Quantum Signal Learning Framework

• Quantum Signal Learning is a property-learning framework that infers target characteristics from the probability distribution of phase-space displacements rather than reconstructing full system details.
• It employs models such as the linear bosonic sensor and Bell QSL protocol to achieve efficient estimation with reduced shot noise and post-hoc data reuse.
• The approach extends to data-centric quantum system learning by integrating classical shadows and neural networks for tasks like state and fidelity estimation.

Quantum Signal Learning (QSL) is a learning-theoretic framework for quantum sensing in which the target is not restricted to a low-dimensional parameter vector, but is formulated as a property of the probability distribution of phase-space displacements induced by an uncertain classical environment (Cotler et al., 19 Feb 2026). In a closely related usage, "Quantum System Learning" denotes data-driven inference of states, fidelities, energies, and other physical quantities from measurement-derived signals, classical shadows, and side information; in that formulation, the operational meaning is the same as learning a mapping from experimentally accessible signals to target physical quantities (Du et al., 2023). In both senses, QSL shifts attention from full reconstruction to task-specific property learning, often with post-hoc reuse of a single experimental dataset.

1. Conceptual scope

The defining move in QSL is to treat sensing as property-learning. For a linear bosonic sensor coupled to an unknown classical field, the object of interest is a functional

$$\Psi\!\left(P_H^{(T)}\right) = \int d^{2n}\alpha\,P_H^{(T)}(\alpha)\,\psi(\alpha),$$

where $P_H^{(T)}(\alpha)$ is the distribution of the induced phase-space displacement and $\psi$ is a property kernel (Cotler et al., 19 Feb 2026). The formal task is

QSL$(H,T,\Psi,\epsilon,\delta)$: estimate $\Psi(P_H^{(T)})$ to absolute error $\epsilon$ with success probability at least $1-\delta$, given query access to $\mathcal E_H^{(T)}$.

This differs from standard quantum metrology in the quantum-Fisher-information style. The latter typically assumes a known parametric family, a low-dimensional parameter, and a prespecified risk function. QSL instead addresses uncertain, distributional signals for which one may wish to infer many downstream quantities from the same data record. Within this framework, single-parameter estimation is recovered when $P_H^{(T)}$ is a delta distribution, while detection, hypothesis testing, and matched filtering become special cases of property-learning (Cotler et al., 19 Feb 2026).

The same operational logic appears in data-centric quantum-system settings. There, a training example is built from measurement-derived input $x$ and target $P_H^{(T)}(\alpha)$0, and learning aims at the map $P_H^{(T)}(\alpha)$1 over a family of related quantum systems rather than exhaustive tomography of each instance from scratch (Du et al., 2023). This suggests a broad research program in which "signal" may mean either a classical field transduced into bosonic displacements or a quantum-system measurement record compressed into a learnable representation.

2. Formal model of signal-induced learning tasks

In the classical-field sensing formulation, the starting point is the linear bosonic Hamiltonian

$P_H^{(T)}(\alpha)$2

Because the coupling is linear in quadratures, finite-time evolution produces only a phase-space displacement (Cotler et al., 19 Feb 2026). Writing

$P_H^{(T)}(\alpha)$3

the Heisenberg equation gives

$P_H^{(T)}(\alpha)$4

Hence deterministic evolution is a displacement $P_H^{(T)}(\alpha)$5. If the field is stochastic, the endpoint displacement is random, and the channel becomes

$P_H^{(T)}(\alpha)$6

This reduction is central because it makes QSL a problem of learning properties of $P_H^{(T)}(\alpha)$7, not of reconstructing the full underlying field history. Examples given for $P_H^{(T)}(\alpha)$8 include quadrature covariances, characteristic-function values, Fourier-domain matched-filter outputs, and control-relevant observables such as cavity phase-rotation drift (Cotler et al., 19 Feb 2026). The framework is therefore more general than a single-purpose estimator: the same sensing channel can support many property queries.

