applying the theorem once per F_i at the common n; if the common-n/rescaling step still requires g_i = (hat-g_i)^q, prove that the q additional compositions preserve the error bound. If Theorem 1.3 instead gives a possibly n-dependent network, then Lemma 2.2 as printed is invalid and Theorem 2.3 needs a new argument. In either case, an n-ablation on the USCT task with n=1,2,4,8 and fixed hyperparameters would show whether the actual algorithm displays the claimed depth-accuracy scaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that self-composing a fixed-size backbone G improves accuracy as the depth n grows. Theorem 2.3, however, is stated as: for any epsilon there exist fixed N, tilde-N, L such that for any n one can find L_n, P_n, G_n with error at most epsilon + C_epsilon/log(n). Because G is allowed to depend on n, this theorem does not show that a single backbone becomes more accurate as n increases; it only bounds the approximation error of the whole class of fixed-size n-fold compositions. The sentence after the proof, 'the deeper, the larger n, the better,' requires either one G working for all n or nested representable classes, neither of which is stated. The proof's bridge to the external depth-approximation result is Lemma 2.2: it first obtains approximations hat-F_i with composition counts n_i, then introduces a common n = qn >= max_i n_i, and finally sets g_i = (hat-g_i)^q. As printed this rescaling is not verified, and if Theorem 1.3 of Zhang-Lu-Zhao only supplies an n_i-dependent network, the argument fails. The paper's own 'Discussion on the convergence rate' concedes that the O(1/log n) bound is slow and that practical gains may come from parameter-sharing regularization and Train-and-Unroll initialization rather than asymptotic n-dependence. Furthermore, Table 1 shows MgNO-self does not beat MgNO on Darcy, and Table 2 lacks a non-self MgNO USCT baseline. The intended claim may be repairable, but the central theoretical mechanism is not yet demonstrated as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-composing neural operator (SC-NO) framework of the form O(v)=P∘(G∘)^n∘L(v), in which a single backbone operator G is applied n times to a lifted latent state, motivated by iterative PDE solvers such as multigrid. The authors prove a universal approximation result for any fixed n (Theorem 2.1) and a quantitative depth-accuracy bound of ϵ + C_ϵ/log(n) for Lipschitz target operators (Theorem 2.3), introduce a Train-and-Unroll strategy that grows composition depth during training, and report experiments on Darcy flow benchmarks and a 300--500 kHz USCT Helmholtz task where the MgNO-self model achieves substantially lower RRMSE than FNO, AFNO, BFNO, and UNet.","tokens_in":20664,"tokens_out":11842,"duration_ms":109747,"significance":"If the theoretical claims were fully established, the paper would make a useful contribution to parameter-efficient neural operator design: the self-composition architecture shares parameters across depth, the Train-and-Unroll curriculum is a practical training scheme, and the USCT results at high frequencies address a genuinely hard benchmark. The paper is also honest in its 'Discussion on the convergence rate,' acknowledging that the proved O(1/log n) rate is slow and that practical gains may come from parameter sharing and initialization rather than asymptotic depth scaling. However, as written, the central depth-accuracy theorem is not demonstrated because of its quantifier structure and its dependence on an unstated external approximation theorem, and the empirical isolation of the self-composition mechanism is incomplete. No code, data split details, or error bars are provided, so the empirical claims are currently not reproducible.","major_comments":[{"comment":"The statement of Theorem 2.3 has the quantifier structure 'for any n there exist L_n, P_n, G_n', so it bounds the optimal error of the n-fold composition class, not the accuracy of a single fixed backbone as n grows. The conclusion after the proof that 'the deeper, the larger n, the better' is not a logical consequence of the theorem. Since the paper's central motivation is that repeatedly applying the same backbone G improves accuracy, either prove a fixed-G version (G independent of n) or explicitly restate the claim as a property of the class and soften the interpretation accordingly. The paper's own 'Discussion on the convergence rate' concedes that the O(1/log n) rate is slow and that