acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Induced Operator Families for Deep Representation Learning

Abstract

The geometry of latent representations governs which components should interact and how information should propagate, making geometry-aware operator design a fundamental ingredient of structured deep representation learning. However, existing neural architectures typically rely on generic operator templates or geometry-specific constructions, creating a need for a unified framework that can derive admissible operators directly from fixed structural information while remaining adaptive to changing contexts. We introduce Geometry-Induced Operator Families (GIOF), a general framework that converts fixed geometry into a structured family of propagation operators and dynamically selects an appropriate member of this family according to the current context. GIOF first transforms geometry-derived interaction channels into reusable generator bases, then combines them through a context-dependent selector and adaptive propagation scale, and finally realizes the selected operator through stable continuous-time propagation and a bottleneck residual layer. We establish theoretical guarantees covering parameter compression, identifiability, stability, locality, compositional structure, and oversmoothing behavior, while controlled experiments validate these mechanisms and experiments on PEMS-BAY and METR-LA achieve the lowest mean MAE across all reported regional-outage settings, improving over the strongest retained baseline by 2.4%–8.8% at 30% missing sensors.

Primary area:
interpretability and explainable AI
Keywords:
Geometry-Induced Operators, Latent Geometry, Deep Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1LQ4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Geometry-Induced Operator Families (GIOF), which aims to provide a general framework for designing geometry-aware neural-network layers for structured representation learning. The method constructs interaction operators from the fixed relationships between sites and learns how to combine them for each input. The experiments focus on synthetic operator checks and traffic-sensor reconstruction.

Critical strengths

None

Critical weaknesses

W1. The biggest issue is that the experiment in the work is very narrowed and doesn’t sufficiently support the paper’s broad claims. This work aims to provide a general framework for designing geometry-aware neural-network layers for structured representation learning, but experiment focuses (i) small-scale synthetic experiments, (ii) two traffic-reconstruction datasets. The synthetic experiments are useful sanity checks, but the real-data evaluation is kind-of weak and doesn’t provide enough evidence for the claimed broader practical applicability

W2. Another major flaw is that the work has terrible readability. The paper is extremely long (>150 pages), the main part contains many definitions, propositions and theorems, it’s hard to get an intuitive understanding of the proposed method. Several appendix proofs also spend excessive space on elementary consequences of the definitions. For example, P.3.2 proof (page 49) follows the fact that every interaction channel is zero outside the prescribed mask, yet its proof explains the multiplication and summation of these zero entries at length (very trivial results) . Theorem 6.1 re-establishes properties already supplied by Proposition C.4. I have not checked every part of the appendix, so additional issues may remain beyond those mentioned above. I suggest that the authors substantially revise and streamline the paper to increase the readability.

Questions for authors

Q1. . The motivation for this work remains kind-of abstract and unclear to me after careful reading. The main text provides very limited intuition for the proposed method, and for the appendix part, the "position of GIOF" paragraph on page 12 only describes how the framework organizes previously studied ingredients, what the new insight and capability does this work provide?

Q2. Why do METR-LA MAE baselines perform worse with 100% of the training days than with 20%? (Table 1)

3 comments

Response to the motivation and the capability added by GIOF.

The missing intuition is not simply that "geometry is useful." It is that admissible geometry and the realized propagation geometry are different objects.

Suppose a fixed sensor support admits several rate patterns, such as proximity-dominated transport and degree-adjusted transport.

A fixed-generator diffusion can propagate for a shorter or longer time, but cannot change the ratio of its infinitesimal rates along these patterns.

GIOF first compiles each geometric pattern into an admissible generator B^r\widehat B_r, then chooses

