acceptodds
Under review as a conference paper at ICLR 2027

VQIL: Visible Quantum Interference Learning for Perturbation Response Ranking on Graphs

Abstract

Learning how local perturbations affect a network is important for identifying which components drive or absorb a system’s response. We formulate this as seed-conditioned response ranking: given a perturbation at one seed node, ranking all other nodes by response strength. This is challenging when the model is given only the interaction topology, while the mechanism governing how perturbations propagate must be learned from data. We present VQIL (Visible Quantum Interference Learning), a trainable Szegedy quantum walk for seed-conditioned ranking. Edge weights, link phases, and per-node phase rotations are optimized end-to-end by backpropagation while the walk operator remains unitary throughout training. Node scores are obtained from a geometrically discounted coherent sum of quantum-walk amplitudes, enabling quantum interference before the final squared-norm readout. Because each learned parameter is attached to a named edge or node, the trained model provides directly inspectable hypotheses about the roles of individual graph components in the propagation mechanism. On AC power-flow perturbations over six transmission grids, with electrical parameters withheld from the model, VQIL improves Spearman rank correlation from for the untrained walk to and outperforms an identically trained random-walk model. On gene perturbation over five protein–protein interaction networks with Perturb-seq supervision, VQIL improves rank correlation from to , outperforming the strongest learned classical and neural baselines. These results show that trainable quantum-walk interference provides a scalable and interpretable inductive bias for learning seed-conditioned response rankings on graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph representation learning, network propagation, quantum walk, perturbation response, interpretable machine learning

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 3 reviews, above 27% of the others.
Reviewer_Kx5zOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a quantum-inspired classical method for ranking node responses to a perturbation at a seed node. It trains a phased Szegedy walk while preserving unitary propagation and computes scores through a geometrically discounted coherent sum of walk amplitudes followed by a squared-magnitude readout. Learned parameters are attached to individual edges and nodes, allowing direct inspection of the fitted model. Evaluation covers six power grids and five Perturb-seq datasets, with perturbation seeds held out during training. Their model showcases improved performance on the trained random-walk baseline and the neural baselines on gene perturbations.

Critical strengths
  1. A conceptually interesting step of including wave phases. I find the combination of a trainable Szegedy walk, a coherent-resolvent readout, and explicit node and edge parameters appealing. It allows the model to learn how propagation paths reinforce or cancel one another while keeping the fitted parameters directly inspectable.

  2. Meaningful empirical gains supported by targeted ablations. Across five Perturb-seq datasets with held-out perturbation seeds, VQIL achieves mean Spearman correlation of 0.620, compared with 0.582 for GCN and 0.570 for trained RWR (Table 2). Ablations further support its design choices: adding node phases improves correlation from 0.587 to 0.620, while coherent geometric accumulation achieves 0.620 versus 0.517 for its square-then-add counterpart (Tables 2–3). Together, these results demonstrate predictive improvements over trained competitors and support the usefulness of node phases and coherent accumulation within the proposed architecture, although their advantage over comparably expressive alternatives remains unresolved.

Critical weaknesses
  1. The conceptual novelty appears limited. VQIL combines an existing phased Szegedy construction with end-to-end ranking supervision and a coherent-resolvent readout. Phased walks, trainable quantum-walk operators, and graph-associated learned propagation parameters have precedents, as acknowledged in Section 3. The contribution therefore appears to lie primarily in the particular combination of parameterization, training objective, and readout, rather than introducing interference or learnable walk dynamics. I consider this combination potentially useful, but do not see a substantial conceptual advance without significant empirical results. This makes convincing evidence of an empirical or interpretability advantage over closely matched existing approaches particularly important; the current comparisons do not yet establish that advantage conclusively.

  2. The comparisons do not sufficiently isolate the advantage of phase-based propagation. Interference and cancellation also occur in classical waves and signed models. While the Szegedy construction justifies "quantum-inspired," the novelty should concern the trainable formulation and coherent readout, rather than interference itself. The O(2)O(2) sheaf baseline uses feature-dependent, MLP-parameterized rotations, so it is not a closely matched, directly graph-parameterized control. Moreover, trained RWR learns edge weights with a fixed restart probability, whereas VQIL additionally learns node phases. A classical model with learned edge weights and per-node restart probabilities, alongside a matched cancellation-capable propagator, would help isolate the contribution. The ablations show that phases and coherent accumulation help within VQIL, but do not establish their advantage over comparably expressive alternatives.

  3. Parameter visibility is better supported than mechanistic explainability. The parameter–annotation associations are useful, particularly the generator-bus result, but edge-weight correlations with co-expression are almost identical for VQIL and RWR (0.11–0.28 versus 0.11–0.27). The node-phase interpretation is not compared against a classical model with learned node parameters. Moreocer the paper does not explain which paths reinforce or cancel for individual predictions, or validate Section 6's stronger claim that phase magnitudes identify explicit sinks where perturbations are halted. Prediction-specific analyses or targeted intervention tests would strengthen the mechanistic interpretation.

  4. The symmetric link phases appear mathematically redundant. In Eqs. (1)–(3), define Dϕ∣i,j⟩=eiϕij∣i,j⟩D_\phi|i,j\rangle=e^{i\phi_{ij}}|i,j\rangle. Since ϕij=ϕji\phi_{ij}=\phi_{ji}, DϕD_\phi commutes with the swap SS, giving Uϕ=DϕU0Dϕ†U_\phi=D_\phi U_0D_\phi^\dagger, where U0U_0 has zero link phases but unchanged weights and node phases. The seed and readout states transform by the same DϕD_\phi, implying xv(ω,ϕ,δ)=xv(ω,0,δ).x_v(\omega,\phi,\delta)=x_v(\omega,0,\delta). Under the equations as written, link phases therefore cannot affect predictions: their lack of empirical benefit reflects an exact redundancy rather than a dataset-dependent observation. This does not invalidate the main vertex-phased model, but requires clarification of the trainable parameter space.

