⚛️ quantum physics

Triple descent in quantum linear regression

This paper demonstrates that sampling noise in quantum linear regression induces a unique triple-descent error profile featuring a classical peak due to measurement ill-conditioning and a distinct, irreducible quantum peak caused by shot noise fitting unphysical features, revealing that treating quantum features as exact both overestimates classical instabilities and misses genuine quantum ones.

Original author: Gabriele Lo Monaco, Salvatore Lorenzo, Alessandro Ferraro, Mauro Paternostro, G. Massimo Palma, Luca Innocenti

Published 2026-11-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original author: Gabriele Lo Monaco, Salvatore Lorenzo, Alessandro Ferraro, Mauro Paternostro, G. Massimo Palma, Luca Innocenti

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 emerging field of quantum machine learning, computers attempt to learn patterns from data that exists in the strange, probabilistic realm of quantum mechanics. Instead of reading numbers on a screen, these machines interact with quantum states—fragile configurations of particles—and then measure them to extract information. This process is inherently noisy. Because quantum measurements are probabilistic, a single glance at a quantum system rarely tells the whole story. To get a reliable reading, researchers must repeat the measurement many times, a process known as taking "shots," and then average the results. The more shots taken, the clearer the picture becomes, and the less the random noise interferes. For years, the prevailing wisdom was that this noise was merely a nuisance, a temporary obstacle that could be smoothed away simply by gathering more data. If the computer made a mistake due to a fluke in the measurement, the solution was assumed to be obvious: just measure again.

A team of researchers has now shown that this intuition is incomplete. In a study of quantum linear regression—a method where a machine learns to predict a value based on quantum measurements—they discovered that the noise from these measurements does not just fade away. Instead, it can fundamentally reshape how the learning process behaves, creating unexpected peaks in error that defy standard expectations. The researchers found that as the number of measurement outcomes increases, the prediction error does not simply drop steadily. It rises and falls in a complex pattern with three distinct stages, a phenomenon they call "triple descent." This behavior reveals that the noise itself can sometimes help the learning process, and at other times, it can create deep, unshakeable traps that more data cannot fix.

The study focuses on a specific type of learning problem where a quantum device produces a set of features, which are essentially probabilities of different measurement outcomes. A classical computer then tries to fit a line through these features to predict a target value. The researchers simulated this process using random quantum measurements and training data, varying the number of measurement outcomes and the number of times each state was measured. They observed that the error in the prediction follows a specific path. Initially, as the number of measurement outcomes grows, the error drops. Then, it hits a first peak. This first peak occurs when the measurement setup becomes just complex enough to theoretically capture all the information about the quantum state, a point known as informational completeness. At this precise moment, the mathematical structure of the measurement becomes unstable, causing the error to spike. This instability is a known issue in classical statistics, where having just enough data points to solve a problem can make the solution extremely sensitive to small errors. The researchers confirmed that this first peak is a "classical" effect, driven by the geometry of the measurement itself.

However, the story does not end there. As the researchers continued to add more measurement outcomes, the error did not simply continue to fall. It rose again to a second, distinct peak. This second peak appears when the number of measurement outcomes matches the number of training examples, a point called the interpolation threshold. Unlike the first peak, this one is purely "quantum" in origin. It arises because the finite number of shots used to estimate the probabilities introduces random fluctuations that create fake features. These fake features do not exist in the true quantum state; they are pure noise. When the number of outcomes matches the number of training states, the learning algorithm tries to fit these noise-only features perfectly. It assigns them huge weights, effectively memorizing the random noise rather than the underlying pattern. This creates a second spike in error that is fundamentally different from the first.

The most surprising finding is how these two peaks respond to more data. The first peak, caused by the measurement geometry, can be suppressed by taking more shots during the training phase. If the researchers increased the number of measurements per training state, the noise was averaged out, and the first peak flattened. However, the second peak, the quantum one, behaves differently. Even if the researchers took an enormous number of shots to train the model, this second peak did not disappear. It only vanished in the theoretical limit where the training features were known with perfect, infinite precision. In any real-world scenario where data is finite, this peak remains. The researchers showed that adding more training shots actually makes this peak worse before it eventually stabilizes, because the algorithm becomes more confident in fitting the noise. This means that simply gathering more data is not a cure-all; in fact, in the quantum regime, more data can sometimes entrench the model's mistakes.

The team used advanced mathematical tools, specifically a field of mathematics called random matrix theory, to derive the exact shape of these error curves. They proved that the height and width of these peaks depend on the balance between the number of measurement outcomes, the number of training states, and the number of shots taken. They also showed that a technique called ridge regularization, which adds a small penalty to prevent the model from assigning too much importance to any single feature, can smooth out these peaks. This suggests that while the quantum noise creates a fundamental instability, it can be managed with the right mathematical constraints. The results challenge the idea that quantum features can be treated as if they were perfectly known. If researchers ignore the noise and assume the features are exact, they will underestimate the instability caused by the measurement geometry and completely miss the new, quantum-specific instability that arises from the noise itself.

This work highlights a subtle but critical relationship between measurement precision and learning stability in quantum systems. It shows that the path to accurate predictions is not a straight line. There are specific thresholds where the system becomes fragile, and the nature of that fragility changes depending on whether the measurement setup is just complex enough to be complete or whether the model is trying to fit the noise. The researchers conclude that designing quantum learning systems requires a careful balance. One cannot simply increase the number of outcomes or the number of shots independently. The choice of how many measurement outcomes to use must be coordinated with the size of the training set and the available shot budget. If these factors are not aligned, adding more measurement outcomes can actually make the predictions worse, even as the measurement itself becomes more precise. The study provides a roadmap for understanding these instabilities, offering a way to predict when a quantum learning model will struggle and how to tune it to avoid the pitfalls of the triple-descent landscape.

