⚛️ quantum physics

Quantum measurement tomography with mini-batch stochastic gradient descent

This paper introduces stochastic gradient descent algorithms for fast and robust quantum measurement tomography that utilize novel parameterization schemes to ensure physically valid POVM reconstructions, demonstrating superior computational efficiency and fidelity compared to state-of-the-art convex optimization methods.

Original authors: Akshay Gaikwad, Manuel Sebastian Torres, Anton Frisk Kockum

Published 2026-02-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Akshay Gaikwad, Manuel Sebastian Torres, Anton Frisk Kockum

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

Imagine you have a mysterious black box in a quantum laboratory. You put different quantum "inputs" (like specific states of light or atoms) into this box, and it spits out "outputs" (measurement results). Your goal is to figure out exactly how the box works inside. In the quantum world, this "how it works" is described by a complex mathematical object called a POVM (Positive Operator-Valued Measure). Think of the POVM as the box's internal instruction manual or its unique fingerprint.

The paper by Gaikwad, Torres, and Kockum is about a new, super-fast way to read that fingerprint.

The Problem: The Old Way Was Too Slow

Traditionally, scientists tried to figure out this fingerprint using a method called Convex Optimization (specifically, using tools like CVX).

The Analogy: Imagine trying to find the perfect recipe for a cake by testing every single possible combination of ingredients in a giant, slow-moving library. You check one combination, write it down, check another, and so on. As the cake gets more complex (more ingredients, or in quantum terms, more "qubits"), the library becomes so huge that you might spend years just looking at the shelves.
The Reality: For small quantum systems, this old method works fine. But as soon as you add a few more qubits (making the system bigger), the computer gets overwhelmed. It takes hours or even days to solve the math, and sometimes it just gives up entirely.

The Solution: The "Mini-Batch" Stochastic Gradient Descent (SGD)

The authors introduce a new method called SGD-QMT. They took inspiration from how modern AI (like the algorithms that recommend videos on your phone) learns.

The Analogy: Instead of reading the entire library to find the best recipe, imagine you are a chef who only tastes a tiny, random sample of ingredients at a time (a "mini-batch").
- You taste a few ingredients.
- You realize, "Hmm, this tastes a bit too salty."
- You make a tiny adjustment to the recipe.
- You taste a different random sample.
- You adjust again.
- You keep doing this, taking small, quick steps based on small samples, rather than one giant, slow step based on everything at once.

This "stochastic" (random sampling) approach allows the computer to learn the recipe incredibly fast. It doesn't need to process all the data at once; it learns by taking thousands of tiny, rapid steps.

Two New "Ways to Walk" (Parameterizations)

The tricky part of quantum mechanics is that the "fingerprint" (POVM) has strict rules: it must be mathematically "positive" (probabilities can't be negative) and "complete" (all probabilities must add up to 100%). If you just guess randomly, you might break these rules.

The authors invented two special "walking paths" to ensure the computer never steps outside the rules:

The Stiefel Manifold (SM): Think of this as a special, curved track where every step you take automatically keeps you balanced and upright. You can't fall off the track because the track itself is designed to keep you valid.
HONEST (Hermitian Operator Normalization via Eigenvalue Scaling): This is like a self-correcting compass. If your recipe starts to look weird (mathematically invalid), this method instantly "rescales" the ingredients to fix it, ensuring the final result is always a valid quantum measurement.

The Results: Speed and Accuracy

The authors tested their new method against the old "library" method on various quantum systems, from simple ones to complex ones with up to six qubits.

Speed: The new method is a rocket ship compared to the old one. For a system with five qubits, the old method took about 15 minutes. The new method did it in seconds (sometimes under 10 seconds). For larger systems where the old method crashed or took forever, the new method finished in about two minutes.
Accuracy: Surprisingly, the new method wasn't just fast; it was also very accurate. In many cases, especially for specific types of measurements (like checking if a photon was detected), the new method found the "perfect recipe" even better than the old method.
The Winner: Among their new tools, the combination of the HONEST path and a specific type of "tasting" logic (called Maximum Likelihood Estimation) was the absolute champion, finding the most accurate results the fastest.

Why This Matters

The paper claims this completes a "trio" of tools. Previously, scientists had fast, smart ways to check the inputs (Quantum State Tomography) and the operations (Quantum Process Tomography), but checking the outputs (Measurement Tomography) was the slow, difficult step.

Now, with this new SGD-QMT tool, scientists can quickly and accurately check all three parts of a quantum experiment. The authors have even made their code available for free on GitHub, so other researchers can use this "fast-forward" button for their own quantum experiments.

In short: They replaced a slow, brute-force method of reverse-engineering quantum detectors with a fast, smart, AI-inspired learning algorithm that learns by taking tiny, random steps, making it possible to analyze complex quantum machines in seconds rather than hours.

Technical Summary: Quantum Measurement Tomography with Mini-Batch Stochastic Gradient Descent

Problem Statement
Quantum Measurement Tomography (QMT) is a critical component of Quantum Characterization, Verification, and Validation (QCVV), aimed at reconstructing the Positive Operator-Valued Measure (POVM) elements $\{\Pi_i\}$ that characterize an unknown measurement apparatus. Unlike Quantum State Tomography (QST) and Quantum Process Tomography (QPT), which have well-established gradient-descent frameworks, QMT has lagged in the adoption of stochastic optimization methods. Existing QMT approaches primarily rely on Constrained Convex Optimization (CCO) using semi-definite programming (e.g., via CVX). These methods face two significant hurdles:

Computational Scalability: The number of parameters scales exponentially with the system size (qubit number), making CCO computationally prohibitive for systems with more than a few qubits.
Robustness: Standard linear inversion methods, often used as a precursor to CCO, are ill-conditioned and highly sensitive to noise, leading to unphysical estimates that degrade subsequent refinement.

