acceptodds
Under review as a conference paper at ICLR 2027

LoKAN: Back prop-Free Node-Local Functional-Edge Learning for Equation Discovery

Abstract

Kolmogorov-Arnold network (KAN) model nonlinear relationships through functions on edges, but their many learnable edge parameters typically require repeated updates through backpropagation, making training costly. Such iterative optimization does not directly exploit the linear structure of edge coefficients under fixed basis functions to improve training efficiency. Fitting edges independently, in turn, ignores the coupling among incoming edges of the same node and generally cannot guarantee the joint optimum for a fixed support. To address these limitations, we propose LoKAN, a single-layer functional-edge model for equation discovery. Given task-specified candidate features, LoKAN constructs anchored piecewise-linear spline edges, jointly solves a ridge-regression system over the active incoming edges of each output node, iteratively prunes edges by their normalized contributions, and refits on the final support, without backpropagating through edge coefficients. In addition, LoKAN uses exact Gram-component partitioning to perform component-wise solves, reducing matrix storage requirements. Separate efficiency and scaling experiments show, respectively, that LoKAN achieves shorter per-fit training times than the compared backpropagation-trained KAN and spline group lasso, and that component-wise solves reduce Gram-matrix storage. Across the experiments, LoKAN demonstrates strong predictive performance and effectively recovers explicit candidate terms.

Primary area:
interpretability and explainable AI
Keywords:
Kolmogorov–Arnold network, BP-free, equation discovery, oint optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_7d1GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LoKAN, a single-layer functional-edge model for equation discovery. Given a task-specified candidate feature map, the method constructs fixed anchored piecewise-linear spline bases and jointly estimates the coefficients of all active incoming edges at each output node through ridge regression. It then iteratively prunes edges according to their normalized output contributions on the training data and performs a final refit on the selected support. Since the model is linear in its edge coefficients once the bases are fixed, coefficient estimation does not require backpropagation. The paper establishes the uniqueness of the solution for a fixed representation, support, and regularization, and characterizes the conditions under which the system can be partitioned into exact component-wise solves based on its cross-Gram structure. The model’s expressivity depends on the candidate features: raw-coordinate candidates yield additive nonlinear models, explicit interactions require corresponding features in the candidate map, and two-knot linear edges allow coefficient estimation for explicit candidate equation terms. The experiments cover noisy dynamical systems, measured system-identification benchmarks, correlated and missing candidate terms, nonlinear function fitting, and joint-estimation ablations. The paper reports the lowest test NRMSE on all six noisy-derivative tasks and all four candidate-term fitting tasks, and demonstrates faster fitting than BP-KAN and Spline Group Lasso in a separate efficiency evaluation.

Critical strengths
  1. The paper presents a clear and implementable pipeline for joint estimation and structural selection. The fixed spline representation reduces coefficient estimation to a linear-system solve; node-level joint estimation preserves cross terms among correlated edges; and contribution-based pruning followed by final refitting connects coefficient estimation with structural selection. This design is practically useful for scientific modeling tasks that require repeated fitting of candidate equations. The joint-estimation ablations also show that independent edge-wise fitting can lead to substantially larger errors.

  2. The experiments cover multiple scientific modeling settings and clearly demonstrate the role of the candidate representation. Experiments on dynamical systems, measured system-identification sequences, correlated candidate terms, and nonlinear function fitting provide evidence that the method is applicable across several settings. In particular, the comparison between raw inputs and inputs augmented with interaction features clearly illustrates how the candidate map determines the expressive capacity of the single-layer model. This explicit discussion of the method’s applicability and limitations helps clarify how LoKAN should be interpreted and used.

Critical weaknesses
  1. The paper does not clearly establish the methodological gap that motivates LoKAN. In the single-layer setting with a fixed candidate map and fixed basis functions, the model is linear in its learnable coefficients. Both direct coefficient estimation without backpropagation and joint estimation that preserves coupling between correlated features can therefore be achieved through standard least squares or ridge regression. In particular, with two-knot linear edges, the method is closely related to sequentially thresholded ridge regression over a candidate library; with multiple knots, it is directly related to additive regression with fixed spline bases. Although the paper discusses these related approaches, it does not sufficiently explain their remaining limitations or how LoKAN’s specific design addresses them. The algorithmic integration is clear, but the substantive advance beyond existing methods remains insufficiently established.

  2. The connection between the structure-selection mechanism and reliable equation identification with correlated candidates is insufficiently justified. Joint ridge regression can stabilize coefficient estimation on a fixed support, but uniqueness of the regularized solution does not imply correct attribution to candidate terms or recovery of the correct equation structure. LoKAN prunes edges according to the magnitude of their fitted outputs. This score reflects the contribution allocated by the current joint fit, rather than whether a candidate remains necessary after the other edges are refitted. Consequently, correlated redundant terms may be retained, while weak but necessary true terms may be removed. The paper does not provide sufficient justification for why this selection rule should improve structural identification under the candidate correlations emphasized in its motivation.

  3. The result-selection and reporting protocol is insufficient to establish a consistent performance advantage. Section 4 states that the lowest test NRMSE among the recorded runs is reported, but it does not clearly specify whether the minimum is selected across random seeds, hyperparameter configurations, or training checkpoints. Reporting only the best-performing run makes it difficult to determine whether LoKAN’s advantage is consistent and may make the comparison sensitive to differences in the number of trials or variability across methods. The relationship between this statement and the validation-based configuration selection followed by final test evaluation described in Section 3.3 also requires clarification. A clearly specified selection protocol and statistics across repeated runs are needed to support the empirical conclusions.

