{"id":"29d442d8-e59c-471a-b067-5bf6088a0789","arxiv_id":"2507.09213","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A constructive wavelet network picks its initial wavelet frequency by estimating the spatial energy of the function and adds high-energy bases to reach a preset accuracy with fewer parameters.","lead":"This paper introduces a constructive wavelet neural network that estimates the dominant frequency of an unknown nonlinear function from data, then adds only the most energetic wavelet building blocks until a preset accuracy is reached. The approach is tested on synthetic mappings and a real milling-machine time-series dataset, showing reduced parameter counts relative to a generic wavelet network.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The basis-expansion step assumes high-energy next-scale wavelets lie near current-scale centers; no theorem supports this and the paper's examples do not test spatially shifted frequency content.","rationale":"The paper's central claim is that the constructive WNN can reach a preset accuracy with fewer parameters by adding only high-energy wavelet bases. For that claim, two algorithm design choices are load-bearing: (i) Algorithm 1's stopping rule, which the reader flags as relying on unimodality (Assumption 2); and (ii) Algorithm 2's rule for which bases in the next subspace to add. I find (ii) more fundamental. Even if the initial frequency is perfectly estimated, the basis-increase step must identify the truly large coefficients at each new scale. The paper's only theoretical tool for that is Corollary 1, which says coefficients outside B_epsilon are negligible; it is silent on the spatial arrangement of the significant coefficients inside B_epsilon. The algorithm answers this silence with an unproven locality heuristic: for each high-energy center in W_m, take the nearest centers in W_{m+1}. This heuristic is not derived from multiresolution analysis; wavelet coefficients at finer scales are not generally concentrated at the same translation indices as at coarser scales, especially for functions with scale-dependent spatial structure. The experiments in Examples 1 and 4 use functions where the energy is spatially localized, so they cannot expose a failure. The concrete test with a two-bump function directly targets this gap: if the nearest-neighbor expansion misses the high-energy bases, the algorithm will only reach epsilon after falling back to broad expansion, undermining the parameter-reduction headline. This is not a rejection of the paper, but it means the claimed generality needs either a theorem about coefficient locations or an experiment covering spatially separated scale-dependent frequencies. The reader's unimodality concern remains valid and is partially related, but the locality assumption is the one I would test first.","tokens_in":28838,"tokens_out":6986,"duration_ms":80464,"concrete_test":"Take f(x) on [0,1] with two well-separated components: f(x)=50 exp(-((x-0.2)/0.02)^2) sin(4 pi x) + 50 exp(-((x-0.8)/0.02)^2) sin(32 pi x). Run Algorithm 1 to get minit, then Algorithm 2 with the paper's settings (e.g., epsilon=0.0025, mu=1/3, kappa=0.36) on a uniform grid of Nsa points. Record the set A of bases added at scale minit+1 and the final parameter count P_CWNN. Compute the true top-K coefficients at scale minit+1 by numerical integration C_n = integral f(x) psi_{minit+1,n}(x) dx using the Sinc wavelet of the paper. If A misses more than half of the top-K coefficients and P_CWNN exceeds the number required by an oracle that adds the true top-K bases by more than 50%, the locality assumption is empirically falsified. Also report whether Algorithm 2 reaches epsilon within a fixed budget; if it stalls until mu_up=1 or m increases, that confirms the failure mode.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim rests on Algorithm 2 (Section IV-C) adding only 'high-energy' bases. The selection rule in Step 2 (Eq. 27) takes each high-energy translation center in W_m and adds the two nearest centers in W_{m+1} in each dimension (via U(.,.)). This implicitly assumes that the significant coefficients at resolution m+1 are located near the significant coefficients at resolution m. Corollary 1 does not provide this: it only bounds the total energy of coefficients outside a time-frequency box B_epsilon, and says nothing about where the large coefficients inside B_epsilon lie. For functions whose dominant frequency content is spatially separated from the low-frequency content (e.g., a low-frequency bump at one location plus a high-frequency bump at a distant location), the nearest-neighbor rule adds bases near the low-frequency region, missing the high-energy next-scale bases. The outer loop (increasing mu_up to 1, then incrementing m) eventually adds more bases, but only at the cost of many extra parameters, negating the claimed reduction versus a conventional constructive WNN. Remark 8 claims to handle non-stationary frequency distributions but offers no analysis of this spatially shifted case. Examples 1 and 4 use smooth functions whose frequency content is concentrated in one region, so they cannot validate the locality assumption. Without a proof or experiment covering spatially separated scale-dependent energy, the parameter-reduction claim is not established for general nonlinear mappings.