practical gains may stem from parameter-sharing regularization and Train-and-Unroll initialization; this concession should be reconciled with the surrounding narrative.","section":"Section 2.3, Theorem 2.3 (Eq. 2.28)"},{"comment":"The proof of Lemma 2.2 is not self-contained or verifiable as written. It invokes Theorem 1.3 of [52] without stating its hypotheses or conclusion, and the rescaling step 'take n = q n ≥ max_i n_i' followed by 'g_i = (g_i∘)^q' is not a valid derivation unless Theorem 1.3 supplies a single fixed-size network whose n-fold composition approximates each F_i for every n. If the external theorem only supplies an n_i-dependent network, then increasing the composition count by q does not preserve the approximation error, and the argument collapses. This lemma is load-bearing for both Theorem 2.1 and Theorem 2.3, so the authors must either state and prove the adaptation of the external theorem in full or provide a self-contained proof. As printed, the notation also mixes n_i and n inconsistently, which further obscures the argument.","section":"Section 2.3, Lemma 2.2 (Eqs. 2.9-2.10)"},{"comment":"The empirical evidence for the benefit of self-composition is incomplete. Table 2 compares MgNO-self with UNet, FNO, AFNO, and BFNO but includes no non-self MgNO baseline, so the large gains on the USCT Helmholtz task cannot be attributed to self-composition rather than to the multigrid-inspired AdaConv backbone. On the Darcy benchmarks in Table 1, MgNO-self is consistently less accurate than MgNO (e.g., L2 error 0.187 vs 0.153 on smooth, 0.371 vs 0.339 on rough, 0.800 vs 0.715 on multiscale), which further weakens the claim that repeated composition improves accuracy. The authors should add a non-self MgNO baseline on USCT and report controlled ablation experiments that isolate composition depth from backbone design.","section":"Section 5, Table 2 and Section 2.4, Table 1"},{"comment":"The reported SOTA results are single point estimates with no error bars, no number of seeds, no data split details, and no training hyperparameters; no code is provided. For a methods paper whose main empirical claim is a large margin over baselines, this is not sufficient to assess robustness, especially given the small baseline set. Please report means and standard deviations over multiple seeds, describe the data split and preprocessing for OpenBreastUS, and release code or provide detailed training configurations.","section":"Section 5, numerical experiments"}],"minor_comments":[{"comment":"The title in the paper header differs from the arXiv title; please align the submitted title with the published version.","section":"Title"},{"comment":"The arithmetic in Eq. (2.44) is inconsistent with the preceding bounds: the first term should be ϵ/2 if it is to match Eq. (2.29), not ϵ/3. Also, Theorem 2.3 should state n ≥ 2, since log(1) = 0.","section":"Section 2.3, Eq. (2.44)"},{"comment":"The proof of Lemma 2.2 uses F_i and eF_i interchangeably in Eq. (2.10); please clarify the notation throughout the proof.","section":"Lemma 2.2"},{"comment":"There are numerous typos, e.g., 'self-comprising' in the Theorem 2.1 preamble, 'naural motivation', 'campactness', and 'comprehensivly benched'; a careful proofread is needed.","section":"Throughout"},{"comment":"Figure 2 is referenced but the visual is not present in the submitted text; ensure the figure and its caption are included and that the axes and labels are described.","section":"Figure 2"},{"comment":"The note 'MgNO-self mode is named after MgNO' is unclear and should be expanded or removed.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong practical motivation and the USCT results are impressive, but the central theoretical mechanism is not yet established as written. I recommend asking for a precise statement of the external theorem of [52], a repair of Lemma 2.2, and a restatement of Theorem 2.3 that either fixes G or explicitly characterizes the optimal n-fold class. The empirical section would be much stronger with a non-self MgNO baseline on USCT and with standard deviations over seeds. I do not see a fundamental circularity or a conflict of interest; the main issue is the mismatch between the claimed depth-accuracy behavior and the theorem's quantifier structure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper deserves a real referee, but the advertised scaling law isn't established by the theorem as stated. The empirical results are interesting enough that the paper should be seen through to review, not desk-rejected.