\qquad P(c)=e^{\tau(c)L(c)}$$ for the current input. The mixture changes the *shape* of admissible propagation, and $\tau(c)$ changes its *scale*. All selected operators retain structural guarantees by construction. This is a precise new modeling capability relative to a single fixed generator with learned diffusion time, not a claim that every geometry-aware architecture lacks adaptive filters. **Manuscript revision.** The Introduction now gives the shape-versus-scale motivation before the theory; the layer section connects it directly to the trainable residual realization. ### Response to whether the experiments substantiate the framework. We agree that empirical breadth and the generality of the mathematical construction should be distinguished. The descriptor-to-family map is defined for directed admissibility, barriers, and multiscale channels; this is a statement about the construction, not a claim of measured cross-domain superiority. The trained benchmark is intentionally controlled: five published spatial modules and several geometry/parameterization ablations replace the spatial component of the same network on PEMS-BAY and METR-LA. The same checkpoints are probed under three missingness mechanisms, extrapolated regional severities, outage durations, and graph-distance strata. For example, on PEMS-BAY the regional MAE decreases from $3.483$ (Static) to $3.070$ (GIOF), and the point-missingness MAE decreases from $1.307$ to $1.154$. These results test concrete consequences of adaptive structured propagation, rather than just checking algebraic definitions. To strengthen the empirical evidence for general applicability, we additionally evaluate reconstruction on a non-traffic spatial grid and examine directed and barrier-constrained geometric descriptors. These experiments complement the original traffic benchmarks by testing GIOF under different interaction structures, while preserving controlled comparisons against the corresponding static operator family. **Manuscript revision.** We explicitly state the boundary between the general mathematical construction and evidence from the controlled reconstruction setting, while moving the most informative existing ablations into the main narrative. ### Response to length, readability, and elementary proofs. The method can be stated as a three-stage pipeline: $$\text{Fixed descriptor} \longrightarrow \text{Admissible generator dictionary} \longrightarrow \text{Context-selected propagation}.$$ A long proof appendix is not itself a distinct contribution, and the reviewer identifies an unnecessarily expanded argument. In particular, Proposition 3.2 follows immediately because every off-diagonal channel entry is masked by the admissibility relation; its proof can be reduced to one line of algebra. Likewise, the projection result is a standard consequence of closed convexity and should be presented as a supporting interpretation, not a new projection theorem. These observations identify concrete opportunities to improve readability and to distinguish supporting mathematical results from the main contribution. **Manuscript revision.** We keep the geometry-to-operator pipeline in the ten-page main text, explicitly distinguish supporting lemmas from contribution-level statements, and shorten the cited support proof and redundant projection derivation in the appendix.

Response to the nonmonotone METR-LA data-budget comparison.

The same 1,600 optimizer updates are used for 20% and 100% training days; the two settings do not use the same number of epochs or visits per training window.

Hyperparameters are chosen on the 20%-day subset and then held fixed, and normalization statistics are fitted on the respective selected training-day sets.

Most importantly, checkpoint selection averages nine mechanism--severity validation conditions, whereas Table 1 reports only the 30% regional condition.

Therefore, the regional MAE is not constrained to be monotone as more training days are made available.

This explains why the reversal is logically possible, but it does not determine whether training exposure, normalization, objective trade-offs, or incomplete convergence is dominant.

That requires trajectory and longer-budget diagnostics.

The extended-budget results show that all six methods improve on METR-LA with 100% training days, but they do not eliminate the data-budget reversal. For example, at the 3,200-update cap, ECC obtains regional MAEs of 4.8354.835 with 20% training days and 5.0285.028 with 100% training days.

Thus, the reversal cannot be attributed solely to the original 1,600-update cap on the basis of these results. Differences in training exposure, normalization statistics, and checkpoint selection remain possible explanations, and we do not claim to have isolated a single cause.

We report the regional-outage MAE at both budgets and the selected validation checkpoint for each method. This comparison tests whether the nonmonotone data-budget behavior can be attributed to insufficient optimization, rather than assuming that additional training days must improve every individual evaluation condition.

Manuscript revision. We have clarified the comparison's fixed-update and aggregate-validation interpretation, without claiming that the current results are convergence-matched.

Dear Reviewer 1LQ4,

Thank you for your feedback. We have carefully addressed your concerns regarding motivation, experimental generality, readability, and the METR-LA results.

We would appreciate it if you could review these revisions and let us know whether your principal concerns have been addressed, or if any further clarification would be helpful.

Thank you for your consideration.

Reviewer_uUbrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Geometry-Induced Operator Families (GIOF): a fixed graph with edge attributes defines a family of Metzler, constant-preserving generators (row sums zero), and a context-conditioned selector chooses a mixture and a propagation scale τ\tau within it, applying P=eτLP = e^{\tau L} as a residual bottleneck layer. The authors prove stability, locality, compression, identifiability and oversmoothing properties of this family for fixed context. Experiments on PEMS-BAY and METR-LA sensor reconstruction under regional outages show 2.4–8.8% lower MAE than GCN/GATv2/APPNP/MoNet/ECC modules inserted into the same network.