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a constructive wavelet neural network (CWNN) for approximating unknown nonlinear mappings with a preset accuracy. The method has two components: Algorithm 1 estimates the dominant frequency of the unknown function by projecting data onto wavelet subspaces and estimating subspace energies; Algorithm 2 starts from the estimated initial resolution and adds wavelet bases in the next resolution that are nearest to current high-energy translation centers, stopping when the training loss falls below epsilon. The theoretical part (Theorem 1, Corollary 1) states that, under energy concentration in a time-frequency box, wavelets outside an epsilon-expanded box have negligible coefficients. Four experiments compare the CWNN with standard WNN and fixed GNN baselines on synthetic and real milling data and report a reduced parameter count.","tokens_in":29113,"tokens_out":10779,"duration_ms":131554,"significance":"If the parameter-reduction claim is established, the framework would be a useful practical contribution to constructive wavelet networks. The paper supplies a clean statement of the time-frequency truncation result, and the four examples address offline, multi-dataset, online time-varying, and real-world settings, with the real milling benchmark being a reasonable test. It also provides a code-release link and a direct comparison against standard WNN and several GNN baselines. However, the paper's central efficiency claim is not backed by an end-to-end guarantee: the theoretical truncation result concerns the full set of bases in a known time-frequency box, whereas the algorithm selects bases via local energy heuristics. The current experiments use smooth functions whose significant scale content is co-located, so they do not exercise the main risk identified below.","major_comments":[{"comment":"The basis-expansion rule selects the two nearest W_{m+1} centers around each high-energy W_m center. Corollary 1 does not justify this locality assumption: it only bounds the total energy of coefficients with centers outside B_epsilon and says nothing about where the large coefficients inside B_epsilon lie. For a function with a low-frequency bump at one location and a high-frequency bump at a distant location, Step 2 adds next-scale bases near the low-frequency region and misses the high-energy next-scale bases; the outer loop eventually repairs this only by adding many extra bases, negating the reported parameter reduction. Remark 8 and Example 3 do not address this case: the frequency shift in Example 3 is a change over time of the whole mapping, not spatially separated scale-dependent energy. Please provide a theorem or a dedicated experiment for spatially separated multiscale content before claiming the efficiency result.","section":"IV-C, Eq. (27)"},{"comment":"Algorithm 1's stopping rule assumes E_m is unimodal in m; if E_m is multimodal, the EMA rule can stop at a local peak and select an initial frequency that misses the dominant spectral content. Assumption 2 is not proved, and the unimodality evidence is a single function in Example 1 (Fig. 5(b)). Since the initial frequency feeds Algorithm 2 and strongly affects the parameter count, the efficiency claim depends on this assumption. Please either prove a sufficient condition, show that the algorithm is robust to multimodal E_m with a dedicated experiment, or characterize the failure mode.","section":"III, Algorithm 1, Assumption 2"},{"comment":"The theoretical result bounds the error incurred by omitting all wavelets outside a specified B_epsilon, assuming the concentration box [-T,T] x ([Omega_0,Omega_1] union [-Omega_1,-Omega_0]) is known. Algorithm 2 does not construct B_epsilon and does not use T, Omega_0, Omega_1; it selects bases by thresholding estimated energies and by nearest-neighbor proximity. No argument is given that the bases selected by Algorithm 2 form a superset of the significant bases inside B_epsilon or that the Theorem 1 error bound applies to the algorithm's output. Consequently the statement in the conclusion that the approximation error is bounded by the function's energy and the number of wavelet bases is not connected to the implemented algorithm.","section":"IV, Theorem 1 and Corollary 1 vs Algorithm 2"}],"minor_comments":[{"comment":"Equation (42) is used in Algorithm 1 Step 1, Step 2, and in Algorithm 2 before it is defined in Appendix C; the energy-estimate definition should be moved to the main text and the cross-references renumbered.","section":"III, Algorithm 1 and IV-C"},{"comment":"The example in Section II-C refers to 'conditions (4), (35), and (36)' and 'Using (34)', but the conditions are (4)-(6) and the construction is (7); the Appendix numbering (33)-(36) is inconsistent with the main text.","section":"II-C and Appendix B"},{"comment":"The introduction and the conclusion mention 'five numerical examples' and 'five simulation examples', while Section V presents four examples (Examples 1-4); the counts should be reconciled.","section":"I and VII"},{"comment":"Remarks 5 and 6 give inconsistent guidance for the termination threshold: Remark 5 recommends 0.01 epsilon to 0.001 epsilon, while Remark 6 says zeta should be '10 times smaller' than the desired