Questions for authors
  1. What specific shortcomings of existing joint ridge regression and candidate-library sparse regression methods remain unresolved in the single-layer, fixed-basis setting considered here? Please identify LoKAN’s key additional mechanism relative to these approaches and explain how it provides a substantive benefit for the equation-discovery tasks studied in the paper.

  2. Why is the magnitude of an edge’s fitted output an appropriate structure-selection criterion when candidate terms are correlated? Please explain how this criterion distinguishes necessary true terms from redundant terms that receive substantial contributions in the current fit, and how it translates stable coefficient estimation on a fixed support into more reliable equation identification.

  3. How is the “lowest test NRMSE” in Section 4 selected, and is the test set used to select configurations or checkpoints? Please specify the selection scope and number of runs for each method, and explain how this protocol corresponds to the validation-based selection procedure in Section 3.3. Could the authors report test means together with standard deviations or confidence intervals across repeated runs after selecting and fixing configurations using the validation set, so that the consistency of the reported advantages can be assessed?

4 comments

W1 / Q1: Additional mechanisms and their practical benefits

LoKAN's additional mechanisms address how to learn functional effects and select sparse structures simultaneously under a fixed representation. Ridge regression on a fixed support provides joint coefficient estimation. Building on this, LoKAN adds contribution scoring at the functional edge level, iterative pruning, and final refitting to determine a separate edge support for each output.

Compared with sparse regression over a candidate library, when K=2K = 2, LoKAN adjusts the ridge scale according to the active design and combines this with threshold-based selection to recover explicit terms and coefficients. When K>2K > 2, each candidate edge contains multiple spline coefficients, and the unit of selection becomes an entire nonlinear effect function. Compared with Spline Group Lasso, LoKAN separates regularized coefficient estimation from explicit structure selection: the ridge parameter controls estimation regularization, the contribution threshold determines edge retention, and estimation in each round uses a direct joint solve.

We evaluate the practical benefits of this design through both search cost and prediction accuracy. In the additional Duffing and Pendulum experiments, BP-KAN and LoKAN use matched forward basis functions, allowing comparison of the two learning procedures under matched representations. HKAN and MatrixKAN provide architectural comparisons.

Table 1. Search-stage runtime and test accuracy across five random seeds.

MethodDuffing search time (s) ↓Duffing test NRMSE ↓Pendulum search time (s) ↓Pendulum test NRMSE ↓
LoKAN0.0272 (0.0026)3.189e-3 (9.1e-4)0.151 (0.0040)9.017e-4 (4.9e-5)
BP-KAN2.51 (0.037)1.386e-2 (7.8e-3)6.27 (0.036)9.518e-3 (9.2e-4)
HKAN0.181 (0.014)4.408e-2 (3.3e-3)0.425 (0.0028)2.713e-2 (8.0e-3)
MatrixKAN2.43 (0.011)2.440e-2 (8.4e-3)4.53 (0.077)3.021e-3 (1.8e-4)

Values are medians across five seeds, with interquartile ranges (IQRs) in parentheses. Search time excludes shared preparation and final retraining. BP-KAN and LoKAN use matched forward basis functions; HKAN and MatrixKAN are independent implementations in the project used as architectural comparisons. Pendulum uses a common nine-column candidate library.

On both tasks, LoKAN achieves the lowest median search time and test NRMSE among the four methods, demonstrating the cost and accuracy benefits of this procedure. Fixed-library recovery with K=2K = 2 is separately evaluated against STLSQ and SR3 as direct comparisons. When K>2K > 2, scalar weighting and functional edge learning use different representation spaces, and computational comparisons are distinguished accordingly.

W2 / Q2: Contribution scoring and structure identification with correlated candidates

The contribution score evaluates each entire edge using its actual function output. LoKAN performs selection after joint fitting and refits the retained edges. On the correlated-distractor and weak-true-term tasks from the original Section 4.2, we compare models with and without pruning using the same data, basis functions, and joint ridge solver. We retain the original validation configurations frozen in advance and report all five paired seeds.

Table 2. Five-seed structure selection ablation under matched conditions.

TaskMethodSupport F1 ↑Exact support ↑False positives ↓Weak-term retention ↑Test NRMSE ↓
Correlated distractorsWithout pruning0.75 ± 0.000/52.0 ± 0.0—0.02134 ± 0.00030
Correlated distractorsContribution-based pruning1.00 ± 0.005/50.0 ± 0.0—0.02132 ± 0.00030
Weak true termsWithout pruning0.60 ± 0.000/54.0 ± 0.05/50.02340 ± 0.00057
Weak true termsContribution-based pruning1.00 ± 0.005/50.0 ± 0.05/50.02334 ± 0.00063

Values are means ± sample standard deviations; counts indicate successes out of five, and “—” indicates that no weak true term is present. Both strategies fix K=2K = 2, use the same base regularization parameter, and refit on the combined training and validation sets. Support is measured using the terms actually retained, excluding the intercept.

Pruning exactly recovers the support on both tasks (5/5), removes all false positives, and retains the weak true terms (5/5). Test NRMSE remains essentially unchanged, so the benefit is directly reflected in structure recovery.

To examine whether a term remains necessary after the other terms are refitted, we perform 20 edge-deletion-and-refitting evaluations from a common initial model, holding the basis functions, standardization, and ridge scale fixed. Deleting a true interaction term increases validation MSE by an average of 0.0213 and 0.00128 on the two tasks, respectively; deleting the most strongly correlated distractor produces an average change of approximately 5×10−75 \times 10^{-7} on each task. Across all five seeds, the contribution scores rank true terms and distractors consistently with the post-refitting loss rankings, supporting term identification in this setting.