\n\nWhat's new: the self-composing formulation O(v)=P∘(G∘)^n∘L for neural operators, the Train-and-Unroll curriculum, and the multigrid-inspired backbone with adaptive convolution for the Helmholtz problem. Universal approximation for any fixed n is a solid formal step. The USCT results are striking—RRMSE 0.0028 at 300 kHz vs 0.0113 for the best baseline—and the parameter efficiency on Darcy (0.17M vs 0.57M for MgNO) is real.\n\nThe soft spot: Theorem 2.3 states that for each n there exist L, P, G of fixed size with error ≤ ε + C_ε/log n. That doesn't mean the same G gets better as n grows; G is allowed to vary with n. The sentence \"the deeper, the larger n, the better\" overreads the theorem. Lemma 2.2's rescaling is compressed, and the argument leans on an external approximation result. To their credit, the authors' own discussion admits the O(1/log n) rate is slow and that practical gains likely come from parameter sharing and T&U, not the asymptotic bound. So the theory as written supports the architecture, not the specific depth-scaling claim.\n\nOther issues: no code or data splits, no error bars, and Table 1 shows MgNO-self slightly underperforms MgNO on Darcy, so the abstract's SOTA claim is too broad. Table 2 lacks a non-self MgNO baseline on USCT, which would isolate the self-composition effect.\n\nWho's this for: anyone working on neural operators, spectral bias, or multigrid-inspired architectures. The gap between theorem and claim is worth discussing in a reading group.\n\nMy recommendation: send it to peer review. The framework and the training strategy are novel, the empirical signal is strong, and the theory can likely be repaired or reframed; a good referee will push for code release and a corrected theorem statement.","headline":"Fresh framework and strong USCT numbers, but the depth-scaling theorem is overstated as written; still worth a serious referee.","tokens_in":21162,"tokens_out":2329,"would_cite":true,"duration_ms":21337,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Txx","65Mxx"],"pacs":[],"model":"deepseek-v4-flash","headline":"Repeatedly applying one fixed neural-operator backbone makes it a universal approximator, with error that shrinks as the number of repetitions grows.","keywords":["self-composing neural operator","operator learning","multigrid","Helmholtz equation","ultrasound computed tomography","train-and-unroll","spectral bias","high-frequency PDE surrogates"],"falsifier":"Take a fixed Lipschitz target operator, such as the Helmholtz map on a simple two-layer medium, train the self-composed model at depths $n=1,2,4,\\dots$ with identical backbone width and depth and identical data, and compare test error. If the error does not decay like $1/\\log(n)$, or at least monotonically, as $n$ grows, or if the backbone size must grow with $n$ to see improvement, the central depth-accuracy claim is refuted. A cleaner check is to test the quoted fixed-size-composition theorem directly: a constant-size ReLU network's $n$-fold composition must reach $O(n^{-1/k})$ error on a known Lipschitz function such as a quadratic; if it plateaus, the proof mechanism fails.","tokens_in":20130,"feed_emoji":"🌊","tokens_out":15985,"duration_ms":134005,"temperature":0.7,"pith_summary":"This paper claims that a neural operator of the form $O(v)=P\\circ (G\\circ)^n\\circ L(v)$ — one fixed backbone block $G$ applied $n$ times between a lifting map $L$ and a projection $P$ — is a universal approximator for every fixed $n$, and that for Lipschitz target operators the approximation error falls at rate $\\epsilon+C_\\epsilon/\\log(n)$ as the composition count $n$ grows, with the backbone unchanged. If true, this turns depth into a free accuracy knob: PDE surrogates can be made more accurate by reusing the same parameter-efficient block instead of adding new parameters. That matters for high-frequency problems like ultrasound computed tomography, where standard Fourier-type operators struggle to represent sharp oscillations and large models strain memory. The paper proves both theorems, introduces a Train-and-Unroll curriculum that increases $n$ during training, and pairs the framework with a multigrid-inspired backbone with adaptive convolutions. On the USCT Helmholtz benchmark the resulting model reaches relative root-mean-square error 0.0028 at 300 kHz, roughly ten times lower than a Fourier neural operator and four times