accuracy; please reconcile the two recommendations.","section":"IV, Remarks 5 and 6"},{"comment":"The prose around Step 3 of Algorithm 2 describes selection of N_{2 mu E_m} bases after one network run, but the pseudocode uses an outer loop with mu_up incremented by mu; the description and pseudocode should be aligned.","section":"IV-C, Step 3 and Algorithm 2"},{"comment":"The empirical mapping from epsilon to alpha is written as alpha = 2 arctan(-lg epsilon)/pi in Eq. (11), but Eq. (12) and the surrounding text use arctan(-lg epsilon) together with pi/2 factors in a way that is not consistent with Eq. (11); the derivation should be cleaned up.","section":"III, Eqs. (11)-(12)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising practical idea, but the main efficiency claim is not yet established because the theoretical truncation result and the implemented basis-selection heuristic are not connected. The extensive equation-numbering and cross-reference errors suggest the manuscript needs careful editorial revision before it is suitable for publication. I would also ask the authors to soften or support the 'for the first time' claim in the abstract, since the related work section does not systematically compare against prior frequency-estimation or basis-selection heuristics for wavelet networks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a sensible core—estimate the dominant frequency of the unknown function from data, then grow the wavelet basis by energy ranking—and it shows convincing parameter reductions on its four examples (420 to 174, 39,366 to 1,510). That is a genuine engineering contribution within constructive wavelet networks. The writing is rough, but the ideas are coherent and the empirical work is honest.\n\nWhat is actually new: the specific combination of Algorithm 1's energy-based initial frequency estimator and Algorithm 2's energy-ranked basis increase is not in the cited literature. The authors also provide a real L2 bound (Theorem 1) connecting approximation error to energy outside a time-frequency box, and Corollary 1 gives a legitimate reason to skip many wavelet bases. For the tested functions, the method does what it claims.\n\nSoft spots, in proportion:\n\n1. The load-bearing assumption in Algorithm 2 is that high-energy wavelet coefficients at scale m+1 sit near the high-energy translation centers at scale m. Corollary 1 does not imply this; it only bounds total energy outside a time-frequency box and says nothing about the spatial arrangement of large coefficients inside it. A function with a low-frequency bump in one region and a high-frequency bump far away would defeat the nearest-neighbor selection rule. The outer loop (increasing mu_up, then m) would eventually add the needed bases, but at much higher parameter cost, undermining the paper's efficiency claim. Remark 8 claims robustness to non-stationary spectra but gives no analysis, and Examples 1 and 4 use functions with concentrated, roughly stationary spectra. Example 3 shifts frequency content in time, not space, and even there the method needed a large delay to recover. This is the biggest gap.\n\n2. Assumption 2 (unimodal subspace energy) supports Algorithm 1's stopping rule. It is justified only by a heuristic bandwidth argument, one plotted example, and a claim that it is \"not restrictive\"—which is not a proof. If subspace energy is multimodal, the initial frequency can land on a local peak and miss dominant spectral content.\n\n3. The efficiency comparisons are against a generic WNN, not against prior constructive wavelet algorithms like Xu and Yan (2011) or Kugarajah and Zhang (1995), which are the natural baselines. Without those, the incremental improvement is unclear.\n\n4. The manuscript has many editorial inconsistencies: equation (42) is cited before it is defined, equation numbers are off, and reference [37] is dated 2027. These are minor but they make replication harder.\n\n5. Convergence to accuracy ε is demonstrated empirically via training loss; Theorem 1 does not guarantee that the greedy selection reaches ε with the claimed parameter count. That is fine as an experimental claim, but it should be stated as such.\n\nWho this is for: researchers working on constructive wavelet networks or adaptive function approximation. It deserves a serious referee, but a conditional accept at best. I would send it to review and require: release the code, add prior constructive WNN baselines, and address the spatially separated frequency-content case, either by a proof of the locality assumption or by experiments specifically designed to break it.","headline":"A useful two-step constructive WNN idea with real parameter savings on its examples, but the efficiency claim leans on an unproved locality assumption that the experiments never stress.","tokens_in":29631,"tokens_out":2478,"would_cite":false,"duration_ms":34798,"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":"The paper claims that a constructive wavelet neural network can estimate an unknown function's dominant frequency from data and then add only high-energy wavelet bases, reaching a preset accuracy with fewer parameters than a generic…","keywords":["constructive wavelet neural networks","wavelet frequency estimation","wavelet basis selection","time-frequency