acceptodds
Under review as a conference paper at ICLR 2027

JumpLie: Discovering and Exploiting Jump-Diffusion Symmetries

Abstract

Jump-diffusion (JD) processes provide a natural description of stochastic systems that combine continuous evolution with abrupt discrete events, a setting increasingly relevant to machine learning, while symmetry offers a powerful means of exposing structure and constructing inductive biases for learning. We introduce *JumpLie*, an end-to-end, data-driven framework for discovering and exploiting Lie symmetries of JD dynamics, building on the point-symmetry theory of A. Nass (2017). Our approach first learns the underlying JD dynamics from trajectory data and then discovers neural symmetry generators by enforcing the corresponding continuous and finite-jump invariance conditions. This formulation highlights a fundamental distinction from conventional diffusion-SDE symmetry: invariance of the continuous stochastic dynamics alone does not guarantee invariance under the discrete jump mechanism. We study recovery of JD symmetry structure from data and demonstrate its utility across three settings: improving out-of-distribution generalization, discovering and exploiting structure in jump-based stochastic generative models through Generator Matching, and symmetry-aware volatility estimation for financial time series. Together, these results establish JD symmetry as a learnable structural prior for systems with intertwined continuous and discrete stochastic dynamics. To the best of our knowledge, this is the first machine-learning work to introduce the data-driven discovery and downstream use of JD Lie symmetry, opening a new direction for symmetry learning beyond purely continuous dynamics.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
jump-diffusion symmetry, symmetry discovery, stochastic differential equations, neural SDEs, scientific machine learning, out-of-distribution generalization, generative modeling, volatility forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_gcKnOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces JumpLie, a framework for discovering Lie symmetries of jump-diffusion dynamics from trajectory data and using them in downstream learning. It first fits and freezes a neural surrogate for the drift, diffusion, jump displacement, and Poisson rate, then learns infinitesimal generators using determining-equation, non-collapse, and Lie-bracket closure losses. The central distinction from diffusion-only symmetry discovery is an additional finite-jump compatibility condition. The paper reports sub-degree generator-subspace recovery on three low-dimensional systems and improved out-of-distribution coefficient estimation when the discovered generators regularize a fresh surrogate. It also explores a 3,072-dimensional CIFAR-10 ship-generation process, using a flow-plus-categorical-jump operator formulation to learn directions for edits without parameter updates at editing time. Finally, it applies discovered log-price symmetry to regularize next-day realized-variance forecasts for SPY, QQQ, and IWM, reporting improvements over a matched Jump-GRU and greater robustness to nominal-price rescaling.

Critical strengths
  1. The extension to finite-jump compatibility addresses a meaningful limitation of diffusion-only symmetry discovery.

A transformation that preserves drift and Brownian diffusion need not preserve the finite jump map. JumpLie explicitly incorporates this additional constraint into a trainable discovery framework, rather than treating jump-containing observations as ordinary diffusion data. The separation between surrogate fitting and generator discovery, together with normalization and bracket closure, gives a coherent approach to learning nontrivial symmetry directions. This is a useful extension of the trajectory-to-generator approach used by LieStoNet, with a clear motivation for systems containing discrete events.

  1. The controlled recovery experiments provide encouraging quantitative evidence, and some downstream comparisons are carefully matched.

Section 4.1 reports maximum principal angles of approximately 0.55°, 0.69°, and 0.84° across five seeds, using a basis-invariant subspace metric rather than requiring a particular generator basis. Table 1 shows lower mean OOD errors for every reported coefficient, including substantial reductions in the E1-B and E1-C jump-map errors. Appendix A.2 matches architecture, initialization, and optimization between the OOD branches. The finance comparison also uses a matched Jump-GRU and reports block-bootstrap confidence intervals. These choices support the potential usefulness of symmetry regularization, even though stronger attribution controls are still needed.

Critical weaknesses
  1. The high-dimensional experiment does not adequately establish that the learned editing directions are symmetries of the stated generative process.

The main formulation uses a deterministic jump displacement driven by a scalar, constant-rate Poisson process. Appendix A.3 instead uses state-dependent coordinatewise rates and categorical jump destinations, enforcing an operator commutator on randomly sampled test functions. This is a reasonable direction for an extension, but its relationship to the main determining equations and the adequacy of the sampled tests need a clearer justification. In particular, retaining the exact expectation over all destination bins does not by itself establish commutation on a sufficiently rich family of functions. Section 4.3 and Figure 5 show one base image and three edited outputs, without quantitative held-out symmetry residuals, finite-transformation compatibility, or comparisons with norm-matched random directions and unedited stochastic continuations. Different final images alone cannot distinguish genuine symmetry actions from arbitrary perturbations or rollout randomness. Consequently, the claimed scaling of symmetry discovery to 3,072 dimensions and its use for controllable edits remain insufficiently validated.

  1. The current controls do not isolate the benefit of discovering JD symmetry, and the finance stress test contains an unexplained consistency issue.

