⚛️ quantum physics

Comment on "Supervised Learning with Quantum Hamiltonian Embeddings" (arXiv:2610.26833)

This paper critiques the novelty and computational feasibility of "Supervised Learning with Quantum Hamiltonian Embeddings" by contextualizing it within prior Quantum Cognition Machine Learning (QCML) developments and challenging the claim of encoding an entire dataset into a single ground state.

Original author: Alexander G. Abanov, Luca Candelori, Kharen Musaelian

Published 2026-11-04

📖 3 min read🧠 Deep dive

CC BY 4.0

Original author: Alexander G. Abanov, Luca Candelori, Kharen Musaelian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly evolving field of artificial intelligence, researchers are constantly searching for ways to make machines learn faster and more efficiently. One promising avenue involves borrowing tools from quantum physics, the branch of science that describes how the smallest particles in the universe behave. A central idea in this approach is to treat data not just as numbers on a screen, but as physical states within a system. Imagine a landscape where every possible piece of information corresponds to a specific position, and the goal of learning is to find the lowest, most stable point in that landscape. This concept, known as a ground state, has long been a theoretical dream for scientists hoping to solve complex classification problems that overwhelm traditional computers. The question driving recent work is whether encoding data into these quantum states offers a genuine advantage over classical methods, or if it is merely a different way of doing the same old math.

A recent paper by researchers at Qognitive, Inc., and Stony Brook University addresses a specific claim made in a study by Molteni and colleagues regarding "Supervised Learning with Quantum Hamiltonian Embeddings." The authors of the new paper argue that the core idea of encoding data into a quantum system was not new. They point out that their own organization had previously developed a method called QCML, or Quantum Cognition Machine Learning, which relies on the same fundamental principle of mapping data to physical states. In fact, the Qognitive team states that they shared this concept, along with the necessary computer code, with the authors of the Molteni paper during a collaboration that included IBM researchers. They note that they had already successfully run their version of this machine learning task on an IBM quantum processor, demonstrating that the basic approach was already in practice.

The specific difference highlighted in the new paper concerns how the data is organized within the quantum system. In the earlier QCML method, each individual piece of data is assigned its own unique physical setup, creating a distinct state for every single item. The new paper proposes a different strategy where all the data is packed into a single, shared physical setup. The authors of the critique suggest that while this new method might offer some computational speed, it fundamentally changes the nature of the work in a way that may not be necessary. They argue that the quantum part of the process becomes largely decorative. In their view, the specific information about each data point is handled by a set of rotations, while the shared state acts as a common background.

The researchers demonstrate that if one were to remove the shared background state entirely, the specific rotations could be adjusted to do the exact same job. This implies that the complex quantum machinery used to create the shared state might be unnecessary. The paper concludes that the authors of the original study have not proven that their method of learning the shared state and the rotations is better than simply learning the combined effect of both directly. If the latter approach works just as well, it suggests that the entire quantum computation could be replaced by a simpler, classical calculation, rendering the quantum aspect of the experiment ornamental rather than essential.

Based on the text provided, here is a detailed technical summary of the document.

Note on Source Material: The provided text is not a research paper presenting new experimental results or a novel methodology. Rather, it is a commentary (or letter) authored by Alexander G. Abanov, Luca Candelori, and Kharen Musaelian of Qognitive, Inc., addressed to the authors of a paper titled "Supervised Learning with Quantum Hamiltonian Embeddings" (arXiv:2610.26833). The text critiques the claims of novelty and technical significance made in the referenced paper.

Problem Statement

The commentary addresses the claim of novelty in the referenced paper [1], which proposes a machine learning approach where data is encoded as the ground state of a Hamiltonian parameterized by that data. The authors of this commentary argue that this core concept is not new, having been previously introduced and developed by Qognitive under the framework of Quantum Cognition Machine Learning (QCML).

Methodology and Technical Critique

The commentary outlines the following technical distinctions and critiques regarding the methodology:

Prior Art (QCML): The commentary asserts that the idea of encoding data as a ground state of a parameterized Hamiltonian was first introduced by Qognitive. This framework was shared with the authors of the referenced paper (IBM) and formed the basis of a shared publication [5] where QCML was successfully executed on IBM's Heron processor for binary classification. The codebase shared included backpropagation algorithms identical to those used in the referenced paper.
The Specific Modification: The commentary acknowledges that the referenced paper introduces a specific modification: encoding the entire dataset into a single Hamiltonian and a single ground state. In contrast, the QCML framework encodes each data point with its own Hamiltonian, resulting in a distinct state for each data point. The authors note that this per-data-point encoding is considered essential to the QCML method.
The "Ornamental" Argument: The commentary challenges the necessity of the referenced paper's approach. It argues that if the ground state $|\psi(\alpha)\rangle$ $∣ ψ (α)⟩$ represents a rotation of a basis shared by all data points, and the data-specific information is encoded via data-dependent rotations $U(x, \alpha)$ $U (x, α)$ , the quantum computation may be redundant.
- The authors propose that one could define a rotation $V(\alpha)$ such that $V(\alpha)|0^n\rangle = |\psi(\alpha)\rangle$ .
- They suggest that $|\psi(\alpha)\rangle$ could be replaced by the standard zero state $|0^n\rangle$ , and the rotation $U(x, \alpha)$ could be replaced by a composite rotation $\tilde{U}(x, \alpha) = V(\alpha)U(x, \alpha)$ .
- Conclusion of Critique: The commentary posits that learning $|\psi(\alpha)\rangle$ and $U(x, \alpha)$ separately offers no demonstrated advantage over learning the combined operation $\tilde{U}(x, \alpha)$ directly, which would eliminate the need for quantum computation entirely.

Key Claims and Contributions of the Commentary

Priority of Concept: The authors claim that the foundational idea of the referenced paper was introduced by Qognitive as QCML, based on concepts from Cognitive Science.
Collaborative History: The commentary highlights a history of collaboration between Qognitive and IBM, including the sharing of codebases and the joint execution of QCML on IBM hardware.
Discussion of the "Single Hamiltonian" Approach: The authors state that the idea of encoding the entire dataset into one Hamiltonian was extensively discussed during Qognitive/IBM collaboration sessions, specifically noting a working session on May 17, 2024.
Skepticism of Quantum Advantage: The commentary concludes that the modification proposed in the referenced paper (single ground state for all data) may render the quantum computation "entirely ornamental" rather than providing a computational advantage.

Results and Significance

No New Experimental Results: The text does not present new experimental data, benchmarks, or future implications.
Significance: The significance of this document lies in its challenge to the novelty and technical necessity of the referenced paper. It asserts that the referenced work builds upon established QCML concepts without adequately distinguishing its contributions or demonstrating why the proposed "single ground state" architecture is superior to existing methods or classical equivalents.

Disclaimer: This summary reflects the arguments and claims made within the provided commentary text. It does not validate the scientific accuracy of the claims against the referenced paper [1], but rather summarizes the position taken by the authors of this commentary.

Problem Statement

Methodology and Technical Critique

Key Claims and Contributions of the Commentary

Results and Significance

More like this