Technical Summary: Triple Descent in Quantum Linear Regression

Problem Statement
Quantum machine learning (QML) models, such as quantum extreme learning machines and quantum fidelity kernels, typically rely on classical linear regression trained on features extracted from quantum measurements. These features (e.g., outcome probabilities) are never known exactly; they are estimated from a finite number of measurement shots, introducing sampling noise. While standard practice treats this noise as a nuisance that can be suppressed by increasing the number of shots, this paper argues that such a view is incomplete. The authors investigate how finite sampling statistics fundamentally reshape the stability and generalization of quantum linear regression (QLR), specifically examining the prediction error as the number of measurement outcomes ( $n_{out}$ ) and training states ( $n_{tr}$ ) vary.

Methodology
The authors analyze QLR protocols where a quantum device produces features as expectation values of fixed measurement operators (POVMs) on input states. The prediction is a linear combination of these features, with coefficients fitted via classical least-squares regression.

The analysis employs a rigorous framework combining:

Random Matrix Theory (RMT): To characterize the spectral properties of the measurement frame operator and the empirical feature matrix in the high-dimensional limit ( $d \to \infty$ ).
Statistical Decomposition: Separating the mean squared error (MSE) into an "ideal-test error" (bias due to imperfect reconstruction of the target observable) and a "test variance" (sensitivity to finite test-shot noise).
Asymptotic Analysis: Investigating limits where the Hilbert space dimension $d$ , the number of outcomes $n_{out}$ , the number of training states $n_{tr}$ , and the shot budgets ( $N$ for training, $M$ for testing) scale proportionally.
Ridge Regularization: Analyzing the effects of explicit ridge regularization ( $\tau$ ) and comparing it to the implicit regularization induced by finite shot noise.

The study distinguishes between two regimes:

Informational Completeness: Where the number of outcomes $n_{out}$ approaches the dimension of the operator space ( $d^2$ ).
Interpolation: Where the number of outcomes matches the number of training states ( $n_{out} = n_{tr}$ ).

Key Contributions and Results

The central finding is the emergence of a triple-descent profile in the prediction error as $n_{out}$ increases, characterized by two distinct peaks with different physical origins:

The Classical Peak (at Informational Completeness, $n_{out} = d^2$ ):
- Origin: This peak arises from the ill-conditioning of the measurement frame operator itself. As $n_{out}$ approaches $d^2$ , the POVM frame operator develops small eigenvalues, making the reconstruction of certain operator directions highly sensitive to noise.
- Nature: It is a "classical" instability, analogous to the double-descent phenomenon in classical linear regression. It persists even if training features are known exactly.
- Behavior: This peak affects only the variance of the prediction. It can be suppressed by increasing the number of test shots ( $M$ ) or by increasing training shots ( $N$ ), which acts as an effective regularizer by damping the inversion of weakly resolved directions.
The Quantum Peak (at the Interpolation Threshold, $n_{out} = n_{tr} > d^2$ ):
- Origin: This peak occurs when the number of outcomes exceeds the dimension of the operator space but matches the number of training states. Here, finite training-shot noise populates "noise-only" feature directions that do not exist in the exact quantum feature representation (which is limited to $d^2$ dimensions).
- Mechanism: At interpolation, the noise-only block of the empirical feature matrix becomes square and nearly singular. Least-squares regression fits these pure-noise features with large weights to satisfy the training labels.
- Nature: This is a fundamentally "quantum" instability driven by stochastic fluctuations. Unlike the classical peak, it cannot be suppressed by increasing the number of test shots ( $M$ ) because the learned model itself is corrupted by fitting the noise.
- Behavior: The peak height in test variance saturates as the training shot budget ( $N$ ) increases, rather than vanishing. The limit of infinite training shots ( $N \to \infty$ ) does not commute with the limit of exact features ( $N = \infty$ ); the peak only disappears in the theoretical limit of exactly known training features.

Regimes and Interactions

Below Informational Completeness ( $n_{out} < d^2$ ): An interpolation peak ( $n_{out} = n_{tr}$ ) can still occur, but it is "classical" in nature. It is driven by the inability of the measurement to access components of the target observable (inaccessible residuals) rather than by noise-only directions.
Regularization: The authors derive analytical expressions for the shape of both peaks, showing that explicit ridge regularization rounds the singularities. The peak height scales as the inverse square root of the ridge parameter.
Shot Budgets: Increasing training shots ( $N$ ) improves the accuracy of the reconstructed observable (reducing ideal-test error) but can paradoxically increase the variance pole of the quantum peak, as the model becomes more sensitive to the specific noise realization in the training set.

Significance and Claims
The paper claims that treating quantum features as exactly known is a flawed approximation that leads to incorrect predictions about learning stability. Specifically:

It overestimates classical instabilities (by missing the regularizing effect of finite shots on the classical peak).
It misses genuinely quantum instabilities (the quantum peak at interpolation), which arise solely from the interplay between finite sampling and the geometry of the feature space.

The authors conclude that reliable analysis of QML generalization must treat training and test statistics separately and account for the constraints imposed by quantum states and measurements. They assert that measurement design, training-set size, and shot budgets cannot be chosen independently; whether adding more measurement outcomes improves or degrades performance depends critically on how $n_{out}$ compares to both the Hilbert space dimension $d^2$ and the number of training states $n_{tr}$ . The work provides a quantitative framework for understanding these trade-offs using random-matrix theory, offering a more complete picture of the "triple descent" phenomenon in quantum learning than previously available.

More like this