The LieStoNet comparison changes both the surrogate model and the symmetry constraints, so it does not isolate the finite-jump condition as cleanly as removing that condition from the same frozen JD surrogate would. For the OOD study, neither Section 4.2 nor Appendix A.2 specifies the exact training/test state regions or clearly states whether the initial surrogate and generator-discovery stages were also restricted to the same in-distribution region. This leaves the scope of the generalization claim unclear.

In finance, the relevant price-scale invariance is already known. Appendix A.4 mentions an oracle-symmetry GRU and HAR-CJ, but does not report their numerical forecasting results or a full forecasting comparison with the Return-only GRU. More concretely, that appendix defines the Return-only inputs entirely through returns and realized-variation features, which are unchanged under a constant rescaling S↦cSS\mapsto cS. Nevertheless, its curve in Figure 6(b) shows nonzero changes in predicted RV. The stated deterministic feature construction would imply invariance; any additional normalization, inference randomness, or preprocessing responsible for this behavior needs explanation. Until these issues are resolved, the downstream results do not convincingly separate the value of learned JD symmetry from that of directly imposing known invariance or ordinary regularization.

Questions for authors
  1. Controlled discovery and OOD protocol:

Can you report a quantitative ablation that removes the finite-jump residuals while keeping the same frozen JD surrogate, generator capacity, and remaining losses? Please also specify the exact ID/OOD regions, the data used at every discovery stage, and where the OOD regularization collocation points are sampled. Were any OOD-region observations used to learn the frozen surrogate or generators? These details would materially affect the interpretation of the jump-specific and generalization results.

  1. High-dimensional symmetry validation:

How does the operator criterion in Appendix A.3 relate to the deterministic-jump determining equations in the main text, and what family of random test functions is used? Can you provide held-out commutator or finite-action compatibility measurements and quantitative editing comparisons against norm-matched random perturbations and unedited continuations, controlling the remaining rollout randomness? Such results would help establish that the learned directions represent symmetries rather than merely perturbations that produce different images.

  1. Finance consistency and baseline results:

Why does the Return-only GRU change its prediction under nominal-price rescaling when all four stated inputs are invariant? Please clarify preprocessing and inference during this test, and report absolute QLIKE values and paired differences for HAR-CJ, Return-only GRU, Jump-GRU, JumpLie regularization, and the oracle-symmetry GRU already mentioned in Appendix A.4. Does learned symmetry provide an advantage over imposing log-price translation directly with the same regularization weight and training protocol?

5 comments
AuthorsOpenReview ↗

Responses to reviewer gcKn (1)

We sincerely thank the reviewer for recognizing the strengths of our work and for their thoughtful and constructive feedback. We appreciate the opportunity to clarify the key aspects of our approach and strengthen the empirical evidence through additional controlled experiments. In the following, W = Weakness, Q = Question.

W1 & Q2. High-dimensional symmetry validation

We thank the reviewer for emphasizing that visually plausible edits alone do not establish genuine symmetries. We have strengthened the high-dimensional experiment with three complementary quantitative tests: 1) held-out symmetry residuals, 2) finite-edit compatibility, and 3) generative-quality preservation. We additionally demonstrate interpretable control of generated images through the discovered symmetry directions.

1) Held-out symmetry residuals. Lower residuals indicate better agreement with the symmetry equations (0 = exact satisfaction).

Evaluation (↓ better)ResultInterpretation
Flow residual ↓0.0115Low error: learned transformations approximately preserve continuous dynamics
Jump residual ↓0.0092Low error: learned transformations approximately preserve the categorical-jump mechanism

2) Finite-edit compatibility. Lower MMD² indicates that editing before versus after stochastic continuation produces more similar distributions.

Evaluation (↓ better)MMD²Interpretation
Learned edits1.68 × 10⁻⁵Near the independent-sampling reference (1.56 × 10⁻⁵), indicating approximate compatibility with stochastic continuation
Matched random edits4.77 × 10⁻⁴28× larger than learned edits, demonstrating that this compatibility is not achieved by arbitrary perturbations

3) Generative-quality preservation. Lower FID indicates better agreement with real images. Results are averaged over eight generators and five random seeds (± standard error).

Editing strengthUnedited FID ↓Learned-edit FID ↓Matched random-edit FID ↓
ε = 0.525.67 ± 0.1125.86 ± 0.3153.94 ± 0.24
ε = 1.025.67 ± 0.1126.04 ± 0.7757.31 ± 0.36

4) Interpretable control through learned symmetries. Beyond mathematical compatibility and generative-quality preservation, the discovered symmetries provide consistent, interpretable transformations of generated images. Revised Figure 5 (page 8) demonstrates that the same learned generator systematically changes the illumination and color of five distinct ship images: negative pseudo-times produce brighter, warmer scenes, while positive pseudo-times produce progressively darker, bluer scenes.