Critical strengths
  • I like the core idea. Deriving the admissible operator family from a fixed geometry descriptor and letting context pick inside it, instead of learning a free d×dd \times d matrix, is a clean way to get locality, directionality, barriers and multiscale structure out of one layer (App. F shows the standard templates are special cases). This is the interesting part of the paper for me. Whether the operator is actually useful and generalizes beyond the one setting tested is my main concern, see Weaknesses.
  • The experimental protocol is careful: shared encoder/decoder/bottleneck, splits, masks and update budget across all baselines; paired per-seed results over ten mask realizations; measured cost; Fig. 2 tests a missingness rate outside the training range. Within the comparison that was run. My concern is not that the experiments are wrong but that they are not aligned with the theory, and that the theory itself is limited to a fixed context, see Weaknesses.
Critical weaknesses
  • The theorems and the experiments are about different objects. All guarantees in Sec. 4 and 6 assume a fixed context cc (fixed LL and τ\tau); the trained model is self-conditioned, and the authors admit several times that the guarantees do not carry over. So the paper proves that the heat kernel on a fixed graph is stochastic, local and contracting, then evaluates a network to which none of this applies. The scope is narrowed repeatedly until the headline claims no longer cover the reported results.
  • I don’t see what is new relative to GRAND / APPNP. With the sparse implementation (H.4) the layer is a truncated heat kernel on a mixture of fixed Laplacians with learned τ\tau, i.e. GRAND with fixed diffusivity or APPNP with a learned step count. App. A calls GIOF “complementary” to these but never says what the difference is. Please add GRAND and APPNP-with-learned-τ\tau as baselines and state the difference explicitly.
  • Generality is claimed, not shown. One descriptor type, one task, two datasets; none of the directed/barrier/multiscale descriptors of App. F is tested. Table 4 attributes most of the gain to fixed diffusion (Static 3.48 → GIOF 3.07), not to the selector. No imputation baselines (GRIN, SPIN, PriSTI); GATv2 ± 0.80 suggests the baselines are not well tuned.
  • Much of the theory is textbook. Thm 4.2/4.5 are uniformization of a CTMC, Thm 4.3/Prop 4.4 are heat-kernel contraction at the spectral gap, Prop 4.1/Thm 4.8 are dimension counting, Thm 6.1 is projection onto a convex set. Correct, but not informative.
Questions for authors
  • Do any of the Sec. 4 guarantees hold for the self-conditioned model you train? If not, what is guaranteed for it?
  • How is GIOF different from GRAND or APPNP with learned τ and a kernel mixture? Can you compare against them?
  • In Table 4, how much of Static → GIOF (0.41) comes from the selector alone?
  • Why no GRIN / SPIN / PriSTI on PEMS-BAY / METR-LA?
  • GATv2 std ± 0.80; is it tuned? Do the conclusions hold if it is?
  • Any results with a descriptor other than symmetrized distance kernels?
6 comments

Response to whether the theory applies to the trained self-conditioned model.

We appreciate the reviewer's concern about whether the fixed-context guarantees extend to the trained self-conditioned model. An important distinction is between the structural properties of the operator selected for each input and the stability properties of the full nonlinear input--output map.

For every representation HH, the trained selector yields PH=eτ(c(H))L(c(H))P_H=e^{\tau(c(H))L(c(H))} with L(c(H))L(c(H)) in the fixed admissible generator family.

Therefore, for every input,

$ P_H\ge0,$ $ P_H\mathbf1=\mathbf1,$

∥PH∥∞→∞=1,\|P_H\|_{\infty\to\infty}=1,

supp(PH)supp(P_H) $\subseteq\mathcal R_{\mathcal E}. $

Each realized propagation is a nonnegative, constant-preserving averaging operator, and products of such realized operators preserve these invariants without requiring the selected generators to commute.

These statements hold for the propagation matrices actually used by the trained model; they are not restricted to one globally fixed cc.

The stronger assertion that the whole self-conditioned network is globally nonexpansive would indeed be unjustified, and is not our claim.

For the canonical map g(h)=eQ(h)hg(h)=e^{Q(h)}h,

Dg(h)[v]=eQ(h)v+D(eQ(h))[v]h.Dg(h)[v] = e^{Q(h)}v + D(e^{Q(h)})[v]h.

