⚛️ quantum physics

Optimal Quantum Likelihood Estimation

This paper proposes an information-theoretic optimization strategy for the Quantum Likelihood Estimation (QLE) algorithm that dynamically selects experimental parameters to maximize mutual information, thereby significantly accelerating Hamiltonian learning in the NISQ era.

Original authors: Alon Levi, Ziv Ossi, Eliahu Cohen, Amit Te'eni

Published 2026-09-03

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Alon Levi, Ziv Ossi, Eliahu Cohen, Amit Te'eni

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 authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the near future, computers may begin to harness the strange rules of quantum mechanics to solve problems that are currently impossible for even the most powerful supercomputers. These machines do not just calculate faster; they process information in a way that allows them to explore many possibilities at once. However, today's quantum computers are still fragile and prone to errors, a stage of development scientists call the noisy intermediate-scale era. Because these machines are imperfect, researchers often use a hybrid approach, pairing the quantum computer with a classical one. In this partnership, the quantum device performs a specific, delicate task to gather raw data, while the classical computer analyzes that data and decides how to adjust the quantum machine for the next step. The goal is to teach the quantum system to learn about its own environment, specifically to figure out the hidden rules, or Hamiltonian, that govern how a quantum system changes over time. Knowing these rules is essential for building better quantum sensors, simulating new materials, and understanding the fundamental behavior of matter.

The challenge lies in how efficiently the system learns. If the quantum computer is asked the wrong questions or measured in the wrong way, it gathers very little useful information, forcing the researchers to repeat the process many times. This is where a new study by Alon Levi, Ziv Ossi, Eliahu Cohen, and Amit Te'eni at Bar-Ilan University offers a significant improvement. The team focused on a specific hybrid method called Quantum Likelihood Estimation, which is designed to identify the correct Hamiltonian from a list of candidates. While the original method works, it can be slow and inefficient because it relies on fixed settings or simple guesses for how to set up each experiment. The researchers proposed a smarter way to run these experiments by treating each step as an opportunity to extract the maximum possible amount of information.

To achieve this, the team developed a strategy that dynamically chooses the best conditions for every single round of the experiment. Instead of sticking to a pre-set plan, their algorithm constantly asks: "What initial state should we prepare, how long should we let the system evolve, and how should we measure it to learn the most about the hidden rule?" They answered this by using a concept from information theory called mutual information, which measures how much knowing the result of a measurement tells us about the unknown Hamiltonian. By maximizing this value, the algorithm ensures that every measurement provides the clearest possible clue. To find the perfect settings for these variables, the researchers used a computational technique called simulated annealing. This method acts like a careful search that explores many different combinations of settings, occasionally accepting a worse option to avoid getting stuck in a local trap, until it finds the global best configuration.

The results of their simulations were striking. When the team tested their optimized approach against the standard version of the algorithm using a set of four simple quantum rules, the improvement was dramatic. The original method, which used a fixed, static setup, required an average of 144 rounds of measurement to confidently identify the correct rule. In contrast, the new, optimized method needed only 9 rounds to reach the same level of certainty. This reduction means the system learns more than fifteen times faster, drastically cutting down the time and resources required. The advantage became even more apparent when the researchers demanded a higher level of confidence in the answer. Furthermore, they tested the method on a more complex set of six different rules, some of which were so similar that the original method failed to distinguish between them entirely. The optimized algorithm successfully identified all six, averaging just four to five rounds per rule.

The study suggests that the key to this speedup was not simply having access to a wider range of settings, but rather the strategy used to select them. To prove this, the researchers compared their dynamic optimization against a version that simply searched through a wide grid of options at every step without the smart annealing process. Even with the wider search, the grid-based method took ten rounds to converge, whereas the smart optimization took nine. This indicates that the intelligence of the selection process itself is the primary driver of efficiency. The researchers note that while their work was demonstrated through computer simulations on single-qubit systems, the underlying logic is robust and can be extended to more complex, multi-qubit systems and even continuous ranges of unknown rules. By treating the learning process as a continuous cycle of preparing, measuring, updating, and optimizing, this approach offers a principled path to making hybrid quantum algorithms practical and scalable for real-world applications.

Technical Summary: Optimal Quantum Likelihood Estimation

Problem Statement
Hybrid quantum-classical algorithms are central to the noisy intermediate-scale quantum (NISQ) era, where quantum resources are constrained, and classical post-processing is essential for extracting meaningful results. However, the performance of these algorithms is often highly sensitive to the choice of quantum parameters (initial state, measurement basis, evolution time) at each iteration. Poor parameter choices can lead to inefficient convergence or failure to distinguish between hypotheses.