We quantify this visual effect across the five images:

Visual attribute (↑ stronger)ε = −0.4ε = +0.9Interpretation
Darkness ↑17.8469.12Substantial darkening across the learned edits
Blueness ↑0.4226.80Stronger blue coloration

In contrast, the displacement-matched random-edit visualization in Appendix A.3 (page 13) does not exhibit the same coherent illumination progression. This provides direct visual evidence that the discovered symmetry enables meaningful, interpretable control of an otherwise opaque high-dimensional generative process, without semantic editing supervision or model retraining.

Together, these results provide complementary evidence for meaningful high-dimensional symmetry discovery: (1) low held-out residuals demonstrate approximate satisfaction of both continuous and jump symmetry constraints; (2) near-reference finite-edit MMD² demonstrates approximate compatibility with stochastic continuation, unlike matched random edits; (3) near-unchanged FID establishes generative-quality preservation; and (4) consistent illumination/color transformations demonstrate interpretable, training-free generative control. Crucially, the discovered directions are not merely quality-preserving perturbations: they reveal useful, controllable structure in the learned 3,072-dimensional stochastic generative process.

Connection to the deterministic-jump equations and choice of test functions. The operator criterion in Appendix A.3 generalizes the main text's finite-jump determining equation to stochastic jump kernels. Specifically, for a deterministic jump map $T_J(x)=x+J(x) $ with constant jump rate, requiring the commutator $[H,\mathcal L_J]F=0 $ for all smooth test functions $F $ is equivalent (for the spatial generator H=ξ⋅∇H=\xi\cdot\nabla ) to:

DTJ(x) ξ(x)=ξ(TJ(x)),DT_J(x)\,\xi(x)=\xi(T_J(x)),

which is precisely the main-text finite-jump condition $ R_J=0 $: applying the symmetry before or after a jump gives the same transformation.

(continues below)

AuthorsOpenReview ↗

Responses to reviewer gcKn (2)

(W1&Q2 continues)

For categorical jumps, we instead enforce operator commutation using smooth random ridge test functions, including linear, sinusoidal, cosinusoidal, and quadratic functions of random projections of the state. This formulation additionally accounts for state-dependent jump rates and destination probabilities. Training uses four smooth probes per state, and held-out evaluation uses 12 freshly sampled probes per state, with exact summation over all 3,072 coordinates and 32 destination bins. Thus, the evaluation tests compatibility with the full categorical-jump mechanism instead of only the sampled training constraints.

These findings establish strong quantitative and visual evidence for nontrivial, interpretable, and quality-preserving approximate symmetries of the learned high-dimensional stochastic generative dynamics. We have incorporated the results into revised Section 4.3 (page 8) and Appendix A.3 (page 13).


W2 & Q1. Necessity of finite-jump constraints and controlled OOD evaluation

We thank the reviewer for identifying the need to isolate the finite-jump contribution and clarify the OOD protocol. We conducted two additional controlled experiments: (1) a same-surrogate jump-constraint ablation and (2) a four-way downstream OOD comparison, with explicitly documented data access at each stage.

1) Isolating the finite-jump contribution. We compare full JumpLie against a variant that removes the finite-jump spatial and nonlocal temporal constraints, while keeping the same frozen JD surrogate, generator architecture, initialization, collocation points, and remaining losses.

Symmetry discoveryRecovery MPA (°) ↓Jump residual MSE ↓
Full JumpLie0.767 ± 0.0030.0133 ± 0.0078
Jump constraints removed88.589 ± 0.5000.3727 ± 0.0080

(Five-seed mean ± standard deviation; smaller is better)

Removing the finite-jump constraints nearly destroys symmetry recovery, increasing the principal angle from 0.77° to 88.59° and the jump residual by approximately 28×. Since both branches use the same frozen JD surrogate, these results demonstrate that the finite-jump constraints are essential for recovering genuine jump-diffusion symmetries, beyond what continuous drift–diffusion constraints alone can achieve.

2) Downstream OOD controls. We additionally train fresh coefficient predictors using four regularization conditions, all with matched architectures and ID-start training data.

RegularizationOOD drift RMSE ↓OOD diffusion RMSE ↓OOD jump RMSE ↓
Full JumpLie0.09880.09090.0915
No symmetry0.14170.12360.1000
Continuous-only generators0.14960.07220.1594
Norm-matched wrong generators0.71500.15400.2820

Full JumpLie reduces OOD RMSE relative to no symmetry by 30.3% (drift), 26.4% (diffusion), and 8.4% (jump). It also outperforms continuous-only generators on drift and jump prediction, although the latter achieves lower diffusion RMSE. The poor performance of norm-matched incorrect generators further shows that the learned symmetry directions matter, not merely the presence of an additional regularization penalty.