Contribution magnitude and the loss after deletion and refitting have different meanings, and the results above apply to the candidate libraries tested. When K=2K = 2, the contribution score is proportional to the coefficient norm, and the retained sets are identical after mapping the thresholds accordingly.

W3 / Q3: Selection procedure and statistics across runs

For each method and task, we conduct runs with five seeds using the same data splits and predefined, comparable search budgets. For each seed, hyperparameters and, where applicable, training checkpoints are selected using the validation set. Once these choices are frozen, the test set is used only for final evaluation. LoKAN's model is determined by the pruning stopping rule and final refitting.

The “lowest test NRMSE” in the original paper refers to the minimum across the five seed runs, rather than selection across hyperparameters or checkpoints based on test error. One-step prediction and rollout are summarized as separate tasks, and their respective minima may correspond to different seeds. We now additionally report the mean and sample standard deviation across all five runs to show typical performance and run-to-run variability.

Based on unrounded means, LoKAN achieves the lowest mean error on 11 of the 17 tasks, with an average rank of 1.59. Equation term recovery fixes K=2K = 2, whereas nonlinear function learning selects K>2K > 2 using the validation set; structural metrics for these two types of tasks are interpreted separately.

Table 3. Test NRMSE (×10⁻³) across five runs, reported as mean ± sample standard deviation. Bold indicates the lowest unrounded mean for each task.

Task / summary metricLoKANSR3Spline GLBP-KANHKANMatrixKAN
LV, direct0.8274 ± 0.150.9264 ± 0.0761.558 ± 0.3513.24 ± 0.97164.2 ± 1513.02 ± 2.0
Duffing, direct1.951 ± 0.561.794 ± 0.272.910 ± 0.187.708 ± 1.477.14 ± 2620.97 ± 2.1
VdP, direct0.8636 ± 0.321.173 ± 0.171.158 ± 0.185.664 ± 0.88263.7 ± 1112.14 ± 0.40
Lorenz-63, direct0.9559 ± 0.101.418 ± 0.202.076 ± 0.3021.02 ± 1.0145.9 ± 6.937.60 ± 0.57
FHN, direct0.9941 ± 0.260.9927 ± 0.251.885 ± 0.0554.019 ± 0.37389.3 ± 6.125.96 ± 0.55
Rossler, direct1.036 ± 0.0611.969 ± 0.132.063 ± 0.1818.93 ± 0.68167.1 ± 2138.72 ± 0.50
LV, state30.28 ± 1.532.59 ± 2.531.40 ± 1.735.71 ± 1.4188.6 ± 9.441.09 ± 5.5
Duffing, state16.90 ± 0.5324.92 ± 2.918.15 ± 0.3825.89 ± 1.1125.1 ± 5217.39 ± 1.5
VdP, state134.5 ± 2.3136.5 ± 0.58137.0 ± 0.57136.0 ± 0.86340.6 ± 8.2140.6 ± 0.30
Lorenz-63, state753.3 ± 3.0753.3 ± 3.1753.2 ± 3.1750.8 ± 1.4759.5 ± 1.2760.2 ± 2.2
FHN, state84.83 ± 1.386.87 ± 0.7490.49 ± 1.987.10 ± 1.8579.9 ± 3.9127.2 ± 0.86
Rossler, state435.1 ± 71475.5 ± 28476.3 ± 28470.9 ± 30482.9 ± 26486.9 ± 24
N-body55.39 ± 1161.19 ± 5.650.42 ± 5.650.80 ± 5.683.26 ± 1151.74 ± 6.2
Silverbox, 1-step37.13 ± 0.05020.73 ± 0.0008241.45 ± 0.02741.13 ± 0.028143.6 ± 9.915.44 ± 0.085
Silverbox, rollout88.50 ± 1.6169.1 ± 0.089113.1 ± 0.097106.0 ± 0.73600.1 ± 5.970.69 ± 0.66
Tanks, 1-step24.64 ± 0.1430.70 ± 4.924.89 ± 0.06563.65 ± 0.6458.94 ± 1.384.02 ± 0.44
Tanks, rollout229.5 ± 4.6487.4 ± 4.6270.7 ± 2.0447.1 ± 161663 ± 500852.6 ± 4.9
Average rank ↓1.592.882.943.475.764.35
Number of task wins ↑1121102

I thank the authors for the detailed response and for conducting the additional five-seed evaluations and pruning ablations.

Acknowledgment of W2 & W3: The newly provided matched-condition ablations (Rebuttal Table 2), edge-deletion analyses, and full five-seed mean ± std statistics across all 17 tasks (Rebuttal Table 3) satisfactorily address my previous concerns regarding experimental reporting and the empirical behavior of the pruning heuristic. Accordingly, I am happy to acknowledge the improved empirical soundness of the paper.

Remaining Core Concern on W1 (Algorithmic Novelty & Necessity of the KAN Framing): Despite the stronger empirical validation, my fundamental concern regarding the methodological contribution remains unresolved. By restricting the model to a single layer with fixed, uniform spline bases, the architecture removes both the compositional depth and the adaptive basis learning that distinguish KANs from classical function approximators. Mathematically, the resulting formulation reduces to standard ridge regression over a fixed spline basis (i.e., a classical generalized additive model) paired with iterative backward elimination. In the K=2K=2 regime, it further collapses into standard sequentially thresholded linear ridge regression, where classic STLSQ remains slightly more accurate and faster (Appendix Table 5).

