acceptodds
Under review as a conference paper at ICLR 2027

What Survives Higher-Order Signal Reduction? Harmonic Unobservability at Vertices

Abstract

Higher-order models represent signals across vertices, edges, faces, and higher-rank cells, yet collapsing these representations to a lower rank raises a fundamental question: which higher-order information remains observable after reduction? We characterize this information pathway for rank-shared boundary-mediated reductions and show that cross-rank transfer necessarily factors through boundary operators. Under the Hodge decomposition, the retained lower-rank representation can access boundary-visible gradient components, whereas harmonic components are structurally unobservable. We further show that this loss is caused by the observation interface itself rather than by the numerical treatment of singular higher-rank blocks. Motivated by this characterization, we introduce HARP, a HArmonic Readout Pipeline that explicitly exposes the missing harmonic information. HARP constructs a topology-preserving Morse complex, computes its harmonic basis, lifts the resulting representatives through a chain-compatible map, and projects them onto the harmonic subspace of the original complex. The resulting basis yields compact, sample-dependent harmonic coordinates that complement the information accessible through lower-rank reduction. Experiments on synthetic and real higher-order signals validate the predicted information pathway and show that HARP recovers useful information specifically when harmonic components are task-relevant. Together, these results identify a precise blind spot of rank-shared higher-order reduction and provide a compact interface for preserving the information hidden beyond it. Code will be released upon acceptance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Deep Learning, Higher-order Graph Learning, Discrete Morse Theory, Higher-order Reduction

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_ssHnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents an extension to the Collapse Effective Operator paper. It analysis how higher order signal information such as curl and harmonic are not preserved with the collapse of the higher order connectivity to the vertex level set. Consequently it then proposes the HARP features, which retain those higher order information. In experiments it then is shown that HARP indeed recovers the curl and harmonics.

Critical strengths
  1. Theoretical Correctness Every rank-shared, boundary-mediated coupling to the retained vertex space factors through B1B_1. Components in ker⁡B1=curl⁡⊕harmonic⁡\ker B_1=\operatorname{curl}\oplus\operatorname{harmonic} therefore cannot reach that space. In Theorem 3.3, the authors show that this is valid by the construction of the operator, when folding the topology to the vertex level.

  2. The experiments support the theory The matched two-dimensional baselines, including random features, train-only PCA, and non-harmonic Hodge modes, allow us to distinguish harmonic information from a generic low-dimensional embedding effect. The pair-safe synthetic protocol and the basis-ambiguity experimental results support the evaluation.

Critical weaknesses
  1. The authors does not clearly separate its contribution from CEO Sections 2–3.2 closely follow the CEO construction. Equation 5 is CEO’s L∗L^\ast with βk=1\beta_k=1, while Equation 3 reproduces its regularised Schur complement with renamed blocks. The transfer map TεT_\varepsilon in Equation 4 is new. Theorem 3.3 specialises CEO’s Proposition 3.10 to the harmonic component and adds the finite-mass case, but the relationship is not attributed clearly. The new contribution appears to be the transfer map together with the Hodge-theoretic interpretation of the interface. Theorem 3.2 is also difficult to assess because assumption (i) already contains much of its conclusion.

In the signed setting the gradient subspace is then defined as: $ M^\downarrow_\varepsilon = \Delta_0-\gamma_1^2\Delta_0(\Delta_0+\varepsilon I)^{-1}. $ This reduced operator is therefore a graph-Laplacian filter with no contribution from cells above rank 1. Theorems 3.1 and 3.3 then follow as corollaries. At a rank-0, curl and all rank-≥2\geq2 signals, along with the harmonic component, are lost.

The authors do give a reason for emphasising harmonics: curl remains visible at the rank-1/rank-2 interface, whereas harmonic signals are annihilated by both adjacent boundaries. That distinction no longer applies once the retained space is rank 0, however, because both components lie in ker⁡B1\ker B_1. The paper should make this change of perspective explicit.

CEO’s combinatorial-complex formulation, including the MANTRA experiment, replaces signed boundary matrices with binary incidence matrices. In that setting B1B2≠0B_1B_2\neq0, as well as higher-order cells can affect the collapsed operator. The authors should therefore explain why the signed setting is the appropriate basis for its critique, state clearly that the same unobservability result does not carry over to the unsigned case, and, ideally, characterise what information survives there.

  1. The experiments instantiate the reduction only at rank 0 and omit the most important competing baseline. CEO allows any rank to be retained. Keeping edges while eliminating vertices and triangles produces an operator of the form f(B1⊤B1)+g(B2B2⊤).f(B_1^\top B_1)+g(B_2B_2^\top). Under CEO’s PSD condition, its kernel is exactly H1H_1. The same reduction framework can expose the harmonic coordinates computed separately by HARP while also retaining curl information. An edge-level collapse, or at least a standard simplicial/Hodge filter, should be included in Tables 2–4 to demonstrate how a edge signal behaves if the operator is collapsed to the edges. Without this, the experiments establish a limitation of choosing rank 0, not of rank-shared reduction in general. Several existing baselines are also disadvantaged by construction. The “Schur-reduced” signal in Equation 19 is introduced by the authors instead of the original of CEO construction. Every backbone in Table 4 receives only B1zB_1z, so the information HARP is designed to recover has already been removed before learning begins.
  1. Finally, Table 5 shows that HARP performs similarly to direct Hodge projection. Schaub et al. (2020) already applied Hodge projection to the same dataset, and the flow-completion setup follows Jia et al. (2019), although it uses a different fitting basis. The paper should position HARP directly against these methods rather than relying primarily on a vertex-level reduction that is provably unable to access the relevant signal.
Questions for authors

Q1. CEO allows the retained rank to be chosen. If the edges are retained while the vertices and triangles are eliminated, the resulting operator has the form f(B1⊤B1)+g(B2B2⊤).f(B_1^\top B_1)+g(B_2B_2^\top). Under CEO’s PSD condition, its kernel is H1H_1. The harmonic coordinates used by HARP are therefore already present as the zero modes of this rank-1 reduction. Could the authors include this reduction, or a comparable simplicial/Hodge filter, in Tables 2–4? More generally, is the unobservability result about rank-shared reduction itself, or only about retaining rank 0?