3) Exact data-access and OOD protocol. In this separate controlled rerun:

  • Surrogate fitting. Training transition starts satisfy t∈[0,2]t\in[0,2], x∈[−1,1]2x\in[-1,1]^2. No shifted-OOD-start observations are used; observed transition endpoints are retained even when they cross the ID boundary.
  • Symmetry discovery. Both learners use the same frozen surrogate and ID collocation points. Finite-jump constraints can evaluate fields at jump destinations outside the ID box.
  • Downstream training. Fresh predictors use the same ID-start transitions. Symmetry regularization is collocated at ID points, with nonlocal evaluations at predicted jump destinations.
  • OOD evaluation. We translate matched ID evaluation states by x1↦x1±3x_1\mapsto x_1\pm3, holding tt and x2x_2 fixed. True coefficients are used only for scoring.

Thus, no deliberately shifted-OOD-start observations are used for fitting or discovery.

Together, these results directly establish the necessity of the finite-jump conditions for accurate JD symmetry discovery and provide controlled evidence of their downstream generalization benefit.


(continues)

AuthorsOpenReview ↗

Responses to reviewer gcKn (3)

Q3. Corrected finance consistency test and forecasting baselines

We thank the reviewer for identifying the inconsistency. The reviewer is correct that Return-only GRU should be invariant to constant nominal-price rescaling.

1) Cause and correction of the inconsistency. We traced the nonzero Return-only curve in the original Figure 6(b) to an error confined to stress-test preprocessing: normalization was applied across the full feature vector, including the nominal log-price channel (before removing that channel). Consequently, rescaling prices altered the normalization of the retained return-based features, introducing an unintended price dependence. However, training and standard forecasting evaluation used the intended preprocessing, thus this error affected the rescaling diagnostic, not the trained models or forecasting scores.

We corrected the stress test to use the original training-time preprocessing with fixed normalization statistics. Across all seven price multipliers c∈0.1,0.25,0.5,1,2,4,10c\in\\{0.1,0.25,0.5,1,2,4,10\\}, we verified zero change in Return-only raw features, normalized inputs, and RV predictions for SPY, QQQ, and IWM.

The corrected worst-case mean relative prediction changes are:

AssetReturn-only ↓Jump-GRU ↓JumpLie ↓Oracle GRU ↓
SPY00.50240.46170.4911
QQQ00.40950.27230.3680
IWM00.81080.22580.5340

(Five-seed means; lower indicates greater robustness)

JumpLie reduces rescaling sensitivity relative to Jump-GRU by 8.1% (SPY), 33.5% (QQQ), and 72.1% (IWM). This establishes improved robustness compared to other methods.

2) Expanded forecasting baselines. We additionally report absolute out-of-sample QLIKE against the classical HAR-CJ model, Return-only GRU, matched Jump-GRU, and Oracle GRU using the analytically known log-price translation.

AssetHAR-CJ ↓Return-only ↓Jump-GRU ↓JumpLie ↓Oracle GRU ↓
SPY0.293570.289260.301570.266360.30597
QQQ0.244550.258770.263700.245530.25552
IWM0.230000.217610.212690.193890.21546

(Lower QLIKE is better; neural entries are five-seed means)

JumpLie improves mean QLIKE over the matched Jump-GRU by 11.7%, 6.9%, and 8.8%, respectively. It outperforms Return-only and Oracle GRU on all three assets, achieving the best overall QLIKE on SPY and IWM (HAR-CJ is marginally better on QQQ).

3) Learned versus known symmetry. The Oracle GRU uses the exact log-price translation, with the same architecture, regularization weight (0.1), and training protocol as JumpLie. The comparison therefore tests whether discovering a symmetry provides practical value beyond simply imposing the known transformation. JumpLie's stronger forecasting results demonstrate an empirical advantage of learned-symmetry regularization in this setting.

We have removed the incorrect finance figure and incorporated the corrected results and preprocessing description into revised Section 4.4 (page 9) and Appendix A.4 (page 15).

Reviewer_Sz8XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes JumpLie, a two-stage framework for discovering and exploiting symmetries of jump-diffusion dynamics. It first learns a differentiable surrogate for the drift, diffusion, and jump components from trajectories. The surrogate is then frozen, and K neural infinitesimal generators are learned by enforcing the continuous SDE symmetry conditions, finite-jump compatibility, non-collapse, and Lie-algebra closure. The resulting symmetries are then used for downstream tasks.