localization","energy-based basis increase","function approximation","computational complexity","nonlinear system identification"],"falsifier":"Construct a test function on a bounded interval whose wavelet-subspace energies have two well-separated and comparable peaks, for instance a sum of two sinusoids with frequencies several octaves apart and similar amplitudes; run Algorithm 1 and check whether it stops at the lower peak while Algorithm 2 then needs more bases than a standard wavelet neural network to reach the same accuracy.","tokens_in":28597,"feed_emoji":"🌊","tokens_out":8059,"duration_ms":87033,"temperature":0.7,"pith_summary":"This paper tries to show that a constructive wavelet neural network need not guess where an unknown nonlinear function lives in frequency. Instead, the network first estimates the energy of the function projected onto wavelet subspaces from input-output data, uses a smoothed energy comparison to pick an initial wavelet frequency, and then adds only the wavelet bases whose coefficients are not nearly zero. If this works, it removes a major practical bottleneck: choosing the wrong starting frequency either slows convergence or wastes most of the added basis functions. The authors claim concrete savings in their examples, including 174 versus 420 parameters on a synthetic mapping and 1,510 versus 39,366 parameters on a nine-dimensional milling-machine regression.","feed_headline":"Frequency-first wavelets hit target accuracy with fewer parameters","feed_subtitle":"A constructive wavelet network estimates an unknown function's dominant frequency, then adds only the bases with real energy.","key_machinery":"The load-bearing object is the enlarged time-frequency box $B_\\varepsilon(T,\\Omega_0,\\Omega_1)$: an orthonormal wavelet $\\psi_{mn}$ with scale $m$ and translation center $n$ is kept only when $2^m$ lies inside the extended frequency band and $|n| \\preceq 2^m T + t_\\varepsilon$. Corollary 1 says wavelets outside this box have nearly zero coefficients, so the basis-expansion loop can skip them. The energy estimates $\\hat E_m$ are computed from the data by training coefficients with one neural-network pass per subspace, then smoothing them with an exponential moving average ($\\alpha$ chosen from the target accuracy $\\varepsilon$) so the algorithm can stop at the first resolution where the smoothed energy stops rising. The separation factor $\\mu$ in Algorithm 2 controls how aggressively the highest-energy bases at the current resolution are used to select the next-resolution bases.","core_discovery":"For functions in $L^2(\\mathbb{R}^d)$ whose energy is concentrated in a finite time-frequency box, the paper proves an approximation bound (Theorem 1): truncating the wavelet expansion to an enlarged box $B_\\varepsilon$ costs at most the energy of the function outside the box plus $2(2\\pi)^{d/2}\\varepsilon\\|f\\|$. Corollary 1 then states that every wavelet coefficient whose time-frequency center lies outside $B_\\varepsilon$ is essentially zero at accuracy $\\varepsilon$. The paper's constructive wavelet neural network turns this into a practical recipe: Algorithm 1 estimates the subspace energies $\\hat E_m$ with one-pass coefficient updates and an exponential moving average to select the initial frequency, and Algorithm 2 repeatedly adds the highest-energy bases from the next resolution level until the loss falls below $\\varepsilon$. The central claim is that this frequency-aware construction reaches the target accuracy with far fewer bases than a standard wavelet neural network, because the omitted bases are exactly the ones whose coefficients are near zero.","pith_inferences":["If the unimodality assumption fails, a practical extension would be to run Algorithm 1's energy scan from both low and high resolutions and keep the higher-energy peak, which would catch multimodal spectra instead of stopping at the first local peak.","The same energy-first selection principle should transfer to other dictionaries such as Fourier packets, curvelets, or shearlets whenever an analogue of the time-frequency box and a coefficient decay estimate exist.","On the milling dataset, the random 80/20 split probably overstates generalization because tool-wear observations are autocorrelated within each trial; a trial-wise split that trains on some machining configurations and tests on unseen ones would provide a sharper test of the claimed parameter savings.","The complexity argument would be stronger with a demonstration that the estimated initial frequency remains reliable as the input dimension grows beyond nine, since the per-dimension parameter cost is the main scaling bottleneck."],"forward_implications":["For any square-integrable mapping whose energy concentrates in a finite time-frequency box, Theorem 1 justifies dropping every wavelet basis outside an enlarged box without exceeding the target error, apart from energy leaked outside the box plus a small constant times $\\|f\\|$.","Starting the network at the estimated dominant frequency should avoid both slow convergence from a too-low starting frequency and wasted bases from a too-high one.","Because the wavelet bases are orthogonal, newly added bases do not disturb already-trained coefficients, so the incremental structure can expand without full