The second term measures dependence through the selector. The existing self-conditioning proposition bounds ∥g(h)−g(h~)∥∞\|g(h)-g(\widetilde h)\|_\infty

by

(1+MKQ)∥h−h~∥∞(1+MK_Q)\|h-\widetilde h\|_\infty

when ∥h∥∞≤M\|h\|_\infty\le M and QQ is KQK_Q-Lipschitz.

This is a conditional bound for the canonical propagation map, not a certified Lipschitz constant for the complete residual encoder--decoder network.

Moreover, its globally pooled context does not automatically satisfy functional ancestor-locality.

Thus the pointwise guarantees remain applicable to the propagation operators used in the trained model, whereas the distinction between pointwise invariants and full-network Jacobian bounds is essential.

Manuscript revision. We have promoted this distinction to the main layer section, alongside the actual residual-layer formula and the appendix conditions.

Response to equivalence with GRAND, APPNP, and learned propagation time.

Using a heat-kernel implementation does not establish equality of operator families.

The following three-site counterexample is exact, not an additional trained experiment.

Let the allowed rates be

\begin{pmatrix} -3 & 3 & 0 \\\\ 0 & 0 & 0 \\\\ 0 & 0 & 0 \end{pmatrix}, \qquad \widehat B_2= \begin{pmatrix} -3 & 0 & 3 \\\\ 0 & 0 & 0 \\\\ 0 & 0 & 0 \end{pmatrix}.$$ with $$L_\alpha = \alpha\widehat B_1 + (1-\alpha)\widehat B_2.$$ Both bases are rate-normalized, admissible generators on a single fixed descriptor. For every $\tau>0$, $$\frac{(e^{\tau L_\alpha})_{12}}{(e^{\tau L_\alpha})_{13}}$$ $$= \frac{\alpha}{1-\alpha}.$$ | Selected mixture $\alpha$ | $P_{12}/P_{13}$ | Time-only changes ratio? | |---|---|---| | $1/4$ | $1/3$ | No | | $3/4$ | $3$ | No | Learning only the time of any *fixed* generator cannot switch between these two ratios. This example distinguishes GIOF from learned-time propagation with a fixed generator. A sufficiently flexible, context-dependent kernel mixture may represent similar operators; GIOF's distinctive specification is the descriptor-compiled admissible family together with pointwise structural guarantees. GRAND may support more general state-dependent dynamics. GIOF instead makes a fixed descriptor-derived generator dictionary, constrained adaptive selection, and per-selection structural certification the modeling primitives. **Manuscript revision.** The main text now states the shape-versus-time distinction and explicitly acknowledges the scope of the comparison. The matched shape-only and time-only controls provide an empirical comparison that complements the operator-family distinction established by the counterexample.

Response to the Static ablation and the claim that only fixed diffusion helps.

We would like to clarify the numerical interpretation of the regional ablation.

The reported PEMS-BAY results (20% training days, 30% missingness) are:

We conduct a matched factorial ablation to separate adaptive generator selection from adaptive propagation time. We additionally include GPR-GNN, which learns multi-hop propagation coefficients over a fixed graph.

MethodPEMS-BAYMETR-LA
APPNP3.2914.966
GPR-GNN3.2384.911
No-spatial3.6965.312
Static3.4835.078
Shape-only3.2174.758
Time-only3.3464.867
GIOF3.0704.668

Regional MAE (↓\downarrow), 20% training days, 30% regional missingness. Shape-only adapts α(c)\alpha(c) with constant τ\tau; Time-only adapts τ(c)\tau(c) with constant α\alpha.

On regional outages, Static provides only 0.2130.213 of the No-spatial-to-GIOF improvement, whereas joint context-dependent shape and time selection provides 0.4130.413 (about 66%66\% of the total).

Notably, Static worsens point-missingness MAE relative to No-spatial (1.3071.307 versus 1.2801.280), while full GIOF improves it to 1.1541.154.

These observations indicate that fixed diffusion alone does not explain the full improvement in the reported setting.

The factorial controls allow us to separate the contributions of context-dependent generator selection and propagation time. Importantly, the comparison with GPR-GNN also tests whether the observed improvement can be explained by learned multi-hop propagation alone.

Manuscript revision. We have included the matched Shape-only and Time-only controls in the main experimental discussion, explicitly distinguishing adaptive generator selection from adaptive propagation time.

Response to baseline choice, tuning, and the scope of generality.

The primary empirical question is which spatial module works better when everything else is held fixed, not which complete imputation system is state of the art.