While decoupling joint ridge estimation from output-norm pruning is a clean engineering pipeline, framing fixed-basis ridge regression as a "backprop-free KAN" feels largely like a repackaging of well-established statistical regression techniques rather than a substantive algorithmic advance for equation discovery.

Conclusion: While I appreciate the authors' thorough empirical follow-up, I believe the algorithmic novelty falls short of the bar for acceptance at ICLR, and therefore I maintain my rating of 2: Weak rejection.

Reviewer_xqF4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LoKAN, a backpropagation-free approach to discovering interpretable equations and fitting nonlinear functions. LoKAN represents feature effects using fixed piecewise-linear spline functions, jointly estimates the incoming edges of each output node through ridge regression, and progressively prunes edges according to their contributions. Experiments on synthetic dynamical systems, real system-identification benchmarks, and candidate-term recovery tasks show the effectiveness of LoKAN.

Critical strengths
  • By replacing iterative gradient-based optimization with direct linear-system solves, LoKAN can substantially reduce fitting time, which is beneficial in both research areas and practice.
  • The proposed LoKAN is fast, simple, and inspectable. The method needs no gradient training, fits in milliseconds, and its univariate edge functions are directly interpretable.
  • Experimental studies show that LoKAN performs well across diverse equation-discovery and system-identification tasks, including noisy dynamical systems and recursive prediction.
Critical weaknesses
  • The novelty is limited: When K = 2, it reduces to SINDy. With linear edges, LoKAN is ridge regression with iterative magnitude-based thresholding and a refit. This is essentially SINDy’s STLSQ, which PySINDy already implements with ridge regularization.
  • STLSQ outperforms LoKAN: In Table 5, STLSQ beats LoKAN on NRMSE, support F1, exact support, coefficient error, and time, but why is it missing from the main experiments?
  • Although the paper demonstrates faster fitting in selected experiments and reduced Gram-matrix storage under exact decomposition, there are some concerns about the efficiency evaluation. The reported timings exclude data generation and hyperparameter search, and the cost of forming and solving dense systems can grow substantially with the number of active basis coefficients.
  • The experimental section reports the lowest test NRMSE among recorded runs, which raises a question about whether test results are used to select among runs.
Questions for authors
  • Why STLSQ is not used in the main experiments?
  • Can you provide a better timing/efficiency comparison?
  • Explain the reported results: whether it contradicts the stated validation-based selection?
3 comments

W1: Methodological novelty

LoKAN integrates functional edge representations, node-level joint estimation, and contribution-driven structure selection into a unified learning procedure without backpropagation. When K=2K = 2, LoKAN adjusts the ridge scale according to the active design and combines contribution-based pruning with joint refitting to recover explicit candidate terms and their coefficients. When K>2K > 2, spline edges additionally learn the nonlinear functional effects of candidate features.

Given the same candidate features, STLSQ assigns scalar weights to candidate terms, whereas LoKAN's spline edges learn the functions through which features affect the target. This extension reduces dependence on predefined analytical forms, allowing the same framework to support both equation term identification and nonlinear function learning. The experimental results in the paper also support this argument.

W2 / Q1: Comparison with STLSQ

STLSQ indeed achieves better results in the appendix when a complete candidate function library is available, but this does not establish that STLSQ is generally superior to LoKAN. Like SR3, STLSQ relies on a correctly specified candidate function library, and its lead in the appendix does not carry over to the tasks in the main paper. The main paper uses SR3 as a representative approach to sparse regression over a candidate library to illustrate the limitations of this class of methods. Following your suggestion, we additionally report STLSQ results on all 17 main tasks in Table 1. These results show that STLSQ does not perform well without a complete, correctly specified candidate function library.

TaskSTLSQ test NRMSE ↓
LV, direct0.00491956
Duffing, direct0.00298270
VdP, direct0.00120077
Lorenz-63, direct0.00196384
FHN, direct0.00207427
Rossler, direct0.00397745
LV, state0.06418093
Duffing, state0.01693124
VdP, state0.13951789
Lorenz-63, state0.76145187
FHN, state0.08313677
Rossler, state0.45503992
N-body0.07337714
Silverbox, 1-step0.01829905
Silverbox, rollout0.16187193
Tanks, 1-step0.02406481
Tanks, rollout1.02853825

W3 / Q2: Computational efficiency and scaling costs

STLSQ and SR3 are important sparse regression baselines for a fixed function library. For fixed-library recovery tasks with K=2K = 2, they are directly relevant comparisons, and the original paper retains the results showing their accuracy and lower computational costs. When K>2K > 2, the STLSQ and SR3 methods tested here learn scalar coefficients for candidate terms, whereas LoKAN learns edge functions through multiple spline coefficients. Thus, identical candidate inputs correspond to different representation spaces.

To evaluate the computational efficiency of KAN learning procedures, we additionally report search-stage runtimes and test accuracy on Duffing and Pendulum. BP-KAN and LoKAN use matched forward basis functions; HKAN and MatrixKAN provide architectural comparisons. Results for all four methods on both tasks are shown below.

