acceptodds
Withdrawn from ICLR 2027

Intrinsic Flow Matching on Quantum Pure-State Manifolds with Phase-Aligned Transport

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum pure-state ensembles live on complex projective space, making flat Euclidean generative modeling geometrically mismatched. We introduce Intrinsic Flow Matching (IFM), a deterministic transport framework on \(CP^d-1\) that learns tangent velocity fields using Pancharatnam phase-aligned conditional paths. IFM replaces local score teachers and reverse-time stochastic sampling with manifold probability flow, while horizontal parameterization removes redundant ambient directions. We show that the IFM objective recovers the induced marginal transport field, represents deterministic projective ensemble flows, and yields endpoint and stability guarantees. Empirically, IFM often improves over ambient Euclidean flow matching across higher-qubit, multimodal, spin-coherent, physics-inspired, and amplitude-encoded MNIST image-vector benchmarks, with strongest gains on high-dimensional and coherence-sensitive tasks but not uniformly across every metric.

Primary area:
generative models
Keywords:
flow matching, quantum machine learning, Riemannian generative models, complex projective space, geometric phase, probability transport

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

The paper proposes a manifold flow matching variant in order to sample and augment quantum states. The method modifies the RFM to a non-linear time schedule and shows in the experiments that it performs better than Euclidean flow matching given the metrics.

Critical strengths

S1) The theory separate the general Riemannian flow-matching results (Theorem 1, ll. 247–258; Propositions 1–2, ll. 261–276) from the results specific to (\mathbb{CP}^{d-1}). These include characterisations of Pancharatnam alignment (Proposition 8, ll. 962–971), horizontal projection as the minimal-norm representative of a projective velocity (Proposition 6, ll. 873–884), and the representative independence and horizontality of the conditional path (Proposition 9, ll. 1001–1014). The paper also proves endpoint transport and an (L^2)-to-Wasserstein stability bound in the Fubini–Study metric (Theorem 2, ll. 281–289; Proposition 3, ll. 294–305).

S2) The experiments include strong controls for the proposed geometric formulation. These contain a matched ambient baseline using the same prior and per-step normalization (D.14; Table 15). It also compares with a exactly (U(1))-equivariant parameterization (D.15; Table 16), ablations over gauge fixing and path choice with both consistent and randomized phase representatives (D.17; Table 18. Lastly a alternative Haar prior (D.16; Table 17), and a 20,000-step baseline run (D.8; Figure 4) is used in the experimental section. The evaluation extends to 12 and 14 qubits. The intrinsic method adds approximately 1 ms per step with little additional memory use (Table 4, ll. 797–818). The paper also report negative results: the generic geodesic baseline performs better on XXZ, and Euclidean flow matching remains competitive on the TFIM and XXZ entanglement (W_1) metric (ll. 413–422).

Critical weaknesses

W1) RFM (Chen and Lipman) uses manifold geodesics with a selectable time schedule. When we use the manifold of \mathbb{CP}^{d-1} the Fubini-Study logarithm first aligns the endpoints and then follows the geodesic curve as it is also presented in Xu et al. This work traces the same curve with a non linear time parameterisation. Hence the the baseline of Geodesic IFM (ours) would be more accurate described as CP-RFM. Further in the appendix the same method has the name gauge-unfixed sphere-lift and this should be also adopted. This is important to note since CP-RFM and IFM are in most reported results similar in performance, considering the variance and for most tables CP-RFM is completely omitted. Hence the newly introduced method with the non linear schedule does not provide any advantage over the RFM version adopted to CP, where this adaptation is minor compared to RFM. Here a proper distinguishment and comparison would be needed to properly evaluate the contribution as it right now dims down to RFM is better than FM on manifolds, which is an already established result.

W2) Xu et al. showed that the MMD is non-characterisitisc as it equals the Hilbert-Schmidt distance between the ensemble-average density matrices. Further the observable metric depends only on the same first moment. Also the per-amplidude fix noise becomes diffuse as the dimension grows and hence the high qubit targets approach Haar. Hence in order to fully evaluate those gains it would be beneficial to also include the specialised metrics of HS-Gaussian MMD or some chordal energy distance, an unchanged-base control. The paper in the current form does not provide convincing evidence of performance improvements over CP-RFM nor about a successful learning of the high dimensional distribution

Questions for authors

NA

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

The paper "INTRINSIC FLOW MATCHING ON QUANTUM PURE-STATE MANIFOLDS WITH PHASE-ALIGNED TRANSPORT" builds upon recent advances in flow matching for specifically strucured manifolds and geometries and introduces intrinsic flow matching for quantum pure-state generation. The authors rely on Pancharatnam-aligned conditional paths and show empirically that their proposed approach outperforms Euclidian Flow Matching for the investigated benchmarks.