3. Bell QSL and sub-vacuum simultaneous readout

The principal protocol proposed for QSL is Bell QSL, which uses two-mode squeezed vacuum (TMSV), passive linear optics, and static homodyne detection (Cotler et al., 19 Feb 2026). For each mode, one prepares

$P_H^{(T)}(\alpha)$9

lets the sensing arm undergo $\psi$0, interferes sensing and idler modes on a $\psi$1 beamsplitter, and measures the commuting EPR quadratures

$\psi$2

The complex Bell outcome is

$\psi$3

A core proposition states that if the channel applies displacement $\psi$4, the Bell record obeys the additive-noise model

$\psi$5

Equivalently, in real quadratures,

$\psi$6

This is the precise sense in which Bell QSL achieves simultaneous two-quadrature readout with shot noise suppressed below the heterodyne or vacuum level. Standard heterodyne gives

$\psi$7

so each real quadrature carries variance $\psi$8, whereas Bell QSL reduces that variance to $\psi$9 (Cotler et al., 19 Feb 2026).

The protocol can be written as a smoothing-deconvolution scheme. If

$(H,T,\Psi,\epsilon,\delta)$0

then Bell data are a Gaussian-smoothed version of the true displacement law. A single record $(H,T,\Psi,\epsilon,\delta)$1 can therefore be reused post hoc to estimate many different properties by choosing different classical recovery maps $(H,T,\Psi,\epsilon,\delta)$2 satisfying $(H,T,\Psi,\epsilon,\delta)$3. This post-hoc reuse is a defining practical feature of QSL (Cotler et al., 19 Feb 2026).

4. Estimators, guarantees, and provable advantage

For Fourier-defined kernels, the characteristic function of the displacement law is estimated by

$(H,T,\Psi,\epsilon,\delta)$4

with $(H,T,\Psi,\epsilon,\delta)$5 (Cotler et al., 19 Feb 2026). Defining

$(H,T,\Psi,\epsilon,\delta)$6

the paper gives the sample-complexity bound

$(H,T,\Psi,\epsilon,\delta)$7

For characteristic-function estimation at $(H,T,\Psi,\epsilon,\delta)$8, this becomes

$(H,T,\Psi,\epsilon,\delta)$9

For polynomial properties, Gaussian deconvolution can be implemented with Hermite polynomials. An important example is covariance learning for electromagnetic signals, where

$\Psi(P_H^{(T)})$0

In the shot-noise-limited regime, the corresponding sample complexity scales as

$\Psi(P_H^{(T)})$1

The strongest lower-bound machinery is the optimal-transport conditioning method. For an experiment $\Psi(P_H^{(T)})$2 and model class $\Psi(P_H^{(T)})$3, the OT ambiguity modulus is

$\Psi(P_H^{(T)})$4

The formal minimax lower bound states

$\Psi(P_H^{(T)})$5

This quantifies when restricted measurements, especially finite-angle homodyne, are ill-conditioned for learning a target property (Cotler et al., 19 Feb 2026).

Three applications organize the paper’s advantage claims. For electromagnetic correlations, Bell QSL estimates quadrature covariance directly, while restricted homodyne becomes blind or ill-conditioned near isotropy. For real-time feedback control of interferometric cavities, Bell QSL estimates an unknown quadrature-rotation angle from the mean Bell outcome $\Psi(P_H^{(T)})$6, with sample complexity scaling as $\Psi(P_H^{(T)})$7. For Fourier-domain matched filtering, the identity

$\Psi(P_H^{(T)})$8

reduces template-bank scoring to characteristic-function estimation. The resulting theorem gives a worst-case exponential separation: any entanglement-free protocol requires

$\Psi(P_H^{(T)})$9

to estimate even one score to constant accuracy, whereas Bell QSL with $\epsilon$0 estimates all $\epsilon$1 template scores with

$\epsilon$2

shots (Cotler et al., 19 Feb 2026).