Search-stage runtime and test accuracy

MethodDuffing search time (s) ↓Duffing test NRMSE ↓Pendulum search time (s) ↓Pendulum test NRMSE ↓
LoKAN0.0272 (0.0026)3.189e-3 (9.1e-4)0.151 (0.0040)9.017e-4 (4.9e-5)
BP-KAN2.51 (0.037)1.386e-2 (7.8e-3)6.27 (0.036)9.518e-3 (9.2e-4)
HKAN0.181 (0.014)4.408e-2 (3.3e-3)0.425 (0.0028)2.713e-2 (8.0e-3)
MatrixKAN2.43 (0.011)2.440e-2 (8.4e-3)4.53 (0.077)3.021e-3 (1.8e-4)

Values are medians across five random seeds, with interquartile ranges (IQRs) in parentheses. Search time excludes shared preparation and final retraining. BP-KAN uses forward basis functions matched to LoKAN; HKAN and MatrixKAN use the project's existing independent implementations as architectural comparisons. Pendulum uses a common nine-column candidate library.

LoKAN achieves the lowest median search time and median test NRMSE among the four methods on both tasks. This comparison presents the computational cost of configuration search alongside final prediction accuracy. In particular, the BP-KAN comparison, with matched forward representations, directly evaluates the difference between the two learning procedures.

Scaling costs depend on the active width. As stated in Section 3.4 of the original paper, for qq active coefficients, Gram matrix construction costs O(Nq2)O(Nq^2), the dense solve costs O(q3)O(q^3), and matrix storage costs O(q2)O(q^2). Pruning reduces the active width in subsequent rounds; when the exact block decomposition conditions hold, solving by components further reduces computation and storage costs. The two-task results above report search-stage costs, while this complexity analysis describes the dependence on scale.

W4 / Q3: Selection of reported results and run statistics

We did not selectively conduct or report runs. For each method and task, we conducted five runs with different random seeds and consistently reported the best NRMSE among them to show each model's performance. To address your concern, we additionally report the mean and sample standard deviation across the five runs. LoKAN achieves the lowest mean error on 11 of the 17 tasks, with an average rank of 1.59.

Table 2. Test NRMSE (×10⁻³) across five runs, reported as mean ± sample standard deviation. Bold indicates the lowest unrounded mean for each task.

Task / summary metricLoKANSR3Spline GLBP-KANHKANMatrixKAN
LV, direct0.8274 ± 0.150.9264 ± 0.0761.558 ± 0.3513.24 ± 0.97164.2 ± 1513.02 ± 2.0
Duffing, direct1.951 ± 0.561.794 ± 0.272.910 ± 0.187.708 ± 1.477.14 ± 2620.97 ± 2.1
VdP, direct0.8636 ± 0.321.173 ± 0.171.158 ± 0.185.664 ± 0.88263.7 ± 1112.14 ± 0.40
Lorenz-63, direct0.9559 ± 0.101.418 ± 0.202.076 ± 0.3021.02 ± 1.0145.9 ± 6.937.60 ± 0.57
FHN, direct0.9941 ± 0.260.9927 ± 0.251.885 ± 0.0554.019 ± 0.37389.3 ± 6.125.96 ± 0.55
Rossler, direct1.036 ± 0.0611.969 ± 0.132.063 ± 0.1818.93 ± 0.68167.1 ± 2138.72 ± 0.50
LV, state30.28 ± 1.532.59 ± 2.531.40 ± 1.735.71 ± 1.4188.6 ± 9.441.09 ± 5.5
Duffing, state16.90 ± 0.5324.92 ± 2.918.15 ± 0.3825.89 ± 1.1125.1 ± 5217.39 ± 1.5
VdP, state134.5 ± 2.3136.5 ± 0.58137.0 ± 0.57136.0 ± 0.86340.6 ± 8.2140.6 ± 0.30
Lorenz-63, state753.3 ± 3.0753.3 ± 3.1753.2 ± 3.1750.8 ± 1.4759.5 ± 1.2760.2 ± 2.2
FHN, state84.83 ± 1.386.87 ± 0.7490.49 ± 1.987.10 ± 1.8579.9 ± 3.9127.2 ± 0.86
Rossler, state435.1 ± 71475.5 ± 28476.3 ± 28470.9 ± 30482.9 ± 26486.9 ± 24
N-body55.39 ± 1161.19 ± 5.650.42 ± 5.650.80 ± 5.683.26 ± 1151.74 ± 6.2
Silverbox, 1-step37.13 ± 0.05020.73 ± 0.0008241.45 ± 0.02741.13 ± 0.028143.6 ± 9.915.44 ± 0.085
Silverbox, rollout88.50 ± 1.6169.1 ± 0.089113.1 ± 0.097106.0 ± 0.73600.1 ± 5.970.69 ± 0.66
Tanks, 1-step24.64 ± 0.1430.70 ± 4.924.89 ± 0.06563.65 ± 0.6458.94 ± 1.384.02 ± 0.44
Tanks, rollout229.5 ± 4.6487.4 ± 4.6270.7 ± 2.0447.1 ± 161663 ± 500852.6 ± 4.9
Average rank ↓1.592.882.943.475.764.35
Number of task wins ↑1121102

I thank the authors for the additional experiments. The mean and std results and the STLSQ comparison are useful. Several concerns remain.

  • Selection protocol. If I read the response correctly, the main tables report the best test NRMSE out of five seeds. Does it mean that they are picking on the test set, even if it’s done for every method, and it doesn’t match the validation-based selection described in the paper?
  • Support recovery. For an equation-discovery problem, I think support F1, exact support, and coefficient error may be more informative than NRMSE, and STLSQ is ahead on all three in Table 5. It would help to see these metrics on the main tasks where the true equations are known, ideally with LoKAN using linear edges compared against STLSQ on the same library.
Reviewer_BBeyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LoKAN is a single-layer functional-edge model that fits fixed spline features using joint ridge regression, prunes edges according to their fitted contributions, and refits the retained support. It avoids backpropagation through the edge coefficients. The paper analyzes fixed-support optimality and exact Gram-matrix partitioning, and evaluates prediction, equation-term recovery, and computational efficiency on synthetic dynamics and measured systems.