The paper focuses on Quantum Likelihood Estimation (QLE), a hybrid algorithm designed for Hamiltonian learning—the task of identifying an unknown Hamiltonian $H_{true}$ from a set of candidates. While standard QLE uses Bayesian inference to iteratively update a probability distribution over candidate Hamiltonians based on measurement outcomes, its efficiency varies significantly depending on the parameters selected at each step. The core problem addressed is how to systematically optimize these parameters to maximize the information gained from each quantum measurement, thereby accelerating convergence.

Methodology
The authors propose an information-theoretic optimization framework that treats each iteration of the QLE algorithm as an independent single-query oracle problem. In this formulation:

The unknown Hamiltonian corresponds to the oracle identity ( $F$ ).
The current weight vector of the algorithm represents the prior probability distribution ( $p_f$ ) over the hypotheses.
The goal is to select the quantum circuit parameters (initial state $|\psi_1\rangle$ , evolution time $t$ , and measurement basis rotation $W$ ) to maximize the mutual information $I(F; Y)$ between the true Hamiltonian $F$ and the measurement outcome $Y$ .

The optimization process relies on the following steps:

Cost Function Formulation: Based on the framework established in Te'eni et al. (arXiv:2409.15549), the mutual information is expressed as the difference between the von Neumann entropy of the marginal state and a basis-dependent version of quantum discord. Since the prior entropy $H(F)$ is constant for a given iteration, maximizing mutual information is mathematically equivalent to minimizing the conditional von Neumann entropy $S(\rho_{FY}|Z)$ .
Parameter Space: The algorithm dynamically selects five parameters at each iteration:
- Initial state orientation ( $\alpha, \beta$ ) on the Bloch sphere.
- Evolution time ( $t$ ).
- Measurement basis rotation angles ( $\theta, \phi$ ).
Optimization Routine: A simulated annealing algorithm is employed to minimize the conditional entropy cost function. The routine iteratively perturbs the parameter vector, accepting changes based on a probabilistic criterion that allows escaping local minima, while gradually cooling the "temperature" to converge on an optimal set of parameters.

Key Contributions

Information-Theoretic Optimization: The paper introduces a principled method for optimizing hybrid quantum algorithms by using mutual information maximization as a guiding cost function for parameter selection, moving beyond heuristic approaches like the Particle Guess Heuristic (PGH).
Dynamic Adaptation: Unlike standard QLE where parameters might be fixed or selected via simple heuristics, the proposed method dynamically re-optimizes the initial state, measurement basis, and evolution time at every iteration based on the current posterior distribution.
Generalization Framework: The authors propose a general scheme extending this approach to broader quantum learning problems. This includes continuous families of Hamiltonians (treated as oracle classification problems) and scenarios involving noise, where noise parameters can be incorporated as hidden variables in the information-theoretic framework.

Simulation Results
The authors evaluated their optimized QLE against a baseline version using a discrete grid search for parameters and the standard PGH for evolution time.

Four-Hamiltonian Case: For a set of four Pauli-based Hamiltonians ( $\sigma_x, 2\sigma_x, \sigma_z, 2\sigma_z$ ), the baseline required an average of 144 iterations to converge (with a weight threshold of 0.99). The optimized version required only 9 iterations.
Stricter Threshold: When the convergence threshold was raised to 0.9999, the optimized method maintained a significant advantage over the baseline.
Six-Hamiltonian Case: For a more complex set of six Hamiltonians involving mixed-Pauli combinations and asymmetric shifts, the original QLE failed to converge for all cases due to indistinguishability under fixed strategies. The optimized QLE successfully converged for all six cases in an average of 4–5 iterations.
Ablation Study: A variant using continuous grid search (without simulated annealing) at each iteration achieved convergence in 10 iterations. This suggests that the primary driver of efficiency is the adaptive optimization strategy itself, rather than simply accessing a wider, unconstrained parameter space.

Significance and Claims
The paper claims that its approach offers a practical method for accelerating Hamiltonian learning in quantum systems by significantly reducing the number of oracle queries (iterations) required for convergence. The authors emphasize that this is achieved without compromising inference accuracy.

The work illustrates how information-theoretic principles can enhance the performance of hybrid quantum algorithms in practical settings. The authors are modest regarding the scope, noting that while the method was demonstrated on single-qubit Hamiltonian learning, it naturally extends to multi-qubit systems (though noting the "curse of dimensionality" in parameter space) and continuous parameter spaces. They suggest that for multi-qubit systems, a fully quantum variant using quantum annealing for the optimization subroutine could potentially mitigate scaling issues, though this requires further investigation. The paper positions QLE as a simplified setting to demonstrate the utility of the information-theoretic framework for improving hybrid algorithms generally.

Technical Summary: Optimal Quantum Likelihood Estimation

More like this