Questions for authors
  1. What evidence separates the benefit of the proposed mechanism from that of adding node-specific parameters? Are matched results available for RWR with learned edge weights and per-node restart probabilities, or a directly graph-parameterized signed/rotation-based propagator? Without such controls, how narrowly should the claimed advantage be interpreted?

  2. Beyond parameter–annotation correlations, what supports Section 6's interpretation of ∣δ∣|\delta| as identifying sinks where perturbations are halted? Does the recurrent-response association persist when calculated on held-out perturbations rather than the training perturbations used in Figure 6?

  3. Could you provide a parameter-matched comparison with a classical signed or rotation-based propagator, using the same number of learnable edge and node parameters, a comparable propagation horizon and readout, and the same training and evaluation protocol? This would help determine whether VQIL’s gains are specific to its propagation rule or could be reproduced by an equally parameterized classical model with node-level flexibility and cancellation capabilities.

  4. Do Eqs. (1)–(3), with ϕij=ϕji\phi_{ij}=\phi_{ji}, imply xv(ω,ϕ,δ)=xv(ω,0,δ)x_v(\omega,\phi,\delta)=x_v(\omega,0,\delta), explaining the lack of link-phase benefit? Can you confirm that setting link phases to zero without retraining leaves predictions unchanged?

5 comments
AuthorsOpenReview ↗

Authors' Response (Part 1)

We thank the reviewer for the detailed and technically precise review. We ran the controls you asked for and answer each point below.

W1& Q1 (empirical evidence against closely matched existing approaches). All controls you name are now available, together with a lazy walk. Each adds one learned scalar per node to Trained-RWR (m+nm+n parameters, as in Vertex-phased-VQIL) and uses the protocol of Tables 1–2 (restart r=0.1r=0.1, same steps, folds, loss and optimiser):

  • (a) Per-node restart: at node vv the walker returns to the seed with probability rv=σ(logit⁡0.1+ρv)r_v=\sigma(\operatorname{logit}0.1+\rho_v).
  • (b) Signed propagator: x←P(τ∘x)+r esx\leftarrow P(\tau\circ x)+r\,e_s with per-node transmission τv=(1−r)cos⁡ψv\tau_v=(1-r)\cos\psi_v, scored by ∣xv∣|x_v|.
  • (c) Rotation-based propagator: the same with τv=(1−r) eiψv\tau_v=(1-r)\,e^{i\psi_v}.
  • (d) Lazy walk: at node vv the walker stays with probability αv=σ(logit⁡0.1+ηv)\alpha_v=\sigma(\operatorname{logit}0.1+\eta_v) and otherwise moves to a neighbor, i.e. P′=P diag(1−α)+diag(α)P'=P\,\mathrm{diag}(1-\alpha)+\mathrm{diag}(\alpha).

Full definitions are given under Q3.

ModelParametersPhase angleVoltagePPI
Vertex-phased-VQILm+n\mathbf{m+n}0.9115 ± 0.04330.8306 ± 0.05640.6196 ± 0.1670
Trained-RWR (weights only)mm0.7772 ± 0.09140.6814 ± 0.09030.5695 ± 0.1500
(a) Per-node restartm+nm+n0.9014 ± 0.06410.7954 ± 0.07380.6048 ± 0.1609
(b) Signed propagatorm+nm+n0.8620 ± 0.06010.7198 ± 0.08700.5670 ± 0.1575
(c) Rotation-based propagatorm+nm+n0.8518 ± 0.08460.7195 ± 0.08680.5676 ± 0.1567
(d) Lazy walkm+nm+n0.8469 ± 0.08510.7401 ± 0.08070.6018 ± 0.1611

Spearman ρ\rho, mean ± sd over 30 cells (power grids) and 25 cells (PPI); mm edges, nn nodes.

The five folds of a task share one graph, so the unit of analysis is the grid (n=6n=6) or the dataset (n=5n=5). We report per-cell win counts together with the number of independent units won. Vertex-phased-VQIL exceeds (b) and (c) in 30/30 (phase angle) and 30/30 (voltage) grid cells and in 25/25 PPI cells, and exceeds (d) in 30/30, 30/30 and 24/25 cells. In all three comparisons it wins in all six grids for each response and in all five PPI datasets. Against (a) it wins in 27/30 voltage cells (5/6 grids) and in 25/25 PPI cells (5/5 datasets), and is level on phase angle (17/30 cells, 3/6 grids). None of the four classical walks matches VQIL on all three tasks at the same parameter count.

On the grids every control has a higher mean ρ\rho than Trained-RWR, so node-specific parameters do account for part of the gap in Table 1. The remaining difference lies in how the node parameter enters the walk. The node phase is the phase θv=π+δv\theta_v=\pi+\delta_v of the reflection at node vv (Eq. 1) and acts inside a unitary step: it does not remove probability, hold the walker, or add moves outside the graph, and it changes the scores only through the coherent readout (Eq. 3), in which amplitudes are summed before the squared norm. The classical parameters instead remove or hold probability mass at the node, as in (a) and (d), or rescale or rotate the transmitted signal of a non-unitary propagator, as in (b) and (c). We therefore state the claim as follows: at the setting of Tables 1–2 and with the same number of parameters, Vertex-phased-VQIL outperforms weight-only RWR and controls (b)–(d) in every grid and PPI dataset, and matches or exceeds the per-node-restart walk (a) without any jump to the seed.

W3 & Q2 (support for the interpretation of |δ|).

Held-out perturbations. Yes, the recurrent-response association persists when it is computed only on the test perturbations of each fold, which were never used to fit δ\delta. Recurrent response is the fraction of perturbations in which a gene is among the 10% strongest responders. We control for degree because degree is known from the graph without training and acts in the opposite direction: ∣δ∣\lvert\delta\rvert decreases with degree (ρ\rho = −0.40 to −0.60), whereas responsiveness increases with it (ρ\rho = 0.16 to 0.38).