Critical strengths

The algorithm is clearly defined, and the fixed-support ridge solution is mathematically justified. The paper also carefully distinguishes coefficient optimality from support recovery and explicitly states its dependence on the supplied candidate features.

The evaluation includes noisy dynamics, measured system-identification data, and recursive prediction. The separate timing experiment shows a substantial per-fit speed advantage over the compared backpropagation-trained KAN, giving the approach potential practical value.

Critical weaknesses

The main evaluation uses test performance to select reported results. Section 4 states that each entry is the lowest test NRMSE across recorded runs, with one-step and rollout results selected separately. This conflicts with the validation-only selection described elsewhere and introduces optimistic selection bias. Without run counts, search budgets, and uncertainty estimates, the reported performance advantages are not reliable evidence of generalization.

The contribution beyond established sparse regression is insufficiently demonstrated. With two-knot edges, LoKAN reduces to ridge fitting on standardized candidate terms followed by coefficient-magnitude pruning and refitting, closely resembling STLSQ. In Table 5, STLSQ has better reported values on every recovery metric and takes 1.13 ms per fit versus LoKAN’s 9.35 ms. Yet STLSQ is absent from the main accuracy comparisons. The backfitting ablation also lacks convergence controls, although converged backfitting should recover the same fixed-support ridge solution.

Questions for authors

Which results were selected using validation data, and which were selected using test error? Please provide validation-selected results across repeated runs, with run counts and uncertainty estimates.

What advantage remains over STLSQ when both methods use identical candidate libraries and comparable validation budgets? Please include this comparison in the main equation-discovery experiments.

What stopping tolerance and iteration budget were used for backfitting? Does it reach the same objective as LoKAN when the support, basis, and ridge scale are held fixed?

Do useful exact Gram components occur in the actual benchmark designs? The current scaling experiment constructs zero cross-component blocks explicitly.

2 comments

We thank the reviewer for recognizing the algorithmic formulation and experiments in our paper. We address your concerns below.

W1 / Q1: Selection of reported results and run statistics

We would like to clarify how the results in the paper were reported. For each method and task, we conducted runs with five random seeds and reported the lowest test NRMSE among the five runs. Reporting one-step prediction and rollout separately means that, for the Silverbox and Tanks benchmarks, one-step prediction and rollout are treated as two separate tasks.

To address your concern, we additionally report the mean and sample standard deviation across the five runs and compute rankings using the means. LoKAN achieves the lowest mean error on 11 of the 17 tasks, with an average rank of 1.59. The complete statistics are provided below.

Table 1. Test NRMSE (×10⁻³) across five runs, reported as mean ± sample standard deviation. Bold indicates the lowest unrounded mean for each task.

Task / summary metricLoKANSR3Spline GLBP-KANHKANMatrixKAN
LV, direct0.8274 ± 0.150.9264 ± 0.0761.558 ± 0.3513.24 ± 0.97164.2 ± 1513.02 ± 2.0
Duffing, direct1.951 ± 0.561.794 ± 0.272.910 ± 0.187.708 ± 1.477.14 ± 2620.97 ± 2.1
VdP, direct0.8636 ± 0.321.173 ± 0.171.158 ± 0.185.664 ± 0.88263.7 ± 1112.14 ± 0.40
Lorenz-63, direct0.9559 ± 0.101.418 ± 0.202.076 ± 0.3021.02 ± 1.0145.9 ± 6.937.60 ± 0.57
FHN, direct0.9941 ± 0.260.9927 ± 0.251.885 ± 0.0554.019 ± 0.37389.3 ± 6.125.96 ± 0.55
Rossler, direct1.036 ± 0.0611.969 ± 0.132.063 ± 0.1818.93 ± 0.68167.1 ± 2138.72 ± 0.50
LV, state30.28 ± 1.532.59 ± 2.531.40 ± 1.735.71 ± 1.4188.6 ± 9.441.09 ± 5.5
Duffing, state16.90 ± 0.5324.92 ± 2.918.15 ± 0.3825.89 ± 1.1125.1 ± 5217.39 ± 1.5
VdP, state134.5 ± 2.3136.5 ± 0.58137.0 ± 0.57136.0 ± 0.86340.6 ± 8.2140.6 ± 0.30
Lorenz-63, state753.3 ± 3.0753.3 ± 3.1753.2 ± 3.1750.8 ± 1.4759.5 ± 1.2760.2 ± 2.2
FHN, state84.83 ± 1.386.87 ± 0.7490.49 ± 1.987.10 ± 1.8579.9 ± 3.9127.2 ± 0.86
Rossler, state435.1 ± 71475.5 ± 28476.3 ± 28470.9 ± 30482.9 ± 26486.9 ± 24
N-body55.39 ± 1161.19 ± 5.650.42 ± 5.650.80 ± 5.683.26 ± 1151.74 ± 6.2
Silverbox, 1-step37.13 ± 0.05020.73 ± 0.0008241.45 ± 0.02741.13 ± 0.028143.6 ± 9.915.44 ± 0.085
Silverbox, rollout88.50 ± 1.6169.1 ± 0.089113.1 ± 0.097106.0 ± 0.73600.1 ± 5.970.69 ± 0.66
Tanks, 1-step24.64 ± 0.1430.70 ± 4.924.89 ± 0.06563.65 ± 0.6458.94 ± 1.384.02 ± 0.44
Tanks, rollout229.5 ± 4.6487.4 ± 4.6270.7 ± 2.0447.1 ± 161663 ± 500852.6 ± 4.9
Average rank ↓1.592.882.943.475.764.35
Number of task wins ↑1121102

W2 / Q2: Comparison with STLSQ

STLSQ depends heavily on a correctly specified function library, much like SR3. Its leading performance in the appendix therefore does not carry over to the main experiments. In the main paper, we use SR3 as a representative approach to sparse regression over a candidate library. Following your suggestion, we additionally report STLSQ results on all 17 main tasks, showing that STLSQ does not maintain its lead on these tasks.