Q2. The theory assumes signed boundary operators satisfying Bk−1Bk=0B_{k-1}B_k=0. CEO’s combinatorial-complex experiments, including MANTRA, use binary incidence matrices, for which B1B2≠0B_1B_2\neq0 and higher-rank cells can affect the vertex-level operator. What remains observable in this setting? Is there an analogue of HARP when a Hodge decomposition is no longer available?

Reviewer_sp3WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes HARP, a harmonic readout pipeline for preserving higher-order signal information that is otherwise lost when reducing signals to vertex-level representations. The authors first show that rank-shared boundary-mediated reduction cannot capture harmonic components, and then introduce a pipeline based on discrete Morse theory, chain-compatible lifting, and harmonic projection to recover this information as compact coordinates. Experiments on both synthetic and real-world higher-order signals support the theoretical findings. The authors also demonstrate improvements in classification and flow completion tasks where harmonic information is useful.

Critical strengths
  • The paper provides a clear theoretical characterization of the higher-order signal information lost during boundary-mediated reduction.
  • The proposed pipeline follows naturally from the theoretical analysis and provides a compact representation of the missing harmonic information.
  • The experiments include useful controls and ablations that support the theoretical findings, including cases where harmonic information offers limited benefits.
Critical weaknesses
  • I am not fully convinced that the use of discrete Morse theory is necessary. The final HARP representation spans the same harmonic subspace as direct Hodge decomposition, and it is unclear what additional benefits the topology-preserving construction and lifting map provide for the evaluated tasks. The computational overhead is also nontrivial. For example, Table 10 shows that DMT takes 17.5 seconds, compared to only 0.71 seconds for sparse Lanczos. The harmonic dimension can also grow substantially for larger complexes.
  • The main unobservability result seems to follow directly from standard properties of boundary operators and Hodge decomposition, and harmonic projection itself is a well-established technique. In particular, I find it difficult to identify a substantial methodological advance in the proposed pipeline. A more direct comparison with Schaub et al. (2020) and Ebli et al. (2024) would help clarify the work's contribution.
  • I find the experiments less convincing in demonstrating the broader practical value of HARP. In the backbone experiments, the baselines only receive boundary-visible inputs, whereas HARP-augmented models are additionally given harmonic coordinates. Given the theoretical analysis, the resulting performance gains are somewhat expected. It is unclear whether HARP offers an advantage over more general higher-order models that can directly process edge features or incorporate alternative cross-rank couplings. I think comparisons with stronger higher-order baselines would make the empirical contribution more convincing.
Questions for authors
  • In Table 5, direct Hodge decomposition and HARP span the same harmonic subspace but achieve slightly different classification results. Could the authors clarify whether these differences mainly come from basis rotation, feature standardization, or regularization?
  • For the flow-completion experiments, could the authors report the relative energy of the gradient and harmonic components for each dataset?
Reviewer_mzRpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper characterizes the information lost when a graded higher-order system is reduced by rank-shared, boundary-mediated Schur complementation: harmonic (and curl) components cannot reach the retained vertex space, under any treatment of the eliminated kernel. HARP supplements the reduced representation with harmonic coordinates obtained via discrete Morse reduction, chain-compatible lifting, and projection.

Critical strengths
  • The distinction between reducing an operator and transmitting a higher-rank input signal is valuable and carefully formulated.
  • Theorem~3.1 rests on an invariant-subspace argument, not merely on B1h=0B_1 h = 0.
  • The projected Morse construction is correct: lifting preserves cycles and projection recovers reference-harmonic representatives.
  • Useful controls (paired synthetic signals, matched-dimensional projections, learned anisotropic couplings, direct Hodge comparison, a gradient-dominated counterexample), and appendices that candidly delimit the result.
Critical weaknesses

My major concern is that the Morse pipeline adds nothing to the downstream tasks. HARP returns an orthonormal basis of the same harmonic subspace as direct Hodge decomposition; on a fixed complex the two differ by an orthogonal rotation. The reduced complex and explicit lift may have independent value, but no experiment uses them: every prediction task consumes only harmonic coordinates. Morse-based signal compression on chain complexes is also prior work. The paper needs a task that requires the Morse complex or the lifting map.

In addition, the theoretical results are narrower than their framing. Theorem 3.2's locality assumption already encodes Y=B1ΓY=B_1\Gamma, so ``rigidity'' follows almost by definition, and the isotropic case Γ=γI\Gamma=\gamma I is assumed rather than derived. The paper's anisotropic experiments show the obstruction disappears outside the rank-shared class. Theorem 3.1 also makes curl equally unobservable, and HARP discards curl too; what survives reduction is the gradient component only. The appendix scope discussion is appreciated, but the main text should match it.

Questions for authors
  • Which downstream operation benefits from the Morse complex or the lifting map, beyond obtaining the harmonic projector?
  • Can Table 4 be rerun with bases that receive the full edge signal, or the reduced vertex signal of eq. (19), rather than B1zB_1 z?
  • How should harmonic coordinates be aligned across complexes? Sign canonicalization does not resolve the O(βk)O(\beta_k) rotational ambiguity for βk>1\beta_k>1.

Comments on the paper

1 comment

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.