Recurrent response vs abs(δ)K562RPE1FrangiehTian-iAdamson
Degree-controlled ρ, training0.4216 ± 0.01490.3994 ± 0.02090.2084 ± 0.03320.4377 ± 0.01200.2683 ± 0.0196
Degree-controlled ρ, held-out0.4137 ± 0.01540.3909 ± 0.02020.2001 ± 0.02980.4272 ± 0.00840.2497 ± 0.0174
AUC (top 10%), training0.8373 ± 0.00790.8705 ± 0.01290.7030 ± 0.00600.8223 ± 0.00570.7452 ± 0.0100
AUC (top 10%), held-out0.8318 ± 0.00800.8676 ± 0.01260.6761 ± 0.00750.8094 ± 0.00870.7291 ± 0.0119

Mean ± sd over five folds.

AuthorsOpenReview ↗

Authors' Response (Part 2)

Intervention tests and "sinks". We intervened on the trained models without retraining. δv\delta_v enters the readout through a factor on every walk passing through node vv (onward transmission, including its phase) and a factor on node vv's own score (response gain).

Spearman ρPhase angleVoltagePPI
Vertex-phased-VQIL (reference)0.9115 ± 0.04330.8306 ± 0.05640.6196 ± 0.1670
δ = 0 at the 5% of nodes with largest abs(δ)0.8655 ± 0.05990.7777 ± 0.07180.5780 ± 0.1583
δ = 0 at 5% random nodes0.9071 ± 0.04580.8204 ± 0.06050.6197 ± 0.1672

Mean ± sd over cells, untruncated readout.

High-∣δ∣\lvert\delta\rvert nodes matter for the prediction: knocking them out costs 0.046–0.051, against 0.004–0.008 for random nodes. δ\delta acts through the node's own response gain, not by blocking onward transmission. We therefore agree that "sinks where perturbations are halted" in Section 6 was imprecise, and we have aligned it with Section 5.2: ∣δ∣\lvert\delta\rvert tracks a node's local response propensity, low at voltage-regulated buses and high at recurrently responsive genes.

Classical model with learned node parameters. We trained RWR with a learned per-node restart probability (same number of parameters, same protocol). On the voltage response, ∣δ∣\lvert\delta\rvert identifies generator buses with AUC 0.7386 ± 0.1003, against 0.6398 ± 0.1111 for the per-node restart (30 cells; higher on 5 of 6 grids).

Edge weights. We agree: the VQIL and RWR weight fields carry the same co-expression information, so the readability of edge weights is shared with the classical walk. What VQIL adds is the node parameter. We now state this explicitly.

W2 & Q3 (parameter-matched classical propagator). We agree with the points of W2 and answer them in turn. All numbers refer to the table under Q1.

Sheaf baseline. We agree that Sheaf-O(2), with feature-dependent, MLP-parameterized rotations, is not a closely matched control. It remains in Tables 1–2 as a neural baseline only. The matched controls are (a)–(d) of Q1, which are parameterized directly on the graph: one scalar per edge and one per node, no features and no MLP.

Learned edge weights with per-node restart. This is control (a). It improves on Trained-RWR by +0.124+0.124 / +0.114+0.114 / +0.035+0.035 (phase angle / voltage / PPI; 6/6 grids for each response, 5/5 PPI datasets), so a learned per-node restart accounts for part of the gap in Table 1. Vertex-phased-VQIL exceeds it in 27/30 voltage cells (5/6 grids) and in 25/25 PPI cells (5/5 datasets), and is level on phase angle (17/30 cells, 3/6 grids).

Matched cancellation-capable propagator. We trained both a signed and a rotation-based propagator, controls (b) and (c). They are classical, non-unitary propagators on the same graph with one learned angle ψv\psi_v per node:

xt+1=P (τ∘xt)+r es,x0=es,scorev=∣xv∣,x_{t+1}=P\,(\tau\circ x_t)+r\,e_s,\qquad x_0=e_s,\qquad \text{score}_v=|x_v|,

where PP is the transition matrix of Trained-RWR and τv\tau_v is the transmission of node vv: τv=(1−r)cos⁡ψv\tau_v=(1-r)\cos\psi_v for the signed propagator and τv=(1−r) eiψv\tau_v=(1-r)\,e^{i\psi_v} for the rotation-based one. A walk through nodes v1,…,vkv_1,\dots,v_k contributes with the factor τv1⋯τvk\tau_{v_1}\cdots\tau_{v_k}, so walks reaching a node along different routes can cancel, by sign in (b) and by rotation in (c). Setting ψ≡0\psi\equiv0 recovers Trained-RWR.

The comparison is matched to Vertex-phased-VQIL as follows.

  • Parameters: mm edge weights with the parameterization of Eq. 2 and nn node angles, initialized and clamped exactly like δv\delta_v (uniform in [−0.3,0.3][-0.3,0.3], clamped to [−π,π][-\pi,\pi]); the count is m+nm+n for all three models.
  • Horizon: restart r=1−γ=0.1r=1-\gamma=0.1, the setting used for Trained-RWR in Tables 1–2.
  • Readout: the signal is accumulated over walk lengths and the modulus is taken once at the end, as in the coherent readout of Eq. 3.
  • Training and evaluation: identical folds, loss, optimizer, early stopping and model selection.

Cancellation helps the classical walk on the power grids: (b) and (c) improve on Trained-RWR by +0.085+0.085 / +0.038+0.038 and +0.075+0.075 / +0.038+0.038 (phase angle / voltage), and neither changes the PPI result. Vertex-phased-VQIL exceeds both in 30/30 phase-angle cells, 30/30 voltage cells and 25/25 PPI cells.