LoKAN's representational advantage over STLSQ is that, for the same candidate features, STLSQ estimates scalar coefficients, whereas LoKAN's spline edges additionally learn nonlinear functional effects. LoKAN thus unifies explicit equation term recovery and nonlinear effect learning within a workflow of joint fitting, contribution-based pruning, and refitting.

TaskSTLSQ test NRMSE ↓
LV, direct0.00491956
Duffing, direct0.00298270
VdP, direct0.00120077
Lorenz-63, direct0.00196384
FHN, direct0.00207427
Rossler, direct0.00397745
LV, state0.06418093
Duffing, state0.01693124
VdP, state0.13951789
Lorenz-63, state0.76145187
FHN, state0.08313677
Rossler, state0.45503992
N-body0.07337714
Silverbox, 1-step0.01829905
Silverbox, rollout0.16187193
Tanks, 1-step0.02406481
Tanks, rollout1.02853825

Q3: Convergence settings for backfitting

Backfitting stops when the maximum absolute change in any edge coefficient over a complete sweep is less than 10−510^{-5}, with a maximum of 100 sweeps.

With the support, basis functions, and ridge scale fixed, fully converged backfitting reaches the same optimum as LoKAN. LoKAN's advantage is its direct solution, which avoids slow iteration caused by correlated candidate terms and prevents finite-iteration errors from affecting subsequent pruning. Table 3 reports performance under the stopping rule and iteration budget described above.

Q4: Exact Gram block decomposition

Table 6 presents a controlled experiment that validates the exact block decomposition mechanism. We have not compiled statistics on the distribution of exact components in the real-world benchmarks.

The theoretical result provides a precise criterion: with the support and a common ridge scale fixed, the centered cross-group Gram blocks are zero if and only if solving the groups separately preserves the joint solution for every right-hand side. The connected components induced by nonzero cross-blocks give the finest exact grouping. This result therefore identifies both the groups of edges that can be solved independently and the joint couplings that must be retained. Its computational benefits are verified by the experiments in Table 6, which satisfy these conditions.

Reviewer_2wcNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduced LoKAN, an equation discovery method that tries to explain the measured data while using a list of pre-specified terms supplied by the user. It begins by using all the terms, removing those that have little contribution, and repeating the procedure. For each term, LoKAN might output a single coefficient, or a curve that describes the effect of the term.

The main benefits of LoKAN include being backpropagation-free and having an epxlicit procedure for simplifying the model by removing redundant terms.

There are experiments in the paper testing the prediction accuracy, recovery of the correct terms in the equation, and the fitting speed. I have some concerns about the experiments which I will mention below.

Critical strengths

1- I like the fitting procedure. Once the equation terms are fixed, eqs. (5) - (7) show that LoKAN aims at finding the best coefficients under its stated fitting rule, and also includes a penalty term to avoid using extremely large coefficients. Appendix A.2 also provides a nice explanation of how the pieces fit together.

2- The variety of the settings included in the experiments section is very nice and useful, and also shows the method's limits. For instance, Table 4 illustrates that LoKAN struggles when the interactions between inputs are not provided to it.

Critical weaknesses

1- One thing that I am quite concerned about is that in the experiments section, the authors mention that the paper reports the lowest test errors among the recorded runs for each method and for each task. This can create a selection bias that makes a method look better than its actual performance. Additionally, this means that if a method is given the opportunity to try for more times, it has a higher chance of getting a better result. I could not find the run counts in the paper, or see a measurement of how much the results vary between runs to actually assess the effect of this, but I am not sure how dependable the results are given this. This is perhaps my biggest concern about the paper.

2- One of the most important things to check in this paper is to see how well LoKAN is able to recover the actual equation. Most of the results in the main text report prediction error, which is not the same as equation recovery. Table 5 is one of the only places that reports this, and in that table, it seems to me that the simpler STLSQ method is performing better than LoKAN while also running faster, so I am a bit confused about the advantage of LoKAN.

Questions for authors

1- Could you please elaborate on how the final reported runs were chosen? (see W1 above)

2- Regarding W2 above, it would be nice if you could provide a comparison of equation discovery between LoKAN and other methods for the main equation discovery tasks so we can see if LoKAN is showing any advantages.

2 comments

We thank the reviewer for recognizing the fitting procedure and the breadth of the experimental evaluation. We address the two concerns below.

W1 / Q1: Selection of reported runs and variability

For each method and task, we conducted five runs with different random seeds and consistently reported the lowest test NRMSE among these runs for every model. We believe this reflects each model's best performance. In response to your concern, we additionally report the mean and sample standard deviation across the five runs. The rankings and numbers of task wins in the table below are based on the five-run means, rather than the minimum from a single run. Test NRMSE (×10⁻³), reported as mean ± sample standard deviation. Bold indicates the lowest unrounded mean for each task.