lower than the best compared Fourier variant.","feed_headline":"Repeating one block slashes 300-kHz wave error","feed_subtitle":"One shared backbone, repeated n times, hits 0.0028 relative error—far below Fourier baselines.","key_machinery":"The central object is the self-composition operator $O(v)=P\\circ (G\\circ)^n\\circ L(v)$, where $L$ lifts the coefficient field into a latent function space, $G$ is one neural operator of fixed network size applied $n$ times, and $P$ projects back to the solution space; this mirrors fixed-point iteration in PDE solvers by treating each application as one solver sweep. The depth-accuracy mechanism is Theorem 2.3, which relies on an external lemma stating that a fixed-size ReLU network's $n$-fold self-composition approximates Lipschitz functions to $O(n^{-1/k})$, giving the $1/\\log(n)$ decay after the base error. For the Helmholtz application, the machinery is the multigrid V-cycle backbone with the adaptive-convolution mechanism $\\operatorname{AdaConv}(k,x)=\\bigl(\\operatorname{MLP}(\\operatorname{Filter}_k*k)\\bigr)\\odot\\bigl(\\operatorname{Filter}_x*x\\bigr)$, which lets the learned smoother and PDE operator modulate their action by local wavenumber values.","core_discovery":"The core discovery is that self-composition alone is expressive: any continuous operator between Hilbert function spaces can be approximated to arbitrary accuracy by $P\\circ(G\\circ)^n\\circ L$ even when the inner network $G$ is reused verbatim $n$ times (Theorem 2.1), and for Lipschitz operators (ones whose response changes by at most a constant factor times the change in input) one fixed $G$ of fixed size achieves error $\\epsilon + C_\\epsilon/\\log(n)$ by increasing $n$ (Theorem 2.3). The proof reduces the operator to finitely many scalar functionals, approximates each scalar function by an $n$-fold self-composition of a fixed ReLU network using a quoted result that avoids growing the network, and then lifts that scalar approximation back to the function space using constant functions. The paper also identifies the practical mechanism: the Train-and-Unroll strategy starts with $n=1$, trains to convergence, then increments $n$ and warm-starts, so the deep self-composed model is reached by a curriculum rather than a single difficult optimization. With a multigrid V-cycle backbone and adaptive convolutions that multiply wave features by wavenumber-dependent masks, the model becomes MgNO-self and sets the reported low errors on the USCT task.","pith_inferences":["A testable extension the paper leaves open is swapping the multigrid backbone for a different local update inside the same $P\\circ(G\\circ)^n\\circ L$ shell; if the depth-scaling benefit persists, self-composition is a general accuracy lever rather than a multigrid artifact.","The paper's own error-versus-depth curves decay faster than the proven $1/\\log(n)$ rate; an editorial inference is that the bound is loose, and the practical depth scaling may be closer to $1/n$ for smoother operators.","Because the same weights are reused at every stage, self-composition imposes a strong weight-tying prior: the same solver sweep is assumed useful everywhere. An ablation that unshares the per-stage weights would isolate how much of the gain is due to sharing itself.","In the large-$n$ limit the composed block approaches a fixed point of $G$, suggesting self-composition could serve as a warm start or regularizer for implicit-style surrogates; the paper does not explore that connection."],"forward_implications":["Increasing the composition count $n$ lowers the guaranteed approximation error to $\\epsilon + C_\\epsilon/\\log(n)$ with no new parameters, so depth becomes a controlled accuracy knob for a fixed backbone.","Because a practitioner can stop training at any intermediate depth, Train-and-Unroll outputs a whole family of surrogates and lets compute be traded against accuracy at deployment time.","For the USCT Helmholtz problem, the self-composed multigrid model reduces relative error to 0.0028 at 300 kHz, 0.0036 at 400 kHz, and 0.0049 at 500 kHz, outperforming FNO, AFNO, and BFNO on both RRMSE and maximum error.","On Darcy and multiscale benchmarks, self-composition cuts parameter counts by about a factor of three (0.17M versus 0.57M for the MgNO backbone) while keeping accuracy essentially on par with the non-shared model."],"supporting_citations":[{"why":"Provides the external theorem (Theorem 1.3) that a fixed-size ReLU network composed