Advantage over comparably expressive alternatives. The ablations in the paper compare variants within VQIL only, as you note. The table under Q1 adds the comparison with classical models of the same parameter count (m+nm+n). At the setting of Tables 1–2, Vertex-phased-VQIL outperforms the matched signed and rotation-based propagators and the lazy walk in every grid and PPI dataset, and matches or exceeds the walk with learned per-node restart.

AuthorsOpenReview ↗

Authors' Response (Part 3)

W4 / Q4 (symmetric link phases).

Thank you for this precise observation. Your derivation is correct: for φij=φji\varphi_{ij}=\varphi_{ji}, the diagonal unitary Dφ∣i,j⟩=eiφij∣i,j⟩D_\varphi|i,j\rangle = e^{i\varphi_{ij}}|i,j\rangle commutes with the swap SS, so Uφ=DφU0Dφ†U_\varphi = D_\varphi U_0 D_\varphi^\dagger, and since the seed and readout states transform by the same DφD_\varphi,

xv(ω,φ,δ)=xv(ω,0,δ)x_v(\omega,\varphi,\delta) = x_v(\omega,0,\delta)

holds exactly. We also confirmed it numerically: setting φ=0\varphi=0 in the trained Link-phased-VQIL models, without retraining, leaves all predictions unchanged up to floating-point precision (maximum deviation of order 10−1610^{-16}).

This does not affect our results or conclusions. Our main model is Vertex-phased-VQIL, which has no link phases: its trainable parameters are the edge weights ω\omega and the node phases δ\delta, and all headline results, the readout ablation and the interpretability analysis use this variant. Link phases appear only in the Link-phased and Graph-phased ablation rows, which we included because the underlying graph-phased Szegedy walk admits them. Your argument explains exactly why those rows coincide with Standard-VQIL and Vertex-phased-VQIL (differences ∣Δρ∣≤0.006|\Delta\rho|\le 0.006, due to random initialization only).

In the revision we state this invariance explicitly in Section 4.1, describe the Link-phased and Graph-phased rows as a consistency check, and remove link phases from the list of trained components in the abstract and contributions.


References

(a) Per-node restart W. Jin, et al. "Supervised and Extended Restart in Random Walks for Ranking and Link Prediction in Networks." PLOS ONE, 2019.

(b) Signed propagator Y. Tian and R. Lambiotte. "Spreading and Structural Balance on Signed Networks." SIAM Journal on Applied Dynamical Systems, 2024.

(c) Rotation-based propagator Y. Tian and R. Lambiotte. "Structural Balance and Random Walks on Complex Networks with Complex Weights." SIAM Journal on Mathematics of Data Science, 2024.

(d) Lazy walk P. Du, et al. "Supervised Lazy Random Walk for Topic-Focused Multi-Document Summarization." IEEE International Conference on Data Mining (ICDM), 2011.


We thank you for the detailed and constructive comments, which helped us isolate the role of the node phase with matched classical controls and sharpen the claims of the paper. We hope the responses above resolve your concerns.

Best regards,

Authors of Submission 61787

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

This paper proposes an information diffusion framework based on the Szegedy Quantum Walk (SQW) to simulate propagation processes on complex graphs. The main contribution lies in incorporating complex link phases and custom node/edge states into the standard SQW architecture to model diffusion dynamics. However, the theoretical novelty and several methodological issues of the proposed framework require further justification. I would be open to raising my rating if the authors adequately address these concerns.

  1. Could the authors clarify what unique dynamics the quantum walk captures that conventional models fail to reveal?

  2. The core algorithm relies on the standard SRSR operators from SQW with added phase rotations. What is the fundamental theoretical breakthrough here beyond a direct parametric extension? How does this differ conceptually from existing phased or coined quantum walk models?

  3. Real-world information diffusion is inherently dissipative (e.g., information fading, memory decay). However, the simulated quantum walk is unitary and probability-conserving. How do the authors reconcile this fundamental contradiction?

  4. The evaluation only compares against Random Walk with Restart (RWR), omitting standard diffusion models like Independent Cascade (IC) or Linear Threshold (LT). Why were IC/LT models excluded?

Critical strengths

This paper introduces the Szegedy Quantum Walk (SQW) framework to the field of information diffusion simulation.

Critical weaknesses
  1. It remains unclear what specific propagation dynamics or real-world diffusion phenomena this complex-phased quantum walk can reveal that conventional classical models fail to capture.

  2. The core algorithm relies on the standard SRSR (Swap and Reflection) operators from standard Szegedy Quantum Walks with added phase rotations. The work appears to be a direct parametric extension rather than a fundamental theoretical breakthrough, and its conceptual distinction from existing phased/coined quantum walk models needs stronger justification.

  3. Real-world information diffusion is inherently non-conservative and dissipative (e.g., information fading, memory decay, and heat reduction). In contrast, the simulated quantum walk relies on pure unitary evolution, which strictly conserves total probability.

  4. The experimental evaluation only compares the proposed framework against Random Walk with Restart (RWR), omitting established information diffusion baselines such as Independent Cascade (IC) and Linear Threshold (LT) models. Without benchmarking against these standard diffusion frameworks, the practical utility and claimed advantages of the proposed approach remain unproven.

Questions for authors
  1. Could the authors clarify what unique dynamics the quantum walk captures that conventional models fail to reveal?

  2. The core algorithm relies on the standard SRSR operators from SQW with added phase rotations. What is the fundamental theoretical breakthrough here beyond a direct parametric extension? How does this differ conceptually from existing phased or coined quantum walk models?

  3. Real-world information diffusion is inherently dissipative (e.g., information fading, memory decay). However, the simulated quantum walk is unitary and probability-conserving. How do the authors reconcile this fundamental contradiction?

  4. The evaluation only compares against Random Walk with Restart (RWR), omitting standard diffusion models like Independent Cascade (IC) or Linear Threshold (LT). Why were IC/LT models excluded?