GRIN, SPIN, and PriSTI are task-specific imputation architectures; comparing their native pipelines directly to a module-substitution experiment changes temporal modeling, capacity, and training objectives.

Accordingly, our results support a controlled comparison of spatial modules, while direct comparisons with GRIN, SPIN, and PriSTI would provide complementary evidence about application-level performance.

To examine the applicability of GIOF beyond the original traffic setting, we consider additional geometry and descriptor configurations.

SettingGeometryStaticGIOF
PEMS-BAYDistance/degree3.4833.070
Weather gridRegular grid0.4320.401
Directed gridDirected edges0.2780.258
Barrier gridBlocked edges0.3390.318

Reconstruction MAE (↓\downarrow). The settings have different signal distributions and numerical scales; comparisons are made within each row, not across datasets.

These controls evaluate whether the benefits of descriptor-induced propagation persist under different admissibility structures, rather than relying exclusively on symmetric traffic graphs.

The quoted GATv2 0.8010.801 standard deviation occurs on PEMS-BAY with 20% days; it is 0.0600.060 with 100% days. GATv2 received the same six-trial pilot search budget as the other spatial modules, with learning rate and depth selected using the shared validation protocol. This provides a matched tuning procedure, although it does not establish that every baseline has fully converged.

More importantly, the strongest PEMS-BAY baseline mean is ECC in both budgets, so GIOF's gain there does not rely on the unstable GATv2 run.

Still, a fixed 1,600-update cap cannot rule out further improvements from longer baseline training, and the reported module-level results do not establish superiority over complete imputation systems or all possible diffusion architectures.

Manuscript revision. We identify our real-data results as controlled module-level evidence, explicitly separate mathematical descriptor generality from empirical cross-domain validation, and clarify the shared selection protocol without claiming unmeasured convergence.

Response to the claim that the theoretical contribution is textbook.

Uniformization, Markov-semigroup contraction, and convex projection are established mathematical tools; we do not claim otherwise.

The contribution is the constructed hypothesis class to which they are applied: a fixed descriptor induces shared interaction channels and a constrained generator dictionary, while a learned selector remains inside that class.

This yields a joint account of dimension reduction relative to edgewise parameters, conditional coefficient identifiability, pathwise reachability and finite-scale influence, and closure of structural invariants under compositions of noncommuting selected propagators.

These are not all novel standalone inequalities, but they characterize different aspects of one explicit architecture rather than an unrelated heat kernel.

The presentation should not elevate elementary consequences to independent theoretical innovations.

Manuscript revision. We prioritize the descriptor-to-family construction and its genuinely discriminating consequences, label projection and uniformization as supporting tools, and compress elementary proofs.

Reviewer_zSBcOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how fixed geometric structure in a latent space (locality, directionality, obstruction, etc.) should determine the operators that propagate information between components of a representation. The authors propose Geometry-Induced Operator Families (GIOF), which convert a fixed geometry descriptor into reusable generator bases, combine them through a context-dependent selector and an adaptive propagation scale, and realize the selected operator through continuous-time (matrix-exponential) propagation followed by a bottleneck residual layer. The paper provides theoretical results on parameter compression, identifiability, stability, locality, composition, and overs-moothing, supported by a very long appendix (154 pages in total). Controlled synthetic experiments probe these mechanisms, and on PEMS-BAY and METR-LA regional-outage reconstruction, GIOF used as a drop-in spatial module achieves the lowest mean MAE against GCN, GATv2, APPNP, MoNet, and ECC, improving over the strongest baseline by 2.4%–8.8% at 30% missing sensors.