While iterative Maximum Likelihood Estimation (iMLE) exists, it often struggles with enforcing physical constraints (positivity and completeness) strictly throughout the optimization process, sometimes requiring post-hoc projections that hinder convergence.

Methodology
The authors introduce a mini-batch Stochastic Gradient Descent (SGD) framework for QMT, completing the trio of gradient-descent-based tomography methods (alongside QST and QPT). The core of the methodology involves reformulating QMT as an iterative function minimization problem. To ensure that the estimated POVM elements remain physically valid (positive semi-definite and complete) at all times during optimization, the authors propose two distinct parameterization schemes:

Stiefel Manifold (SM) Parameterization:
- This approach leverages the geometry of the complex Stiefel manifold $St(k \cdot d, d)$ .
- POVM elements are constructed as $\Pi_i = T_i^\dagger T_i$ , where the matrices $T_i$ are vertically stacked into a single matrix $T_k$ .
- The completeness condition $\sum \Pi_i = I$ is intrinsically enforced by maintaining the orthonormality constraint $T_k^\dagger T_k = I$ .
- Optimization is performed using Vanilla Gradient Descent (VGD) with a retraction procedure (Cayley transform) to keep the parameters on the manifold.
HONEST Parameterization (Hermitian Operator Normalization via Eigenvalue Scaling Technique):
- This approach constructs unnormalized positive semi-definite operators via Cholesky decomposition ( $T_i^\dagger T_i$ ).
- Completeness is enforced via a projective normalization step. The sum of the unnormalized operators $S = \sum T_i^\dagger T_i$ is computed, and its inverse square root $S^{-1/2}$ is derived via eigenvalue scaling (with spectral regularization to handle numerical instabilities).
- The final POVM elements are obtained via "sandwiching": $\Pi_i = S^{-1/2} T_i^\dagger T_i S^{-1/2}$ .
- This scheme is optimized using the momentum-based Adam optimizer.

Loss Functions
The framework evaluates two loss functions to guide the optimization:

Mean Squared Error (MSE): Equivalent to the $L_2$ norm, minimizing the discrepancy between observed and predicted probabilities.
Average Negative Log-Likelihood (MLE): Inspired by Maximum Likelihood Estimation, minimizing the negative log-likelihood of the observed data.

Key Contributions

Algorithmic Completion: The work provides the first dedicated SGD algorithms for QMT, completing the gradient-descent toolkit for all three fundamental components of quantum experiments (states, processes, and measurements).
Physical Constraint Handling: The introduction of SM and HONEST parameterizations ensures that physical constraints are satisfied during the optimization process, eliminating the need for post-hoc projections that can disrupt convergence.
Scalability: The mini-batch approach allows for efficient handling of large datasets and scales significantly better than traditional CCO methods as system dimension increases.
Open Source Implementation: A Python implementation is made publicly available to facilitate adoption and further research.

Results
The authors benchmarked their SGD-QMT algorithms against state-of-the-art CCO methods (using CVX) on both Discrete-Variable (DV) and Continuous-Variable (CV) systems.

Discrete-Variable Systems (up to 6 qubits):
- Computational Efficiency: For systems with 4 or more qubits, SGD-QMT methods achieved speedups of several orders of magnitude compared to CCO-CVX. For example, reconstructing a 5-qubit system with 32 POVM elements took ~1000 seconds with CCO-CVX but only tens of seconds with SGD-QMT. CCO-CVX failed to converge within a reasonable time for 6-qubit systems, whereas SGD-QMT succeeded in approximately two minutes.
- Reconstruction Fidelity: While CCO-CVX offered slightly higher fidelity for small systems (up to 3-4 qubits), SGD-QMT maintained high accuracy (Frobenius error $\sim 10^{-4}$ to $10^{-5}$ ) for larger systems where CCO failed.
- Best Performer: The HONEST-MLE combination (HONEST parameterization with MLE loss) consistently yielded the highest reconstruction accuracy, particularly for rank-1 projective measurements.
Continuous-Variable Systems (Photon Detection and Counting):
- The methods were tested on photon detection (2-outcome) and photon counting (multi-outcome) scenarios with a truncated Hilbert space dimension of 32.
- Photon Counting: This scenario proved computationally demanding, requiring up to 10,000 iterations. HONEST-MLE achieved an average Frobenius norm of $\sim 10^{-4}$ , significantly outperforming other variants. CCO-CVX failed to produce a solution within an hour for these cases.
- Accuracy: Heatmaps of reconstructed operators showed near-perfect recovery of diagonal elements and high accuracy for off-diagonal elements, validating the physical validity of the reconstructions.

Significance and Claims
The paper claims that SGD-QMT offers a versatile, efficient, and robust framework for characterizing quantum measurement devices, particularly as quantum systems scale beyond the capabilities of traditional convex optimization. By completing the "tomography trio" with gradient-descent methods, the authors provide a unified approach for QCVV protocols.

The authors emphasize that while their methods are computationally efficient and robust to noise, they do not claim to replace CCO for small-scale systems where CCO remains faster and slightly more accurate. Instead, the significance lies in enabling QMT for larger, high-dimensional systems (e.g., 5-6 qubits and CV systems) where standard methods become intractable. The work suggests that HONEST-MLE is the most effective configuration, especially for projective measurements, and highlights the potential for future unified optimization problems that combine state, process, and measurement tomography.