acceptodds
Under review as a conference paper at ICLR 2027

Scaling Universality: A Framework for Function Representations

Abstract

A fundamental question in function approximation theory is whether physical operator scale bounds permanently restrict representational capacity or merely reparameterize it. Standard universal approximation theorems evaluate function classes under fixed input and output conventions, leaving a theoretical gap for scale-bounded computational substrates even when target functions remain fully recoverable via output rescaling. We resolve this gap by introducing scaling universality, an algebraic framework that incorporates input dynamic ranges and output scale bounds directly into the representation criterion. The central theoretical insight is the formulation of an output scale threshold (): rather than requiring representation at a fixed output scale, scaling universality requires representations to hold across all output scales exceeding . This threshold requirement elevates output scaling from an ad-hoc heuristic to a closed algebraic framework, guaranteeing structural closure and representation existence for both exact and approximate function classes under linear precomposition, output scaling, addition, and composition across operational depth. To demonstrate the framework in the norm-bounded regime of quantum computing, we instantiate scaling universality via Generalized Quantum Signal Processing (GQSP), proving exact scaling universality for real polynomials on via real-valued GQSP primitives and deriving the analytical normalization threshold governing this space. Extending to multivariate domains, we combine GQSP primitives with Linear Combinations of Unitaries (LCU) and a scale-compatible recursive superposition architecture to prove approximate scaling universality for continuous functions on . This work establishes the representational completeness and algebraic closure of scale-bounded primitives, providing a rigorous theoretical foundation for functional representation decoupled from downstream resource trade-offs such as sampling complexity, compilation depth, and parameter optimization.

Primary area:
learning theory
Keywords:
Scaling universality, function representation, quantum function representation, Generalized Quantum Signal Processing, polynomial approximation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iFGoOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

In the paper "Scaling Universality: A Framework for Function Representations," the authors introduce scaling universality, a framework for studying the representational capacity of function classes subject to input and output scaling constraints. The central idea is that a bounded function class can represent a broader class of functions through appropriate rescaling. The authors formalize this idea by requiring that representations exist for all sufficiently large output scales, allowing these guarantees to extend through operations such as addition and composition. They establish closure properties for this framework and apply it to quantum signal processing, proving exact scaling universality for real polynomials and approximate scaling universality for multivariate continuous functions.

Critical strengths

The paper provides a systematic mathematical treatment of scaling constraints in function representations. The authors establish closure properties under several elementary operations and connect these results to a concrete quantum computing formulation. The theoretical development is detailed, with proofs supporting the main claims, although I cannot independently assess their correctness. The mathematical analysis, particularly the quantum signal processing results, is outside my primary area of expertise.

Critical weaknesses
  1. My main concern is that I do not understand what problem the proposed framework ultimately resolves. At its core, scaling universality allows a bounded function class to represent functions of arbitrary magnitude by introducing an external scaling factor. The authors then establish that these representations preserve certain properties under addition and composition. While this is a reasonable mathematical construction, it is unclear what we learn from these guarantees that could not already be established through standard normalization arguments or existing universal approximation results.

    The authors motivate the framework through physical constraints on computational primitives, but the analysis ultimately concerns the existence of representations under arbitrary output rescaling. I would like to see a more concrete explanation of what limitation this resolves and why it matters.

  2. The paper does not include empirical results or numerical examples. I do not consider experiments strictly necessary for a theoretical paper, but a simple demonstration of how the scaling threshold behaves for representative functions or composed architectures would help motivate the framework. As written, the practical implications of the theoretical results remain unclear.

Questions for authors
  1. What concrete limitation of existing universal approximation theory does scaling universality resolve? Can the authors provide an example where this framework establishes a result that would not follow from conventional normalization or approximation arguments?

  2. What does the scaling threshold tell us about a representation? How does it change with polynomial degree, function complexity, or composition depth, and does it provide useful information about the resulting representation?

  3. What is the main new insight from applying this framework to quantum signal processing? Does it establish previously unknown representational capabilities, or primarily provide a different formulation of existing results?

Reviewer_DdmAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Hello PCs, SACs, ACs, and fellow reviewers,

I share that I have found serious hallucinations and miscitations on the references in this manuscript, grave enough to be considered desk rejection in my standard. I respectfully request the ACs, SACs, and PCs to inspect this submission and I strongly recommend desk rejection. To the fellow reviewers, I respectfully request to reflect this grave ethics violation issue on your reviews. Below are the hallucinations and miscitations I have found so far:

Hallucinations

  • "Rahul Sarpeshkar. Analog versus digital: Computational efficiency, precision, and dynamic range. IEEE Micro, 18(6):40–51, 1998." in the manuscript. This one does not exist. I have checked that the corresponding pages of the cited journal issue are of completely different papers.
  • "Rui Chao Chao, Dawei Ding, Andrew Girth, Richard Kueng, Guang Hao Low, and Martin Suchara. Finding angles for quantum signal processing with polynomial root finding. arXiv:2003.02831, 2020." in the manuscript. The title and author list are both wrong. The actual paper seems to be: Rui Chao, Dawei Ding, András Gilyén, Cupjin Huang, and Mario Szegedy. Finding Angles for Quantum Signal Processing with Machine Precision. arXiv:2003.02831, 2020. The link is: https://arxiv.org/pdf/2003.02831.
  • "Yulong Dong, Xiangling Meng, K Birgitta Whaley, and Lin Lin. Efficient phase-factor determination in quantum signal processing. Physical Review A, 103(4):042419, 2021." in the manuscript. Wrong title and author list. The actual paper is at https://journals.aps.org/pra/abstract/10.1103/PhysRevA.103.042419.
  • "Andr´ as Gily´ en, Yuan Su, Guang Hao Low, and Nathan Wiebe. Quantum singular value transforma- tion and its algorithmic applications. In Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing (STOC), pp. 193–204, 2019." in the manuscript. Wrong title. The actual paper is at: https://dl.acm.org/doi/abs/10.1145/3313276.3316366
  • "Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Meng, Mark Tang, Andrew Howard, Dmitry Allen, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2704–2713, 2018." in the manuscript. Wrong author list. Actual paper is at https://ieeexplore.ieee.org/abstract/document/8578384?casa_token=407DEV-2Cd4AAAAA:-Q2NysNUgFl6iTOxchwrMgBrGvItio91_gFOBa3X1drCHxdBwXsJGu5Vrd5bRSz3EZ-UjpGxcvw
  • "John M Martyn, Trevor M Rossi, Andrew K Tan, and Isaac L Chuang. Grand unification of quantum algorithms. PRX Quantum, 2(4):040203, 2021." in the manuscript. Wrong author list, actual paper is at https://journals.aps.org/prxquantum/abstract/10.1103/PRXQuantum.2.040203

Miscitations Below are the list of miscitations in the manuscript where I find the authors have either over/under stated or misused the references in their citations. I omit detailed explanations, and I am open to discuss and provide detailed reasons why they are miscitations:

  • page 1, lines 48-52, "they treat scaling as a static operational step rather than a structural prop- erty of the function class itself." : substantial overstatement on the three citations
  • page 9, lines 448-449, "Alternative superposition representations—such as those by Kolmogorov (1961), Sprecher (1965), or Doss (1977)—do not generally satisfy the hypotheses of Theorem 2.18. ": this statement is not supported by the citations
  • page 9, lines 471-472, "for high-degree multivariate systems (Chao et al., 2020; Dong et al., 2021)": the Chao and Dong papers' setting is univariate QSP.

I think the hallucination and miscitation in the references are too many and serious in this manuscript. Therefore, I respectfully request desk rejection on this manuscript.

Critical strengths

N/A

Critical weaknesses

Aside the grave hallucination and miscitation issues, I find the content of the manuscript is also flooding with faults.

  • Theorem 2.18 checks domain compatibility for architecture A, then claims universality for a differently scaled B, whose immediate values can leave the allowed domain.
  • Theorem 4.2 invokes closure without explicitly verifying the required intermediate range conditions.
  • The polynomial universality does not guarantee computational complexity nor parametric/sample efficiency. In principle, large enough MLP can be trained to be GPT 6 Astra, but that will never happen within the universe's life span. That is why the convergence rate and complexity / efficiency analysis is important. Very amateur.
Questions for authors

I am very convinced that the authors have AI-generated the references and maths without even proof reading once.

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

The paper defines exact and approximate scaling universality by requiring reconstruction of each input-rescaled target at every sufficiently large output scale. It develops elementary and architecture-level closure statements, characterizes a real-polynomial class obtained from GQSP through Chebyshev coefficients, and derives its exact normalization threshold. It then combines univariate approximation and a superposition representation to claim multivariate continuous-function universality. The contribution is theoretical and concerns representation existence rather than training or computational efficiency.

Critical strengths
  1. The univariate normalization result is precise and useful within the specified construction: Proposition 3.18 (p. 7; proof pp. 19–20) identifies the necessary and sufficient scale using the unit-circle norm of the polynomial formed from Chebyshev coefficients. The real-part closure argument is also supplied. This is more informative than merely choosing a loose sufficient coefficient bound.

  2. The paper explicitly separates representation from operational cost: The signed reconstruction factor, elementary addition/composition arguments, and circuit constructions provide a coherent language for combining bounded primitives. The manuscript appropriately avoids claiming efficient training, efficient measurement, or a quantum speedup from existence alone.