Critical strengths
  1. The separation between fixed geometry (what may interact), admissible generators, and context-dependent selection is clean and gives a useful organizing view of geometry-aware operator design.
  2. The theoretical results are carefully qualified. The authors explicitly distinguish algebraic identifiability from recovery, discuss exponential aliasing and noncommuting composition, and provide counterexamples where guarantees do not hold, which I appreciate.
  3. The controlled experiments include failure regimes (singular probes, geometry mismatch, excessive smoothing) and not only favorable demonstrations.
  4. The real-data comparison is well controlled: all baselines are inserted into the same encoder-bottleneck-decoder network with shared masks, splits, and update budgets, and the ablations (Static, Shuffled, One-step, GeoMLP, FreeEdge) go well beyond a simple no-spatial ablation.
  5. The authors are upfront about the computational cost (GIOF is slower than GCN/GATv2/APPNP and comparable to MoNet) and frame the result as an accuracy-structure trade-off rather than a universal gain.
Critical weaknesses
  1. The paper is extremely long (~140 pages of appendix), and the main text reads largely as a summary, making it hard to see what is new. Many ingredients are well known, e.g., diffusion operators, geometry-conditioned filters (MoNet, ECC), continuous-depth propagation (GRAND [1]), and learned diffusion time (DiffusionNet [2]), so the contribution lies mainly in their combination. Closely related work such as DiffusionNet [2], GDC [3], and GPR-GNN [4] is not discussed.
  2. My main concern is the limited evaluation: one task on two traffic datasets with three seeds, and no learned multi-hop baseline (APPNP uses fixed propagation weights; the others are two- to four-layer local modules). Can author(s) add a learned-diffusion baseline (e.g., GRAND-l [1] or GPR-GNN [4]) and at least one non-traffic geometry?
  3. All methods share a fixed 1,600-update budget with configurations chosen on 20% of the training days. On METR-LA, every baseline gets worse with 100% of the days while GIOF improves, suggesting under-tuned baselines; selected hyperparameters, convergence curves, and paired per-seed differences are not reported.
  4. The Static ablation removes both the adaptive mixture and the adaptive propagation time, so it is unclear which one drives the gain.
  5. Minor: the sup-norm contraction claimed on p. 2 does not extend to reverse-mode gradients, which pass through P⊤P^\top, whose ∞\infty-norm can exceed one unless PP is balanced.

[1] B. P. Chamberlain, J. Rowbottom, M. Gorinova, S. Webb, E. Rossi, and M. M. Bronstein. GRAND: Graph Neural Diffusion. ICML, 2021.

[2] N. Sharp, S. Attaiki, K. Crane, and M. Ovsjanikov. DiffusionNet: Discretization Agnostic Learning on Surfaces. ACM Transactions on Graphics, 2022.

[3] J. Klicpera, S. Weißenberger, and S. Günnemann. Diffusion Improves Graph Learning. NeurIPS, 2019.

[4] E. Chien, J. Peng, P. Li, and O. Milenkovic. Adaptive Universal Generalized PageRank Graph Neural Network. ICLR, 2021.

Questions for authors
  1. Which hyperparameters were selected for each method and dataset, and were the baselines still improving at the 1,600-update cap? Does the METR-LA 100% ranking hold under a longer training budget?
  2. Can the authors report shape-only and time-only adaptation ablations, and show how the selected mixture and propagation time vary across outage mechanisms and severities?
  3. How do training and inference costs scale with the number of nodes and generator channels, e.g., for graphs one or two orders of magnitude larger than PEMS-BAY?
  4. What is the realized rank of the generator dictionary on the real datasets, and is the coefficient interpretation meaningful there?
  5. Can the authors report paired per-seed differences or confidence intervals for the Table 1 improvements?

Any discussion around these will help improve the paper over its current state.

4 comments

Response to novelty and the relation to learned diffusion.

We agree that diffusion and continuous-depth propagation are established techniques. The key distinction lies in the operator family that GIOF constructs and the way context selects within this family. A fixed-generator diffusion with learned time has the form exp⁡(τ(c)L0)\exp(\tau(c)L_0): time rescales all infinitesimal transition rates together.

GIOF constructs admissible, rate-normalized bases B^r\widehat B_r from a fixed geometry descriptor and selects