2 comments
AuthorsOpenReview ↗

Authors' Response (Part 1)

We thank the reviewer for the careful reading and the constructive questions. We clarify the main points and respond to each question below.

W1 & Q1 (what the quantum walk captures).

Whereas a classical random walk is driven by a transition matrix with nonnegative elements, the original Szegedy quantum walk is driven by a real-valued unitary operator. Although in the end both produce a probability distribution, the quantum walk deals with real-valued amplitudes during the evolution rather than with probabilities, so interference effects that are not possible in classical walks can arise. Thus, from a mathematical point of view, a quantum walk offers a richer expressivity. Moreover, the complex-phase extension makes the dynamics driven by a complex-valued operator, which increases the expressivity further.

Despite this increased expressivity, we acknowledge that there is no deep theoretical proof that a quantum walk always beats any classical model. However, treated as heuristic algorithms, quantum walks have proved useful in the literature, outperforming their classical counterparts in classification problems [1,2]. For this reason quantum walks, and in particular the complex-phased Szegedy walk, were chosen as a good candidate for the VQIL algorithm.

We feel that this question is clarified by the text in the introduction, lines 67–73.

W2 & Q2 (difference from existing phased or coined quantum walks).

A coined quantum walk can be described by a swap (also referred to as shift) and a coin operator. A natural problem is how to construct the coin operator so that it represents a sensible dynamics on the graph. In principle, any unitary operator with a block-diagonal structure would be valid, with one unitary coin block for each node. For those blocks, the usual choices are Hadamard coins for linear graphs and Grover coins for general graphs. However, these coins carry no information about the graph beyond its backbone structure.

A solution to the question of how to construct the coin operator comes from the Szegedy quantum walk. It is indeed a coined quantum walk, and it provides a sensible recipe for constructing the coin as the reflection operator RR. This operator is block-diagonal, and each block is a reflection around the state given by the transition probabilities of a given node. Thus, the coin is constructed according to further properties of a weighted graph and captures relevant information during the evolution.

Adding complex phases does not modify the inherent coined feature of the quantum walk, but it increases the capability of representing graph information in the coin operator. Beyond transition probabilities, further edge parameters can be embedded in the link phases, and node parameters in the arbitrary phase rotations. So, again, the complex-phase extension provides a recipe to construct a sensible coin that captures graph information.

Although the way the operator is extended may look like a purely parametric extension, we want to highlight that the extension follows the mathematical properties of the operator, which has to be unitary, and aims at capturing crucial information of the graph.

As usual in applied quantum-walk algorithms, the relevance of the results comes from actual simulation results that show the advantages, more than from fundamental theoretical results. In our case, we use this model, which enables the codification of graph properties in the dynamics, as a heuristic, and we show through training simulations its enhanced performance and its visibility properties.

The introduction in lines 73–77 has been revised to address this question and to clarify the choice of the graph-phased Szegedy quantum walk.

AuthorsOpenReview ↗

Authors' Response (Part 2)

W3 & Q3 (unitary evolution versus dissipative diffusion).

First, we want to point out that a random walk, despite not being reversible, also conserves the total probability. Thus, conserving the total probability does not imply that there are no dissipative effects. For example, a random walk can converge to the uniform distribution and thereby erase all information about the initial state.

Regarding a quantum walk, it is true that, in a bare sense, it implies unitary evolution, so that the final quantum-state amplitudes do not dissipate any information. However, at the end of any quantum algorithm there is also a measurement process that produces the probability distribution. The final probability distribution cannot be inverted to obtain the initial state, so information has, in some sense, also been dissipated.

Moreover, the previous comment applies only to a simple quantum walk that starts from a state and gives the final distribution. In our case we obtain the amplitudes of the quantum-walk states at different time steps and take a weighted sum prior to the measurement, which introduces additional dissipation. We thus want to remark that the quantum walk performs a unitary evolution, but it is used as a subprocess in a more complex quantum-inspired classical algorithm whose results do not come from a simple unitary process, and it is therefore compatible with real-world information diffusion.

The disclosure in lines 240–246 has been revised to address this question and to clarify how unitary propagation within VQIL is compatible with predicting responses in dissipative systems.

W4 & Q4 (baselines).

We chose trained random walk with restart (RWR) as the matched classical control for seed-conditioned response ranking. RWR and VQIL use the same edge-weight parameterization, loss, optimizer, seed-disjoint splits, and early-stopping protocol. We set the restart probability to r=1−γr = 1 - \gamma to match the geometric discount across walk lengths. This comparison lets us evaluate probability-based propagation and coherent amplitude aggregation under closely matched training conditions. Furthermore, the evaluation also includes GCN, GAT, and O(2)O(2) neural sheaf diffusion. In the revision we also added three classical walks with one learned parameter per node (lazy, per-node-restart and signed random walks), trained under the same protocol (Tables 1–2). To make the task and the model easier to follow, we added an overview figure (Figure 1) that shows the seed-conditioned ranking task, the learnable parameters of VQIL, and the training and readout of the learned parameters.

The reason why we did not include Independent Cascade (IC) or Linear Threshold (LT) is that our targets are continuous perturbation-response magnitudes, whereas IC and LT model node activation. To obtain a comparable node-level score, we would need the probability that each node becomes active given the seed. Computing these probabilities exactly is intractable on general graphs. Estimating them from sampled cascades would introduce sampling noise into the ranking loss and require gradient estimation through discrete activation decisions, which would considerably increase the difficulty of training such a model.


References

[1] L. Bai, et al. "Quantum Kernels for Unattributed Graphs Using Discrete-Time Quantum Walks." Pattern Recognition Letters, 2017.

[2] H. Saarinen, et al. "Disease Gene Prioritization with Quantum Walks." Bioinformatics, 2024.