Critical weaknesses
  1. Theorem 2.18 is false as stated because its range condition concerns AA while its conclusion concerns BB: On p. 4, Eq. (10) controls pre-composition ranges of A[G]A[\mathcal{G}], although AA and BB may differ by arbitrary nonzero internal scales. This does not ensure that B[G]B[\mathcal{G}] is defined on D1D_1. A concrete counterexample is as follows. Let D_1=\{0\},\qquad D_2=\{0\}\cup Q.$$ Let $\mathcal{F}=\{f_c:\mathbb{R}\to\mathbb{R},\ f_c(x)=c:c\in\mathbb{R}\}$ and $\mathcal{G}=\{g_q:D_2\to\mathbb{R},\ g_q(x)=q:q\in Q\}$. The domain of each $g_q$ is exactly $D_2$. For any input scale $\alpha\in\mathbb{R}$, $f_c(\alpha x)=c$. If $c\neq 0$, for every $\beta\neq 0$ choose $$\begin{aligned} j&=\left\lceil\frac{|\beta|}{|c|}\right\rceil,\qquad q=\frac{\operatorname{sgn}(c/\beta)}{j},\\ \|\beta g_q-f_c(\alpha\cdot)\|_{D_2} &=|\beta q-c|\leq\frac{|c|^2}{|\beta|}. \end{aligned}$$ Thus, for any $\varepsilon>0$, the threshold $\tau=|c|^2/\varepsilon$ suffices. If $c=0$, take $q=1/j$ with $j>|\beta|/\varepsilon$. Hence $\mathcal{G}$ is approximately scaling-universal for $\mathcal{F}$ on $D_2$. With the identity input map, take $$A=\square\circ\bigl(1\cdot(\square\circ\mathrm{id})\bigr),\qquad B=\square\circ\bigl(\sqrt{2}\cdot(\square\circ\mathrm{id})\bigr).$$ Both sets are compact, $\mathcal{F}$ is continuous and input-scaling-invariant, and $A\sim B$. The pre-composition ranges for $A[\mathcal{G}]$ lie in $D_2$, while those for $A[\mathcal{F}]$ lie in $\mathbb{R}=\operatorname{int}(\operatorname{dom}(\mathcal{F}))$; the identity input map also sends $D_1$ into $D_2$. However, every inner scaled output of $B$ at $0$ is $\pm\sqrt{2}/j\notin D_2$. No member of $B[\mathcal{G}]$ is defined at $0$, although $A[\mathcal{F}]$ contains all constant targets.

The intended application also needs an explicit normalized architecture: Theorem 4.2 (p. 9) invokes Theorem 2.18 after a superposition inclusion, but does not specify the scaling factors that establish its hypotheses. The following remark suggests a repair, but does not demonstrate it. Please state conditions on the actual realizing architecture and verify them for the multivariate construction. The counterexample does not itself disprove the claimed quantum universality result.

  1. The conceptual and quantum novelty is insufficiently distinguished from existing results: The unit-circle realizability condition is already the central GQSP result (Motlagh and Wiebe, Corollary 5); translating it through Chebyshev coefficients gives the paper's threshold characterization. Moreover, for an output-contraction-closed class G\mathcal{G}, a reconstruction β0g\beta_0g with β0≠0\beta_0\neq 0 already extends to every ∣β∣≥∣β0∣|\beta|\geq|\beta_0| by replacing gg with (β0/β)g(\beta_0/\beta)g. The paper's real GQSP class has this property. Thus, in its main example, the new output-scale quantifier needs a clearer explanation of what it adds beyond this closure and ordinary normalized approximation.

    Earlier work already establishes multivariate quantum approximation with explicit circuit-size bounds and quantitative error bounds for Lipschitz targets (Yu et al., Theorems 1–3), and fixed-encoding, single-qubit uniform approximation of continuous multivariate functions bounded in magnitude by one, using a circuit matrix element as output (Pérez-Salinas et al., Theorem 1). Recursive polynomial block-encoding networks also predate this submission (Ivashkov et al., QKAN, Sections III–IV). These are different architectures, so they do not establish duplication. Nevertheless, the introduction's broad gap claim needs a direct comparison showing the additional guarantee provided here. A nontrivial example where the proposed framework resolves a limitation of existing normalization/composition arguments would materially strengthen the contribution.

Questions for authors
  1. How will Theorem 2.18 be corrected to exclude the counterexample above? Please provide the revised range conditions and the explicit normalized architecture used in Theorem 4.2, including why its scales can be fixed independently of the target and approximation tolerance.
  2. For the contraction-closed GQSP class, which representation result or architectural guarantee is unavailable from existing normalized approximation and block-encoding constructions? Please compare assumptions and conclusions directly with the three approximation/network works cited above.
  3. Definitions E.3–E.6 allow α=0\alpha=0. Consequently, a bounded constant representation class can satisfy existential scaling universality for all continuous targets merely by reconstructing f(0)f(0). Is this degeneracy intended, and how does correcting it affect the motivating comparison?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.