\qquad P(c)=e^{\tau(c)L(c)}.$$ Consequently, changing $\alpha(c)$ changes relative transition rates while respecting the same descriptor-imposed admissibility constraints; changing $\tau(c)$ determines the amount of propagation. These are separate degrees of freedom whenever the compiled bases contain genuinely different generator directions. APPNP and GPR-GNN propagate using polynomials of prescribed graph operators; DiffusionNet adapts diffusion scales for a prescribed surface Laplacian; GDC constructs a diffused propagation graph. GRAND is broader and can itself incorporate state-dependent dynamics: we do *not* claim that GIOF is universally more expressive than GRAND. The distinct contribution is a reusable descriptor-to-generator construction whose structural properties are certified for *every selected operator*, not the invention of matrix exponentiation. **Manuscript revision.** The Introduction and related-work appendix now foreground the shape-versus-scale distinction and explicitly position GIOF against APPNP, GRAND, GPR-GNN, GDC, and DiffusionNet. ### Response to evaluation scope and what the ablations actually show. The real-data tests are not restricted to a single regional-outage mechanism. The same trained models are evaluated under regional outages, independent sensor outages, independent point missingness, varying outage duration, graph distance from observed sensors, and regional missingness up to 60% (outside the training-rate range). These evaluations provide broader evidence within the reconstruction task. The additional grid-based experiments reported below extend the geometric settings, although they do not establish generality across unrelated tasks. The regional ablation results suggest a more nuanced interpretation of the role of fixed diffusion. On PEMS-BAY (20% training days; 30% regional missingness), No-spatial, Static, and GIOF yield MAEs $3.696$, $3.483$, and $3.070$. Static contributes $0.213$ relative to No-spatial, while replacing the constant mixture and scale by context-conditioned selection contributes an additional $0.413$, or $66\%$ of the total $0.626$ reduction along this ablation path. GeoMLP ($3.390$) and FreeEdge ($3.384$) retain the same support and exponential-action mechanism while relaxing fixed-channel sharing; neither matches GIOF. These results show that fixed diffusion alone does not account for the full regional improvement. To disentangle the two forms of adaptation, we additionally evaluate Shape-only and Time-only controls. On PEMS-BAY, Shape-only and Time-only achieve MAEs of $3.217$ and $3.346$, respectively, compared with $3.483$ for Static and $3.070$ for full GIOF. On METR-LA, the corresponding values are $4.758$, $4.867$, $5.078$, and $4.668$. Both adaptive components contribute to the observed improvements, with their joint effect evaluated by the full GIOF model. Since the variants are trained separately, these differences should not be interpreted as an additive causal decomposition. **Manuscript revision.** We have made this measured decomposition and its precise scope explicit in the main experimental discussion. The revised experiments include matched shape-only and time-only controls, together with a learned multi-hop GPR-GNN baseline. These comparisons directly address whether adaptive generator selection contributes beyond learning the propagation scale or multi-hop coefficients.

Response to optimization fairness, the METR-LA reversal, and statistical uncertainty.

All spatial modules share the encoder, decoder, masks, chronological splits, per-seed initialization of shared weights, and 1,600 final optimizer updates.

The selection procedure is not unspecified: For clarity, each method and dataset uses six 500-update validation pilot configurations on the 20%-day subset. The selection is based on a common validation objective averaging three masking mechanisms at three severities. The selected configuration is then held fixed across the two training-day budgets.

We now report the validation-selected hyperparameters for each of the original five spatial baselines and GIOF on both datasets, in addition to the common six-trial search protocol. The selected configurations are held fixed across the two training-day budgets.

The extended-budget comparison provides an additional optimization diagnostic. Nevertheless, matched update budgets and longer runs do not by themselves establish that all methods have converged.

The METR-LA 100%-day behavior therefore cannot establish poor implementation or under-tuning by itself: both the number of passes per example and training-only normalization statistics change with the day budget, and the aggregate checkpoint criterion differs from the single regional test condition.

Equally, these mechanisms explain why monotonicity is not guaranteed; they do not identify the actual cause of the reversal. The appropriate checks are recorded validation trajectories, a longer-update comparison, and the selected hyperparameters.

The ten test masks are averaged within each of three model seeds; they are not 30 independent training replicates. We therefore report paired seedwise MAE differences across the three matched model seeds, rather than inferring uncertainty from the marginal standard deviations in Table 1.

DataComparatorBaseline (1,600 updates)GIOF (1,600 updates)Baseline/GIOF (3,200 updates)Paired Δ\Delta (Mean ±\pm SD)
PEMS 20%ECC3.1983.0703.155/3.0380.128±0.0220.128\pm0.022
PEMS 100%ECC3.1473.0723.111/3.0310.075±0.0320.075\pm0.032
METR 20%ECC4.8974.6684.835/4.6060.229±0.0510.229\pm0.051
METR 100%GATv25.0444.5994.956/4.5510.445±0.0230.445\pm0.023

Regional MAE (↓\downarrow) at 30% missingness. Comparators are the strongest baselines at 1,600 updates. Paired differences use baseline MAE minus GIOF MAE, computed across the same three model seeds.

Manuscript revision. We have added the selected hyperparameter configurations, extended-budget comparisons, and paired seedwise differences. We also clarify that the primary comparison uses a fixed optimization budget rather than convergence-matched training.