Critical strengths
  1. The core idea is interesting, the paper doesn't propose a way to learn jump diffusion; however it utilizes jump diffusion estimator to learn about possible symmetry transformation to enable downstream tasks like OOD dynamic learning.
  2. Mathematically coherent methodology and theoretical construct.
Critical weaknesses
  1. Slight concern on novelty, as the author has mentioned, similar application was used in another paper, and they have extended into jump diffusion case.
  2. The paper shows that the learned symmetry directions can produce visually plausible edits, but that does not establish that these transformations preserve the quality or distribution of the generator. Since the entire conceptual claim is that these are symmetries of the learned generative dynamics, this is exactly what should be tested quantitatively. For example, FID score before/after transformation? Or even simpler we can do high dimension 8-gaussian?
  3. Although I find the paper generally interesting, most of the empirical examples are more or less proof of concepts.
Questions for authors
  1. How sensitive is symmetry recovery to surrogate estimation error? Can the authors vary data quantity, observation noise, surrogate capacity, or jump observability and measure the effect of aggregation of errors over different stages?
  2. Why must K generally be fixed in advance? Could the method instead use a sufficiently large Kmax⁡K_{\max} and estimate the effective symmetry dimension from the rank or eigenvalue spectrum of the generator Gram matrix?
  3. Related to 2) What happens under deliberate over-specification and under-specification of K? In particular, can the current non-collapse objective force spurious generators when the true algebra has lower dimension? Is it possible to enforce sparsity in the Gram Matrix (or in its eigen values) such that we make K to be dynamically learned?
  4. Can the author just supply the symmetry learner with real jump diffusion to estimate the effect of error in surrogate learning?
5 comments
AuthorsOpenReview ↗

Responses to reviewer Sz8X (1)

We sincerely thank the reviewer for their positive assessment of our work and for their thoughtful and constructive feedback. We particularly appreciate the insightful questions regarding methodological novelty, empirical validation, and robustness, which have helped us further clarify and strengthen our contributions. In the following, W = Weakness, Q = Question.

W1. Novelty relative to LieStoNet

We appreciate the concern and agree that JumpLie builds on the useful two-stage learning strategy introduced by LieStoNet. However, we argue that the extension to jump-diffusion dynamics is not simply applying the same symmetry constraints to a different dataset.

1) A different mathematical symmetry requirement. LieStoNet addresses continuous, Wiener-driven SDEs. In a JD process, satisfying the continuous drift and diffusion determining equations is not sufficient: the transformation must also preserve finite jumps. The symmetry action must be compatible with the jump map $T_J(t,x)=x+J(t,x) $, imposing a nonlocal condition involving the generator at both xx and x+J(t,x)x+J(t,x).

JumpLie builds on established JD symmetry theory to make these additional constraints learnable from a fitted stochastic model. We do not claim to have invented the underlying JD determining equations - our contribution is their data-driven discovery and downstream exploitation.

2) The JD-specific constraint is empirically essential. We directly test this distinction by removing the jump constraints while keeping the frozen JD surrogate and all remaining discovery conditions identical:

Discovery methodRecovery MPA (°) ↓Jump residual ↓
Full JumpLie0.7670.0133
Continuous constraints only88.5890.3727

The continuous-only learner nearly fails to recover the actual JD symmetry space. This shows that finite-jump compatibility changes what can be correctly discovered instead of merely improving an existing diffusion-symmetry objective.

3) New downstream capabilities. Beyond low-dimensional recovery, JumpLie extends discovery to a 3,072-dimensional continuous-flow/categorical-jump generative process. There, discovered directions enable interpretable, training-free image edits while approximately preserving stochastic compatibility and FID (e.g., 25.86 learned-edit FID versus 25.67 unedited and 53.94 matched-random FID at editing strength 0.5). The same symmetry-learning framework also provides useful regularization for real-market volatility forecasting, improving QLIKE over the matched Jump-GRU across SPY, QQQ, and IWM.

The novelty is therefore the learnable enforcement of genuinely JD-specific symmetry structure, validated through a controlled jump-constraint ablation and applied to high-dimensional generation and real financial forecasting. This goes beyond transferring diffusion-only symmetry discovery to another stochastic model.


W2. Preservation of generative quality under learned symmetries

We agree that visually plausible edits alone do not demonstrate preservation of generative quality. We have therefore added quantitative FID and finite-edit compatibility evaluations, comparing all eight learned generators against displacement-matched random controls.

1) Generative-quality preservation. We measure FID relative to the same real-image reference, averaging over eight generators and five random seeds (± standard error). Lower FID indicates better generative quality.

Editing strengthUnedited FID ↓Learned-edit FID ↓Random-edit FID ↓
ε = 0.525.67 ± 0.1125.86 ± 0.3153.94 ± 0.24
ε = 1.025.67 ± 0.1126.04 ± 0.7757.31 ± 0.36

The learned edits increase FID by only 0.74% and 1.44%, respectively, whereas equally large random edits increase FID by approximately 110% and 123%. Thus, the discovered transformations retain nearly the original generative quality even at substantial editing strengths.

2) Distributional compatibility with stochastic generation. We additionally compare the distributions obtained by editing before versus after stochastic continuation. Lower MMD² means better agreement.

Comparison (MMD² ↓ better)ResultInterpretation
Learned edits1.68 × 10⁻⁵Near the sampling reference: approximate finite-action compatibility
Independent-sampling reference1.56 × 10⁻⁵Baseline discrepancy from sampling variability
Matched random edits4.77 × 10⁻⁴28× larger discrepancy than learned edits

These complementary results show that JumpLie discovers transformations that preserve generative quality while approximately respecting the stochastic dynamics, rather than merely producing visually plausible perturbations. The quantitative results are included in revised Section 4.3 (page 8), with complete evaluation protocols in Appendix A.3 (page 13).