We thank you for the careful reading and the constructive questions, which helped us clarify the motivation for the quantum walk, its relation to existing coined and phased walks, and how unitary propagation fits dissipative systems. We hope the responses above resolve your concerns.

Best regards,

Authors of Submission 61787

Reviewer_jy5FOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates seed‑conditioned response ranking: given a perturbation at one seed node, rank all other nodes by response strength, using only interaction topology. It proposes VQIL, a trainable Szegedy quantum walk on the directed‑edge space of dimension mm, with step operator U=S⋅RU=S\cdot R (a register swap times a sum of Householder reflections R=∑i((1−eiθi)∣ψi⟩⟨ψi∣−I)R=\sum_{i}\big((1-e^{i\theta_i})|\psi_i\rangle\langle\psi_i|-I\big) ) whose unitarity holds for every parameter value by construction. Three parameter sets are trained end‑to‑end by backpropagation through KK walk steps: symmetric edge log‑weights ω\omega (defining a degree‑normalised stochastic matrix pp), symmetric link phases φ\varphi, and per‑node phase rotations δi\delta_i with θi=π+δi\theta_i=\pi+\delta_i. Node scores come from a coherent geometrically‑discounted resolvent readout ∣χs⟩=∑k=0KγkUk∣ψs⟩|\chi_s\rangle=\sum_{k=0}^{K}\gamma^{k}U^{k}|\psi_s\rangle followed by a single squared‑norm readout xv=∣⟨ψv∣χs⟩∣2x_v=|\langle\psi_v|\chi_s\rangle|^{2}, with γ=0.9\gamma=0.9 and K=109K=109 chosen so the discarded tail γK≤10−5\gamma^{K}\le 10^{-5}. Training minimises a soft‑Spearman loss (Blondel et al., 2020) on log‑transformed standardised scores, with AdamW, lr 0.050.05, ≤300\le 300 epochs, early stopping on validation Spearman, in float64/complex128. Two application domains are used: AC power‑flow perturbation ranking on six IEEE/PEGASE grids (∣Δθ∣|\Delta\theta| and ∣ΔV∣|\Delta V| responses to +20+20 MW / +6+6 MVAr injections) and Perturb‑seq gene knockdown ranking on five PPI networks (K562, RPE1, Frangieh, Tian‑i, Adamson from scPerturBench). Evaluation is seed‑disjoint 5‑fold. Headline results: grid ρ\rho improves 0.1470.147 to 0.8310.831 (untrained to Vertex‑phased‑VQIL) and PPI ρ\rho 0.1710.171 to 0.6200.620, beating a matched classical Trained‑RWR control (0.7770.777 / 0.6810.681 / 0.5700.570) and, on PPI, GCN (0.5820.582), GAT (0.5280.528) and Sheaf‑O(2) (0.3650.365). Six ablation arms separate the contributions of weights, link phases and node phases, a 2×22\times2 readout factorial isolates the coherent‑vs‑square‑then‑add and uniform‑vs‑geometric choices, and the trained fields are validated against withheld network properties (bus type, generator status, recurrent responsiveness).

Critical strengths
  1. The matched classical control is the right first baseline, and VQIL genuinely beats it. Trained‑RWR fixes the weight parameterisation (Eq. 2), loss, optimiser, splits and early stopping, with propagation as the key difference; VQIL wins by +0.134 (grid angle), +0.149 (grid voltage) and +0.050 (PPI) in ρ\rho. This is the paper's cleanest test of the propagation change: it isolates the unitary walk from the parameterisation, and the margins are consistent across all three tasks.

  2. The 2×22\times2 readout factorial is a genuinely useful ablation (Table 3). It separates the coherent (add‑then‑square) from the square‑then‑add readout and the geometric from the uniform discount at matched horizon (0.703 / 0.782 / 0.911 / 0.839 on grid angle), and the geometric ×\times coherent cell is best without the two factors being redundant. It shows that the gain is not a single tuning artefact but a specific combination of design choices, and it is the part of the empirical programme I would keep in a revision.

Critical weaknesses
  1. The matched classical control is the right first baseline, and VQIL genuinely beats it. Trained‑RWR fixes the weight parameterisation (Eq. 2), loss, optimiser, splits and early stopping, with propagation as the key difference; VQIL wins by +0.134 (grid angle), +0.149 (grid voltage) and +0.050 (PPI) in ρ\rho. This is the paper's cleanest test of the propagation change: it isolates the unitary walk from the parameterisation, and the margins are consistent across all three tasks.

  2. The 2×22\times2 readout factorial is a genuinely useful ablation (Table 3). It separates the coherent (add‑then‑square) from the square‑then‑add readout and the geometric from the uniform discount at matched horizon (0.703 / 0.782 / 0.911 / 0.839 on grid angle), and the geometric ×\times coherent cell is best without the two factors being redundant. It shows that the gain is not a single tuning artefact but a specific combination of design choices, and it is the part of the empirical programme I would keep in a revision.

Questions for authors
  1. On the missing per‑node classical control. A classical walk with per‑node parameters (for example, per‑node restart probabilities) and a tuned‑restart RWR would help separate the per‑node phase δ\delta from an additional per‑node bias (Jin et al., 2019). Is such a control feasible, and what does it show?

  2. On oversmoothing and the selection of γ\gamma. Please report Dirichlet energy, MAD, rank, or a smoothness‑versus‑depth curve at γ=0.9\gamma=0.9 and γ=0.7\gamma=0.7, and state the criterion used to select γ\gamma. If the mechanism claims cannot be measured, it would be better to present them as hypotheses.

  3. On significance testing. Please report the Wilcoxon tests at the level of independent units (n=6n = 6 grids, n=5n = 5 datasets) rather than folds, or replace the p‑values with per‑cell win counts and a clear statement of the unit of analysis.