retraining.","The framework is claimed to work in offline, combined-dataset, online time-varying, and nine-dimensional real-data settings, implying the frequency-estimation step transfers across those regimes.","The parameter counts reported in Examples 1 and 4 give the concrete promise: 174 versus 420 parameters on the synthetic mapping, and 1,510 versus 39,366 on the milling regression."],"supporting_citations":[{"why":"Daubechies' time-frequency localization results supply the concentration boxes and coefficient decay that Theorem 1 and Corollary 1 extend to several dimensions.","marker":"[28]"},{"why":"Kugarajah and Zhang's multidimensional wavelet frames give the single-scaling construction of the d-dimensional wavelet bases and the Lemma 1 decay bounds used in the proof of Theorem 1.","marker":"[31]"},{"why":"Mallat and Hwang's theorem on wavelet transforms and function regularity is cited to justify adding new bases at higher resolutions near the previous translation centers.","marker":"[30]"},{"why":"Mallat's multiresolution analysis defines the scaling subspaces $V_m$ that Algorithm 2 uses alongside $W_m$ for the low-frequency part of the approximation.","marker":"[29]"},{"why":"Xu and Yan's constructive wavelet network is the baseline whose resolution increase from $m=4$ to $m=5$ motivates the energy-based basis-increase mechanism.","marker":"[32]"},{"why":"The milling dataset supplies the real-world high-dimensional time series used in Example 4.","marker":"[37]"},{"why":"Lu et al.'s feature-engineering approach defines the eight input features and the normalization used for the milling-data regression.","marker":"[38]"}],"fun_headline_variants":["Energy-aware wavelet net cuts bases to hit accuracy","Frequency-first basis selection trims wavelet net","Wavelet net adds only high-energy bases, less compute","Constructive wavelet net targets dominant frequency","Smart wavelet selection: add only needed bases"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 2: the energy of the unknown function projected onto wavelet subspaces rises and then falls with resolution, so the first time the smoothed energy stops increasing the algorithm has already found the dominant frequency; if the energy has several comparable peaks, early stopping can miss the main spectral content.","fun_headline_variants_meta":{"raw":{"variants":["Energy-aware wavelet net cuts bases to hit accuracy","Frequency-first basis selection trims wavelet net","Wavelet net adds only high-energy bases, less compute","Constructive wavelet net targets dominant frequency","Smart wavelet selection: add only needed bases"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1239,"prompt_tokens":947,"completion_tokens":292,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":223}},"tokens_in":563,"tokens_out":292,"duration_ms":4389,"temperature":1.0,"reasoning_tokens":223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:11.636825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a test function on a bounded interval whose wavelet-subspace energies have two well-separated and comparable peaks, for instance a sum of two sinusoids with frequencies several octaves apart and similar amplitudes; run Algorithm 1 and check whether it stops at the lower peak while Algorithm 2 then needs more bases than a standard wavelet neural network to reach the same accuracy.","supporting_citations":[{"cited_title":"The wavelet transform, time-frequency localization and signal analysis,","cited_arxiv_id":null,"evidence_quote":"Daubechies' time-frequency localization results supply the concentration boxes and coefficient decay that Theorem 1 and Corollary 1 extend to several dimensions."},{"cited_title":"Multidimensional wavelet frames,","cited_arxiv_id":null,"evidence_quote":"Kugarajah and Zhang's multidimensional wavelet frames give the single-scaling construction of the d-dimensional wavelet bases and the Lemma 1 decay bounds used in the proof of Theorem 1."},{"cited_title":"Singularity detection and processing with wavelets,","cited_arxiv_id":null,"evidence_quote":"Mallat and Hwang's theorem on wavelet transforms and function regularity is cited to justify adding new bases at higher resolutions near the previous translation centers."},{"cited_title":"Multiresolution approximations and wavelet orthonormal bases of l2(r),","cited_arxiv_id":null,"evidence_quote":"Mallat's multiresolution analysis defines the scaling subspaces $V_m$ that Algorithm 2 uses alongside $W_m$ for the low-frequency part of the approximation."},{"cited_title":"Adaptive learning control for finite interval tracking based on constructive function approximation and wavelet,","cited_arxiv_id":null,"evidence_quote":"Xu and Yan's constructive wavelet network is the baseline whose resolution increase from $m=4$ to $m=5$ motivates the energy-based basis-increase mechanism."},{"cited_title":"Mill data set,","cited_arxiv_id":null,"evidence_quote":"The milling dataset supplies the real-world high-dimensional time series used in Example 4."},{"cited_title":"Weak monotonicity with trend analysis for unsupervised feature evaluation,","cited_arxiv_id":null,"evidence_quote":"Lu et al.'s feature-engineering approach defines the eight input features and the normalization used for the milling-data regression."}],"review_version":1}