(continues)


AuthorsOpenReview ↗

Responses to reviewer Sz8X (2)

(continues)

W3. Empirical validation beyond proof of concept

We appreciate this concern. Beyond the quantitative high-dimensional symmetry validation and generative-quality comparisons reported in our response to W2, we have strengthened the empirical contribution by demonstrating that the discovered symmetries provide interpretable, training-free control over an otherwise opaque stochastic generative model.

1) Interpretable transformations across different images. Revised Figure 5 (page 8) shows the same discovered generator (Generator 1) applied to five distinct generated ships over a range of symmetry pseudo-times ϵ∈[−0.4,0.9]\epsilon \in[-0.4,0.9]. Rather than producing unrelated visual changes, the generator induces a coherent illumination and color transformation: negative pseudo-times produce brighter, warmer scenes, while positive pseudo-times produce progressively darker, bluer scenes. Importantly, this effect is consistent across different starting images, suggesting that the learned direction captures a meaningful and reusable transformation of the generative process.

2) Quantitative evidence of interpretability. We measure the visual changes using darkness and blueness scores averaged across the five images:

Visual attribute (↑ stronger)ϵ\epsilon = −0.4ϵ\epsilon = +0.9Interpretation
Darkness ↑17.8469.12Consistent, substantial darkening
Blueness ↑0.4226.80Systematic increase in blue coloration

These measurements support the visual interpretation: the discovered generator produces a consistent, controllable change in scene appearance, rather than arbitrary image perturbations. In contrast, the displacement-matched random-edit visualization in Appendix A.3 (page 13) does not exhibit the same coherent illumination progression.

3) Practical value beyond qualitative examples. Beyond enabling controllable edits, the discovered symmetry directions also provide interpretability into the otherwise largely opaque stochastic generative process. By revealing coherent transformations, such as illumination and color changes, that approximately preserve the learned dynamics, JumpLie exposes meaningful structure in how the generative model represents and transforms images. Importantly, this interpretability emerges without semantic labels, editing supervision, model retraining, or test-time optimization. As demonstrated in W2, these transformations also preserve nearly the original FID and exhibit substantially better dynamical compatibility than displacement-matched random edits. Thus, symmetry discovery provides both a principled way to understand the structure of a high-dimensional generative process and a practical mechanism for controlling it without sacrificing generative quality.

Together, these results strengthen the high-dimensional experiment beyond proof of concept in two complementary ways: the quantitative evaluations in W2 validate that the learned directions are meaningful approximate symmetries, while the new visual and perceptual evaluations demonstrate a concrete downstream capability - discovering interpretable, quality-preserving controls for stochastic generative processes. The revised Figure 5 (page 8) and Appendix A.3 (page 13) document these findings.


Q2 & Q3. Choosing the number of generators and sensitivity to K

We agree that fixing KK to the known symmetry dimension is undesirable for unknown systems, and that the current non-collapse objective does not automatically suppress excess generators. We investigated both deliberate under-/over-specification and an independent method for estimating the symmetry dimension from the governing JD equations.

1) Detecting the symmetry dimension before neural training. We use a local-jet rank test on the complete JD determining system, including finite-jump and temporal compatibility conditions. Intuitively:

  1. List possible local transformations. At a representative point, treat the unknown generator's values and derivatives as unknown numbers, called local jets.
  2. Impose all compatibility requirements. The drift, diffusion, and finite-jump determining equations constrain those numbers. Differentiating the equations removes apparent solutions that work at one point but cannot extend consistently to a neighborhood. Jump-related evaluations at xx and x+J(t,x)x+J(t,x) are linked.
  3. Count independent compatible choices. A matrix-rank calculation counts how many lower-order generator directions can be extended consistently to higher derivatives. We check that this count stabilizes as compatibility depth, jet order, and numerical-rank settings vary.

(continues below)


AuthorsOpenReview ↗

Responses to reviewer Sz8X (3)

(Q2&Q3 continues)

How is the symmetry dimension computed? We express the full JD determining equations as a system of linear constraints on the unknown symmetry generator and its derivatives. Some constraints involve the generator at the same point, while the finite-jump conditions additionally link its values before and after a jump.

Satisfying the equations at a single point is not enough: a candidate may appear valid locally but fail to extend to a genuine symmetry in a neighborhood. We therefore differentiate the determining equations to introduce additional compatibility constraints, progressively ruling out such invalid candidates.

We then use matrix rank to count how many independent generator directions remain compatible with all these constraints, allowing higher-order derivatives to take whatever values are needed for consistency rather than counting them as additional symmetries. We repeat this calculation at increasing differentiation orders, different base points, and numerical rank thresholds. When the count stabilizes, we take it as the estimated symmetry dimension $K $.