  4. On the PPI interpretability claim. The ∣δ∣|\delta| annotation is computed from the same training perturbations used to fit δ\delta. Please report the direct correlation between ∣δ∣|\delta| and training‑set responsiveness, and a held‑out‑perturbation or held‑out‑node evaluation rather than the in‑sample annotation.

  5. On the GCN and GAT voltage entries. Table 1 reports identical 0.920±0.0410.920\pm0.041 for both; please confirm whether these are independent runs and report the number of seeds and standard errors.

3 comments
AuthorsOpenReview ↗

Authors' Response (Part 1)

We thank the reviewer for the careful reading and the constructive questions. We noticed that the Weaknesses field of the review contains the same text as the Strengths field, so we could not identify specific weaknesses to respond to there. We have therefore treated Questions 1–5 as the reviewer's main concerns, ran the requested controls, and answer each point below.

Q1 (per-node classical control). We agree that this control was missing. The node phase has no exact classical counterpart: it enters a unitary reflection and acts locally at its node, without removing amplitude and without adding moves that are absent from the graph. We therefore distinguish a local classical analogue, whose node parameter changes only the transition rule at its own node, from the non-local per-node restart you suggest, whose node parameter acts through a jump to the seed. Both controls add one learned scalar per node to Trained-RWR (m+nm+n parameters, as in Vertex-phased-VQIL) and use the protocol of Tables 1–2 (restart r=0.1r=0.1, same steps, folds, loss and optimizer).

Local control: lazy walk [1]. At node vv the walker stays with probability αv=σ(logit⁡0.1+ηv)\alpha_v=\sigma(\operatorname{logit}0.1+\eta_v) and otherwise moves to a neighbour with the learned edge probabilities, i.e. P′=P diag(1−α)+diag(α)P'=P\,\mathrm{diag}(1-\alpha)+\mathrm{diag}(\alpha). Scores follow the restart iteration x←(1−r)P′x+r esx\leftarrow(1-r)P'x+r\,e_s. The nn parameters ηv\eta_v are initialized at 00 and trained jointly with the mm edge weights.

Non-local control: per-node restart [2]. At node vv the walker returns to the seed with probability rv=σ(logit⁡0.1+ρv)r_v=\sigma(\operatorname{logit}0.1+\rho_v), with ρv\rho_v trained jointly with the edge weights.

ModelParametersPhase angleVoltagePPI
Vertex-phased-VQILm+nm+n0.9115 ± 0.04330.8306 ± 0.05640.6196 ± 0.1670
Trained-RWR (weights only)mm0.7772 ± 0.09140.6814 ± 0.09030.5695 ± 0.1500
Lazy walk (local control)m+nm+n0.8469 ± 0.08510.7401 ± 0.08070.6018 ± 0.1611
Per-node restart (non-local control)m+nm+n0.9014 ± 0.06410.7954 ± 0.07380.6048 ± 0.1609

Spearman ρ\rho, mean ± sd over 30 cells (power grids) and 25 cells (PPI).

Both controls improve on Trained-RWR in every grid and dataset, so node parameters account for part of the gap in Table 1, as you anticipated. Following Q3, the independent unit is the grid (n=6n=6) or the PPI dataset (n=5n=5), and a cell is one unit × one fold. Vertex-phased-VQIL is above the lazy walk in 30/30 cells (6/6 grids) for both grid responses and in 24/25 cells (5/5 datasets) on PPI. Against per-node restart it is above in 27/30 cells (5/6 grids) on voltage and in 25/25 cells (5/5 datasets) on PPI, and level on phase angle (17/30 cells, 3/6 grids). We will add these controls to Table 1 and state the claim as: at the setting of Tables 1–2, VQIL outperforms weight-only RWR and the lazy walk, and matches or exceeds the per-node-restart walk. What the node phase adds is therefore node-level flexibility inside a local, unitary step: with the same number of parameters it is more accurate than the local classical control, and it matches or exceeds the per-node restart without any jump to the seed.

Q2 (smoothness metrics and the criterion for γ\gamma). Thank you. We ran the requested diagnostic (new Appendix Figure 3, summarized in the table below) and revised Section 4.2.

Criterion for γ\gamma. Both γ\gamma values were fixed before any experiment and were never tuned on validation or test data. They correspond to restart probabilities r=1−γ=0.1r = 1-\gamma = 0.1 and 0.30.3, both standard choices for random walk with restart [3,4], and they are the values at which our RWR control is run. We use γ=0.9\gamma = 0.9 as the main setting because it is more accurate in every block (table below), while a longer horizon is costly: the number of walk steps for a given truncation tolerance grows as 1/(1−γ)1/(1-\gamma), each step costing O(m)O(m) forward and backward (about 2×2\times at γ=0.95\gamma = 0.95, 10×10\times at γ=0.99\gamma = 0.99). γ=0.9\gamma = 0.9 is a fixed operating point, not a tuned optimum.

We report MAD, the mean average cosine distance between node representations (a node's vector of within-query rank scores over the held-out seeds), as a function of the readout depth (Appendix Figure 3).

Spearman ρ, γ = 0.7Spearman ρ, γ = 0.9MAD trained, γ = 0.7MAD trained, γ = 0.9MAD untrained, γ = 0.7MAD untrained, γ = 0.9
Power grids, phase angle0.8488 ± 0.08530.9115 ± 0.04330.2491 ± 0.01020.2343 ± 0.03250.2478 ± 0.00860.2072 ± 0.0452
Power grids, voltage0.7039 ± 0.08020.8306 ± 0.05640.2204 ± 0.01950.1793 ± 0.02900.2478 ± 0.00860.2072 ± 0.0452
PPI networks0.5870 ± 0.16050.6196 ± 0.16700.1347 ± 0.03020.0596 ± 0.02650.1242 ± 0.02370.0272 ± 0.0075
AuthorsOpenReview ↗

Authors' Response (Part 2)

On the power grids MAD rises with depth and plateaus. It does not collapse as the readout deepens, so we see no oversmoothing. On the PPI networks MAD peaks within a few steps and then falls, so the predictions trend toward a seed-agnostic ranking as the horizon grows. The untrained walk falls further, so training restores part of the seed specificity. A longer horizon therefore buys accuracy at the price of seed specificity. The two domains differ because the PPI graphs are far denser (average degree 20–76 and diameter 6–10, against 2.5–5.8 and 6–25; Appendix Table 4), so a walk covers them within a few steps.