Task / summary metricLoKANSR3Spline GLBP-KANHKANMatrixKAN
LV, direct0.8274 ± 0.150.9264 ± 0.0761.558 ± 0.3513.24 ± 0.97164.2 ± 1513.02 ± 2.0
Duffing, direct1.951 ± 0.561.794 ± 0.272.910 ± 0.187.708 ± 1.477.14 ± 2620.97 ± 2.1
VdP, direct0.8636 ± 0.321.173 ± 0.171.158 ± 0.185.664 ± 0.88263.7 ± 1112.14 ± 0.40
Lorenz-63, direct0.9559 ± 0.101.418 ± 0.202.076 ± 0.3021.02 ± 1.0145.9 ± 6.937.60 ± 0.57
FHN, direct0.9941 ± 0.260.9927 ± 0.251.885 ± 0.0554.019 ± 0.37389.3 ± 6.125.96 ± 0.55
Rossler, direct1.036 ± 0.0611.969 ± 0.132.063 ± 0.1818.93 ± 0.68167.1 ± 2138.72 ± 0.50
LV, state30.28 ± 1.532.59 ± 2.531.40 ± 1.735.71 ± 1.4188.6 ± 9.441.09 ± 5.5
Duffing, state16.90 ± 0.5324.92 ± 2.918.15 ± 0.3825.89 ± 1.1125.1 ± 5217.39 ± 1.5
VdP, state134.5 ± 2.3136.5 ± 0.58137.0 ± 0.57136.0 ± 0.86340.6 ± 8.2140.6 ± 0.30
Lorenz-63, state753.3 ± 3.0753.3 ± 3.1753.2 ± 3.1750.8 ± 1.4759.5 ± 1.2760.2 ± 2.2
FHN, state84.83 ± 1.386.87 ± 0.7490.49 ± 1.987.10 ± 1.8579.9 ± 3.9127.2 ± 0.86
Rossler, state435.1 ± 71475.5 ± 28476.3 ± 28470.9 ± 30482.9 ± 26486.9 ± 24
N-body55.39 ± 1161.19 ± 5.650.42 ± 5.650.80 ± 5.683.26 ± 1151.74 ± 6.2
Silverbox, 1-step37.13 ± 0.05020.73 ± 0.0008241.45 ± 0.02741.13 ± 0.028143.6 ± 9.915.44 ± 0.085
Silverbox, rollout88.50 ± 1.6169.1 ± 0.089113.1 ± 0.097106.0 ± 0.73600.1 ± 5.970.69 ± 0.66
Tanks, 1-step24.64 ± 0.1430.70 ± 4.924.89 ± 0.06563.65 ± 0.6458.94 ± 1.384.02 ± 0.44
Tanks, rollout229.5 ± 4.6487.4 ± 4.6270.7 ± 2.0447.1 ± 161663 ± 500852.6 ± 4.9
Average rank ↓1.592.882.943.475.764.35
Number of task wins ↑1121102

W2 / Q2: Equation recovery and advantages over STLSQ

STLSQ estimates scalar coefficients over a predefined candidate library, and its performance is limited when the library is incomplete or misspecified. Its advantage in Appendix Table 5 therefore does not readily generalize to other tasks. LoKAN combines explicit term recovery with learning nonlinear effects: linear edges yield equation terms and coefficients, while nonlinear edges learn the functional effects of the given candidate features, reducing dependence on predefined analytical forms. In the main paper, we use SINDy-SR3, a method similar to STLSQ, as a representative approach to sparse regression over a candidate library. To address your concern, we additionally provide STLSQ results on the tasks in Table 1 for reference.

TaskSTLSQ test NRMSE ↓
LV, direct0.00491956
Duffing, direct0.00298270
VdP, direct0.00120077
Lorenz-63, direct0.00196384
FHN, direct0.00207427
Rossler, direct0.00397745
LV, state0.06418093
Duffing, state0.01693124
VdP, state0.13951789
Lorenz-63, state0.76145187
FHN, state0.08313677
Rossler, state0.45503992
N-body0.07337714
Silverbox, 1-step0.01829905
Silverbox, rollout0.16187193
Tanks, 1-step0.02406481
Tanks, rollout1.02853825
The equation discovery tasks in this paper are not limited to Appendix Table 5. Section 4.1 examines equation identification under noise on six known dynamical systems, and Section 4.2 further introduces correlated distractors, dense distractors, weak true terms, and missing terms. To address your concern that the main paper primarily reports NRMSE, which does not directly evaluate recovery on the main tasks, we additionally report structural metrics from the same runs as the NRMSE values in the original Table 1. These results demonstrate LoKAN's advantage in equation recovery.
MetricLoKAN
--
Average support F1 ↑0.9889
Average exact support ↑94.44%

Thank you for the response and the additional experiments.

Regarding W1, I appreciate the additional results, which suggest that LoKAN remains competitive. However, the mean results are less favourable than the originally reported minimum test errors (e.g., 11 task wins rather than 13). This reinforces my concern that selecting the best run based on test performance presents an overly optimistic picture. I am also unclear why the mean and standard deviation were not reported originally, given that five runs had already been conducted. I strongly encourage replacing the minimum-based results with the five-run statistics in the paper.

Regarding W2, the additional structural recovery results are encouraging. However, my main concern was about comparing equation recovery across methods, whereas the new comparisons with STLSQ still focus on prediction error. Therefore, I believe this concern remains only partially addressed.

Overall, I appreciate the additional evidence, but I will maintain my original assessment. Hopefully, the authors will find the suggestions above helpful for improving their paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.