Response to graph scaling, realized dictionary rank, and reverse-mode stability.

The propagation core uses a sparse Poisson/uniformization action, not a dense matrix exponential. With m=d+∣Eoff∣m=d+|\mathcal E_{\rm off}|, RR channels, truncation order KK, and bottleneck width CC, its per-window cost is

O(Rm+KmC),O(Rm+KmC),

with O(Rm)O(Rm) fixed channel storage and O(KdC)O(KdC) uncheckpointed recurrence activations.

For bounded-degree graphs this is linear in dd when R,K,CR,K,C are fixed. It is an analytical statement, not a measured claim for graphs 10--100 times larger.

For the eight-channel dictionaries, the measured numerical rank is eight on both PEMS-BAY and METR-LA, under the stated singular-value threshold.

We additionally benchmark graph-size scaling using tiled PEMS-BAY supports. Inference latency increases from 2.312.31 ms at 325 nodes to 10.8010.80 ms at 3,250 nodes and 99.4099.40 ms at 32,500 nodes.

These experiments provide controlled evidence about computational scaling. Since graph tiling does not create a larger connected road network, we do not interpret these measurements as evidence of predictive generalization to larger real graphs.

Finally, the point about reverse-mode gradients is well taken and motivates a distinction between forward and dual-norm stability. Row stochasticity implies

∥P∥∞→∞=∥P⊤∥1→1=1.\|P\|_{\infty\to\infty} = \|P^\top\|_{1\to1}=1.

The direct reverse-mode map at fixed context is nonexpansive in the dual ℓ1\ell_1 norm, not necessarily ℓ∞\ell_\infty; selector derivatives and the residual bottleneck require separate treatment.

Manuscript revision. The Introduction now states the correct dual-norm stability distinction. We also report realized dictionary ranks and controlled graph-scaling measurements, while distinguishing computational scaling from predictive generalization.

Dear Authors,

Thank you for the response. I am not very satisfied with the responses, especially regarding the core contribution being a combination, which the authors have further accepted. Further, the additional results still don't help answer point 2 above. Hence, I am lowering my score.

Response to the follow-up on methodological novelty and empirical evaluation.

We thank the reviewer for the follow-up. We would like to clarify two points, as our previous response may not have sufficiently highlighted the relevant distinctions and additional evidence.

First, we have not conceded that GIOF's central contribution is merely a combination of existing techniques. We acknowledged that diffusion and continuous-depth propagation are established mathematical tools. The contribution of GIOF is the explicit construction of a descriptor-induced family of admissible generators, together with context-dependent selection of both propagation shape and scale while preserving structural constraints for every selected operator.

This distinction is substantive. As demonstrated by our three-site counterexample, changing the generator mixture modifies relative transition rates, which cannot be achieved by varying the propagation time of a single fixed generator. We do not claim that GIOF subsumes all adaptive diffusion architectures, but its descriptor-to-generator construction and constrained operator selection are the central methodological contributions, rather than a direct assembly of existing diffusion layers.

Second, we would like to highlight the additional evidence addressing the original evaluation concerns. We introduced GPR-GNN as a learned multi-hop propagation baseline under the same reconstruction backbone. At 30% regional missingness, GPR-GNN obtains MAEs of 3.238 and 4.911 on PEMS-BAY and METR-LA, respectively, compared with 3.070 and 4.668 for GIOF.

We also added non-traffic and alternative-geometry reconstruction experiments. The Weather grid, Directed grid, and Barrier grid settings yield Static/GIOF MAEs of 0.432/0.401, 0.278/0.258, and 0.339/0.318, respectively. These experiments support applicability beyond the original traffic geometry, although we do not present them as exhaustive cross-domain validation.

Furthermore, the appendix reports the actual context-dependent selector behavior. For example, mean propagation time increases from 0.49 to 1.24 as regional missingness increases from 10% to 60%, accompanied by changes in the selected generator mixture. This directly addresses the question of how propagation shape and scale respond to different missingness conditions.

We recognize that broader evaluations would further strengthen the empirical evidence. Nevertheless, the revised manuscript now contains a learned multi-hop comparison, additional geometric settings, separate shape/time ablations, and measured selector behavior.

We would appreciate clarification on which specific aspect of the original evaluation concern remains unresolved, so that we can address it precisely.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.