We agree that the sentence in Section 4.2 (γ\gamma kept below the mixing time so that the readout retains the seed's identity) was stated without measurement. It is now presented as a hypothesis: the measurements are consistent with it on the power grids, while on the PPI networks the readout trends toward seed-agnostic predictions.

Q3 (p-value reporting). We agree. The five folds of a task share one graph and are not independent, so the unit of analysis is the grid (n = 6) or the dataset (n = 5), not the fold. Following your suggestion, we have removed all p-values from these comparisons and state the unit of analysis in Section 5.1. Each comparison now reports per-cell win counts together with the number of independent units won. The conclusions are unchanged: Vertex-phased-VQIL exceeds Standard-VQIL in 27/30 (phase angle) and 28/30 (voltage) grid cells and in 25/25 PPI cells, and exceeds the trained RWR control in 30/30, 30/30 and 25/25 cells. In both comparisons it wins in all six grids for each response and in all five PPI datasets.

Q4 (in-sample |δ| annotation). Thank you for raising this. We computed both quantities you asked for.

(i) Direct correlation with training-set responsiveness. The Spearman correlation between ∣δ∣\lvert\delta\rvert and a gene's mean response over the training perturbations is 0.05–0.29 across the five datasets. This raw value understates the association because node degree acts in the opposite direction: ∣δ∣\lvert\delta\rvert decreases with degree (ρ\rho = −0.40 to −0.60), whereas responsiveness increases with degree (ρ\rho = 0.16 to 0.38). Degree is known from the input graph without training, so herewe report degree-controlled values (0.22–0.47), which show what ∣δ∣\lvert\delta\rvert carries beyond degree.

(ii) Held-out perturbations. We recomputed the annotation using only the test perturbations of each fold, which were never used to fit δ\delta. The results are essentially unchanged (table below). We report three quantities:

  • Mean response is a gene's mean response over the perturbations. This is the training-set responsiveness you asked.
  • Recurrent response is the fraction of perturbations in which the gene is among the 10% strongest responders. This is the quantity used for the annotation in the paper.
  • AUC measures how well ∣δ∣\lvert\delta\rvert separates the 10% most recurrent responders from all other genes, computed within degree deciles and averaged. This is the statistic reported in the paper.
Association with abs(δ)K562RPE1FrangiehTian-iAdamson
Mean response: degree-controlled ρ, training0.4234 ± 0.01490.4138 ± 0.02080.2157 ± 0.03840.4682 ± 0.00940.2657 ± 0.0246
Mean response: degree-controlled ρ, held-out0.4194 ± 0.01610.4120 ± 0.01900.2045 ± 0.03730.4612 ± 0.01140.2582 ± 0.0237
Recurrent response: degree-controlled ρ, training0.4216 ± 0.01490.3994 ± 0.02090.2084 ± 0.03320.4377 ± 0.01200.2683 ± 0.0196
Recurrent response: degree-controlled ρ, held-out0.4137 ± 0.01540.3909 ± 0.02020.2001 ± 0.02980.4272 ± 0.00840.2497 ± 0.0174
Recurrent response: AUC (top 10%), training0.8373 ± 0.00790.8705 ± 0.01290.7030 ± 0.00600.8223 ± 0.00570.7452 ± 0.0100
Recurrent response: AUC (top 10%), held-out0.8318 ± 0.00800.8676 ± 0.01260.6761 ± 0.00750.8094 ± 0.00870.7291 ± 0.0119

Values are mean ± sd over five folds.

We note that δ\delta is a per-node parameter learned transductively, so the annotation applies to genes present in the training graph. We now state this scope explicitly and report the held-out numbers in the appendix.

AuthorsOpenReview ↗

Authors' Response (Part 3)

Q5 (identical GCN/GAT entries). These are independent runs. The match is a coincidence of rounding to three decimals. On the voltage response, GCN scores 0.9199 ± 0.0405 and GAT 0.9195 ± 0.0411 (mean ± sd over 30 cells). The two models differ in all 30 cells (mean ∣Δρ∣=0.008|\Delta\rho| = 0.008) and in their per-grid results (mean ± sd over five folds):

case30case57case118case145case300case1354All (30 cells)
GCN0.8780 ± 0.03420.8961 ± 0.05470.9066 ± 0.00350.9142 ± 0.00620.9645 ± 0.00120.9603 ± 0.00130.9199 ± 0.0405
GAT0.8870 ± 0.04160.8879 ± 0.05050.9067 ± 0.00420.9096 ± 0.01740.9657 ± 0.00140.9603 ± 0.00120.9195 ± 0.0411

References

[1] P. Du, et al. "Supervised lazy random walk for topic-focused multi-document summarization." IEEE International Conference on Data Mining (ICDM), 2011.

[2] W. Jin, et al. "Supervised and Extended Restart in Random Walks for Ranking and Link Prediction in Networks." PLOS ONE, 2019.

[3] J. Gasteiger, et al. "Predict then Propagate: Graph Neural Networks meet Personalized PageRank." International Conference on Learning Representations (ICLR), 2019.

[4] S. Erten, et al. "DADA: Degree-Aware Algorithms for Network-Based Disease Gene Prioritization." BioData Mining, 2011.


We appreciate your thoughtful feedback, which has helped us clarify the classical controls, smoothness analysis, and statistical reporting of our work. We hope the revisions and additional experiments above address your concerns.

Best regards,

Authors of Submission 61787

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.