Critical strengths

The authors propose a solid extension to the current set of flow matching approaches. Their approach is based directly on the structure of quantum-pure state manifold and the deviation is clearly outlined in the paper. Moreover the authors have shared their code which enhances reproducibility.

Critical weaknesses
  • From a pure machine learning perspective the novelty of the paper is limited in the Reviewers' opinion and it is very specific to its application domain. There however it definitely is of interest.
  • The paper is missing a discussion of the computational costs of the new method and a comparison to Euclidean Flow Matching here.
  • The Reviewer is not an expert on benchmarks for quantum benchmarks, but the ones presented all seem to be on the smaller side. The authors compare them to MNIST and in the appendix the number of training steps does also not seem particularly large in the Reviewers opinion.
  • Moreover, the models described in the Appendix are quiet small and are based on a MLP backbone only. This is a large contrast to the architectures used for many of other state-of-the-art flow matching approaches.
Questions for authors

See weaknesses and:

  • What was the number of parameters for the models employed ?
  • Could you please comment on the training costs of the new method compared to Euclidean Flow Matching ?
Reviewer_8d5POpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies generative modeling of quantum pure-state ensembles. They propose their original method called a "Intrinsic Flow Matching (IFM)", which is designed to learn tangent velocity fields directly on complex projective space. The conditional paths are built with special methods in terms of Pancharatnam phase alignment, and the velocity is parameterized in the horizontal space, which are argued to remove the redundant directions of the ambient space. Instead of local score-based generative models and reverse-time stochastic sampling, the authors claim that IFM uses a deterministic probability flow on the manifold. In the experiments, IFM often outperforms ambient Euclidean flow matching, with the largest gains on high-dimensional and coherence-sensitive tasks, although the improvement is not uniform across all metrics.

Critical strengths

The biggest strength of this paper is that the authors present a flow matching method on the projective space CPd−1CP^{d-1}, which can serve as a model of pure quantum states. Going beyond the model spaces that Riemannian flow matching is based on, they claim to model the vector field efficiently through horizontal modeling in the tangent space. The conditioning of the flow follows the formal approach that existing methods have proposed and explored. In the experiments, the method outperforms other Euclidean or Riemannian methods, which empirically demonstrates the underlying argument the authors make.

Critical weaknesses

This paper raises and tackles an important issue well in its first part, but the core claims all fall apart once the main method is introduced. For example, the paper introduces the horizontal space and motivates this modeling with the sentence “This removes both radial motion and unphysical phase rotation.” However, a formal introduction should come before this in the first place: the horizontal/vertical distributions, the general connection form and the resulting properties of the principal bundle, how these have been used and analyzed in the existing physics literature, the connection between the topological obstruction of projective space and learnability, and so on.

Many such parts are not introduced in a systematically connected way at all. Although I'm experts on this field, i'm unable to fully understand the definition of (7) deeply due to the informal definition of it without careful definition of all mathematical objects!

Next, the following sentence also reads as very unscientific.

“All IFM vector fields, losses, and numerical updates are defined in this tangent-space representation. Compared with diffusion-based quantum generative models (Ho et al., 2020; Song et al., 2021b; Xu et al., 2026), this replaces score learning, local diffusion teachers, and reverse-time stochastic simulation with direct velocity matching and deterministic manifold flow integration.”

How the modeling in the bundle coordinates described above this sentence connects to the proposed vector field and the resulting learning algorithm is not even described, and there is no discussion of whether the method replaces the existing score-based methods. There is only the claim.

An even worse problem, though this is my personal and subjective judgment (I would like the AC and the other reviewers to check this part), is that I think Figure 1 is very likely to have been generated by AI. The colors in the figure are arranged randomly, the shapes are not aligned to their positions and are placed randomly, there are empty gaps on the left and right, and I cannot understand at all what the figure is trying to explain with the curves it shows.

In the theoretical section, the discussion from Theorem 1 onward is not systematically organized and presents the results without a clear flow. Each result is understandable on its own, but the consistency between them is very poor. In addition, the content itself is only an extension of existing Riemannian flow matching results at the level of an example problem, and the rebranding of this as “population optimality”, a non-scientific term that is not clearly defined, is also a main concern in the theory part. When the term “population” is used in a scientific sense, the associated methods are typically built on measure-theoretic results. For all theoretical results, the authors should show the historical context and choose their words precisely.

Questions for authors

Please see the above "Critical Weaknesses" section for details!

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

Related papers

Discussion (0)

Sign in to comment.