n times approximates Lipschitz functions with error O(n^{-1/k}) while the network size stays independent of n; Lemma 2.2 and Theorem 2.3 depend on it.","marker":"[52]"},{"why":"Establishes the nonlocal neural operator universal approximation framework and the abstract deep operator notation that Theorem 2.1 extends to self-composition for n=1.","marker":"[31]"},{"why":"Introduces the MgNO multigrid parameterization of linear operators and the abstract fully connected neural operator framework used throughout; also serves as the non-shared baseline in Table 1.","marker":"[21]"},{"why":"Supplies the Darcy flow dataset and the FNO2D baseline; its spectral-bias limitations are the motivating target for high-frequency and multiscale surrogates.","marker":"[33]"},{"why":"Provides the multiscale benchmark and prior analysis of spectral bias that the paper uses as a testbed in Table 1.","marker":"[34]"},{"why":"Supplies the OpenBreastUS ultrasound computed tomography dataset and Helmholtz forward-simulation benchmark on which MgNO-self reaches RRMSE 0.0028.","marker":"[51]"},{"why":"Establishes the original link between multigrid V-cycles and convolutional networks, which the paper adapts into the learnable V-cycle backbone for the Helmholtz application.","marker":"[22]"}],"fun_headline_variants":["One block repeated tames 500-kHz wave error","Self-composed operator beats Fourier nets on USCT","Curriculum unrolling: deep surrogate from one block","Single backbone, many passes: PDE error drops","Self-composition proves universal, cuts wave error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is a quoted external result: one fixed-size network, applied $n$ times to itself, can approximate any function whose output changes by at most a constant factor times the input change, with error shrinking like $n^{-1/k}$, without growing the network. If that result fails in the form used, the paper's guarantee that deeper self-composition reduces error collapses.","fun_headline_variants_meta":{"raw":{"variants":["One block repeated tames 500-kHz wave error","Self-composed operator beats Fourier nets on USCT","Curriculum unrolling: deep surrogate from one block","Single backbone, many passes: PDE error drops","Self-composition proves universal, cuts wave error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1506,"prompt_tokens":1031,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":415}},"tokens_in":647,"tokens_out":475,"duration_ms":4912,"temperature":1.0,"reasoning_tokens":415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:41:30.747960+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed Lipschitz target operator, such as the Helmholtz map on a simple two-layer medium, train the self-composed model at depths $n=1,2,4,\\dots$ with identical backbone width and depth and identical data, and compare test error. If the error does not decay like $1/\\log(n)$, or at least monotonically, as $n$ grows, or if the backbone size must grow with $n$ to see improvement, the central depth-accuracy claim is refuted. A cleaner check is to test the quoted fixed-size-composition theorem directly: a constant-size ReLU network's $n$-fold composition must reach $O(n^{-1/k})$ error on a known Lipschitz function such as a quadratic; if it plateaus, the proof mechanism fails.","supporting_citations":[{"cited_title":"Zhang, J","cited_arxiv_id":null,"evidence_quote":"Provides the external theorem (Theorem 1.3) that a fixed-size ReLU network composed n times approximates Lipschitz functions with error O(n^{-1/k}) while the network size stays independent of n; Lemma 2.2 and Theorem 2.3 depend on it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the MgNO multigrid parameterization of linear operators and the abstract fully connected neural operator framework used throughout; also serves as the non-shared baseline in Table 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Darcy flow dataset and the FNO2D baseline; its spectral-bias limitations are the motivating target for high-frequency and multiscale surrogates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multiscale benchmark and prior analysis of spectral bias that the paper uses as a testbed in Table 1."},{"cited_title":"He and J","cited_arxiv_id":null,"evidence_quote":"Establishes the original link between multigrid V-cycles and convolutional networks, which the paper adapts into the learnable V-cycle backbone for the Helmholtz application."}],"review_version":2}