Applied to E1-C, this full-JD procedure consistently detects the correct dimension:

Governing dynamicsDetected dimensionTrue dimension
Exact E1-C coefficients22
Learned E1-C surrogate22

2) What happens if K is misspecified? We independently train neural generators with K=1,2,4K=1,2,4, using the exact E1-C coefficients.

Requested KRecovered rankMissing true symmetry ↓Spurious learned content ↓
1 (too few)150.0%~0%
2 (detected dimension)20.0041%0.0041%
4 (too many)48.21%54.11%

With K=1K=1, the learner finds one valid direction but necessarily misses half the true symmetry space. With K=4K=4, the non-collapse objective encourages four independent fields, producing substantial spurious content. With the independently detected K=2K=2, recovery is essentially complete.

We thus use the results to both confirm the reviewer's concern and provide a principled way to address it: the current NC/BK objective should not be interpreted as automatic dimension selection, but the full-JD determining equations can independently supply the appropriate KK in this tested system.


Q1 & Q4. Sensitivity to surrogate error and discovery with exact dynamics

We thank the reviewer for suggesting a direct exact-dynamics control. We conducted this comparison on E1-C, whose true symmetry algebra has dimension two, and additionally varied training-data quantity and jump observability to examine how errors propagate through the two-stage pipeline.

1) Exact versus learned dynamics. We train the same symmetry learner using either the exact JD coefficients or a neural surrogate estimated from 5,000 observed-jump trajectories. Recovery is evaluated on held-out points against the true dynamics.

Dynamics suppliedRecovery MPA (°) ↓True-dynamics jump MSE ↓Recovered rank
Exact coefficients0.377 ± 0.2222.21 × 10⁻⁵2
Learned surrogate0.837 ± 0.0200.0011452

(Five-seed means; uncertainty shown for MPA)

The exact-coefficient control isolates the symmetry-learning stage from surrogate estimation. Both conditions recover the full two-dimensional symmetry space with under one-degree MPA, showing that surrogate estimation introduces some error but does not substantially disrupt geometric recovery in this benchmark.

2) Training-data quantity and hidden jumps. We next fit surrogates on nested training sets and rediscover the symmetries:

Training trajectoriesJump countsDrift RMSE ↓Recovery MPA (°) ↓True-dynamics jump MSE ↓
500Observed0.30254.5000.00691
1,500Observed0.22393.8350.00317
5,000Observed0.10620.8370.00115
5,000Hidden0.14301.6120.04117

With observed jump counts, increasing the training set tenfold improves recovery MPA from 4.50° to 0.84°, alongside lower surrogate coefficient error and true-dynamics jump residuals. When jump counts are hidden, a zero/one-jump mixture surrogate still permits accurate geometric recovery (1.61° MPA).

These experiments separate the error sources directly: exact-coefficient discovery measures error intrinsic to neural symmetry learning; exact-versus-fitted comparison measures the additional effect of surrogate estimation; and data-size/jump-observability variations reveal how the fitted dynamics influence downstream recovery. Together, these evidences show robustness of JumpLie to the tested variations


Reviewer_uRvVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper incorporates (Lie) Symmetries into a Jump diffusion framework. First, the determining equations are derived, then the algorithm combines them with a non collapse and Lie bracket loss. Usually the JD framework is learned beforehand, using the standard jump diffusion losses, then frozen and then trained to discover symmetries. The algorithm is first benchmarked on some toy examples with known ground truth, to a jump process on cifar10 and then on financial data.

Critical strengths
  1. Jump diffusion frameworks are an underresearched topic, and discovering symmetries of those is interesting.

  2. The results are more theoretical in the sense that more focus is laid onto validating the algorithm than having super shiny examples, but this is good. However in particular the price rescale experiment is very interesting.

Critical weaknesses
  1. I am worried about identifiability. To me it is a priori unclear whether the algo should be able to match the exact symmetries (or even functions). The MPA seems good, but I am wondering if this is to be expected (in the infinite data limit).

  2. I do not understand what section 4.3 is supposed to tell me. What are the discovered symmetries? I do not see the value in training free edits if they are not explained or shown? what is a pseudotime?

Questions for authors

see weaknesses

4 comments
AuthorsOpenReview ↗

Responses to reviewer uRvV (1)

We sincerely thank the reviewer for their thoughtful and constructive feedback and for recognizing the strengths of our work. We particularly appreciate the insightful questions regarding symmetry identifiability, theoretical guarantees, and the interpretation and practical utility of discovered symmetries, which provide an opportunity to further clarify our contributions. In the following, W = Weakness.

W1. Identifiability and recovery in the infinite-data limit

We thank the reviewer for raising this important distinction. The natural identifiable object is the symmetry space (Lie algebra), not a unique set of generator functions.

1) Why individual generators need not match exactly. If H1,…,HKH_1,\ldots,H_K form a basis of a symmetry algebra, any invertible linear combination of these generators spans the same symmetry space. Hence, even with exact dynamics and perfect optimization, a learned generator need not equal a particular analytic generator pointwise. This is why we evaluate the maximum principal angle (MPA) between the learned and analytic generator subspaces, rather than direct parameter or function-wise errors.

2) What should happen with infinite data? There are two separate requirements:

  • Identification of the dynamics. The observed trajectories must determine the relevant JD drift, diffusion law, and jump mechanism sufficiently accurately. This depends on observability, model specification, and available sampling information (hidden jumps can make estimation harder).
  • Identification of the symmetry space. Given the correct JD dynamics, the full determining equations must characterize the intended symmetry class, and optimization must recover the complete, nontrivial solution space rather than only a subspace or redundant generators.

Under suitable coefficient-identifiability, surrogate-consistency, regularity, and optimization assumptions, we expect the recovered symmetry space (Lie algebra) to converge to the true one as data and optimization accuracy improve.

3) Direct empirical evidence. We additionally separate symmetry-learning error from surrogate-estimation error on the analytic E1-C system:

EvaluationRecovery MPA (°) ↓Recovered dimension
Exact JD coefficients0.377 ± 0.2222
Learned JD coefficients (5,000 trajectories)0.837 ± 0.0202

As shown, both recovered MPAs are close to zero, and an independent rank test of the full JD determining equations detects dimension 2 (detection mechanism specified in Q2&Q3 response to reviewer Sz8X) for both exact coefficients and the learned surrogate. Thus, the estimated dynamics preserve the correct symmetry-space structure in this benchmark, and the neural learner accurately recovers it.

The key conclusion is that exact recovery of a particular generator basis is neither expected nor required - recovery of the underlying symmetry space is the appropriate objective. Our under one-degree MPAs and independent dimension-detection results provide strong finite-sample evidence for this objective.


W2. Interpretation, pseudo-time, and practical value of discovered symmetries

We thank the reviewer for requesting a clearer interpretation of the discovered transformations. We have replaced the original limited visualization with a five-image symmetry-flow visualization in revised Figure 5 (page 8), which demonstrates how the same learned generator consistently transforms different images as pseudo-time varies.

1) What symmetry is discovered? As shown in Figure 5 (page 8), generator 1 learns an interpretable illumination/color transformation: negative pseudo-times produce brighter, warmer ship scenes, whereas positive pseudo-times produce darker, bluer scenes. Crucially, this direction behaves consistently across five distinct starting ships.

We quantify the transformation using two image-level color measurements averaged across the five images: darkness, reflecting lower overall pixel brightness, and blueness, reflecting the prominence of the blue color channel relative to red and green.

Visual attribute (↑ stronger)$\epsilon $ = −0.4$\epsilon $ = +0.9Interpretation
Darkness ↑17.8469.12Images become substantially darker
Blueness ↑0.4226.80Images acquire a stronger blue atmosphere

(continues below)

AuthorsOpenReview ↗

Responses to reviewer uRvV (2)

(W2 continues)

2) Meaning of pseudo-time. Pseudo-time ϵ\epsilon controls how far and in which direction an image moves along the discovered symmetry transformation. It is distinct from the generative time tt.

We pause generation at t=0.3t = 0.3, integrate the learned vector field according to dxε/dε=ξa(t,xε)dx_\varepsilon/d\varepsilon=\xi_a(t,x_\varepsilon) while holding tt fixed, and then resume the original frozen generative process. Positive and negative ϵ\epsilon move in opposite directions along the learned symmetry orbit. No model retraining or test-time optimization is required in this process

3) Why are these transformations useful? We verify that the interpretable edits also preserve generative quality and approximately commute with stochastic generation.

Evaluation (↓ better)Learned editsMatched random editsInterpretation
FID at ε = 0.525.8653.94Learned edits retain nearly the original image quality
FID at ε = 1.026.0457.31Quality preservation persists under stronger edits
Finite-edit MMD²1.68 × 10⁻⁵4.77 × 10⁻⁴28× better compatibility with stochastic continuation

The unedited FID is 25.67, and the finite-edit independent-sampling MMD² reference is 1.56×10−51.56 × 10^{-5}. Thus, the learned edits remain close to both the original generative-quality baseline and the sampling-compatibility reference.

These findings show that JumpLie learns a coherent, interpretable editing direction that provides training-free control while approximately preserving the stochastic generative dynamics and image quality. Revised Figure 5 (page 8) makes the transformation visually explicit, while Appendix A.3 (page 13) provides the full quantitative results and random-edit controls.

Reviewer_uRvVOpenReview ↗

response

I thank the authors for their rebuttal and will keep my weak accept. i think it is a good work.

AuthorsOpenReview ↗

Response to Reviewer uRvV

We sincerely thank the reviewer for their encouraging response and positive assessment of our work. We greatly appreciate their constructive suggestions, which have helped us clarify the theoretical aspects and better demonstrate the interpretability and practical value of the discovered symmetries.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.