5. Relation to data-centric Quantum System Learning

A related but broader use of QSL appears in "Quantum System Learning," which is defined as an umbrella for data-driven methods that extract useful information about quantum systems from measurement data without reconstructing a full exponentially large classical description whenever that is unnecessary (Du et al., 2023). In that setting, the supervised dataset is

$\epsilon$3

with examples sampled from an underlying distribution $\epsilon$4. The input $\epsilon$5 contains classical shadows $\epsilon$6 or local shadow-derived features from $\epsilon$7 snapshots, together with easy-to-access side information $\epsilon$8; the label depends on the task, such as $\epsilon$9 in quantum state tomography or $1-\delta$0 in direct fidelity estimation.

The proposed data-centric paradigm, ShadowNet, combines classical shadows with neural networks. The learning objective is standard empirical risk minimization,

$1-\delta$1

but the point is representational rather than merely architectural: classical shadows provide memory-efficient, statistically meaningful training features, while side information such as Hamiltonian or device parameters supplies task-relevant context (Du et al., 2023). The paradigm is trained offline and then applied to unseen systems from the same family without per-instance optimization.

Two concrete tasks were studied. In quantum state tomography, the input is the reconstructed shadow matrix, the network outputs a denoised density matrix constrained by a Cholesky layer,

$1-\delta$2

and performance is evaluated by quantum fidelity and ground-state energy error. In direct fidelity estimation, the feature vector is built from local inverse snapshots plus side information, with dimension scaling linearly in qubit number. The paper reports numerical studies up to $1-\delta$3 qubits for DFE, and an ablation in the $1-\delta$4-qubit GHZ setting shows average test loss about $1-\delta$5 with full features, compared with $1-\delta$6 using only noise parameters and $1-\delta$7 using only shadows (Du et al., 2023). In this usage, QSL emphasizes family-specific generalization and a shadow-based notion of faithfulness rather than metrological advantage.

6. Acronym ambiguity and field-specific meanings

The acronym QSL is heavily overloaded across quantum science, and context is therefore essential. In condensed-matter physics it most often denotes quantum spin liquid. Representative examples in the supplied literature include thermodynamic evidence that $1-\delta$8 ($1-\delta$9) lie close to the Kitaev quantum spin liquid regime, signaled by a two-peak magnetic heat capacity and an entropy shoulder near $\mathcal E_H^{(T)}$0 (Mehlawat et al., 2017); the identification of Ce$\mathcal E_H^{(T)}$1Zr$\mathcal E_H^{(T)}$2O$\mathcal E_H^{(T)}$3 as a strong candidate for an octupolar $\mathcal E_H^{(T)}$4 QSL, with field-induced Anderson-Higgs physics and octupolar spin waves invisible to neutrons but visible thermodynamically (Gao et al., 2022); a Dirac-type nodal spin liquid in the square-lattice $\mathcal E_H^{(T)}$5-$\mathcal E_H^{(T)}$6 Heisenberg model (Nomura et al., 2020); multiple competing $\mathcal E_H^{(T)}$7 QSL regimes in $\mathcal E_H^{(T)}$8-TaS$\mathcal E_H^{(T)}$9 inferred from $P_H^{(T)}$0SR and specific heat (Mañas-Valero et al., 2020); and gapless QSL phases emerging from deconfined-critical settings in frustrated square-lattice antiferromagnets (Liu et al., 2022, Liu et al., 2021).

In quantum dynamics, QSL can instead mean quantum speed limit. In that usage, the term refers to a lower bound on evolution time, and a recent formulation introduced an attainable QSL for finite-dimensional open and closed systems based on a new state distance and a Fourier-projective decomposition of density operators (Mai et al., 22 Jun 2025).

In distributed quantum machine learning, QSL can also denote quantum split learning. That framework partitions a quantum model across clients and server, introduces cross-channel pooling, and was reported to achieve a $P_H^{(T)}$1 higher top-1 accuracy than QFL on MNIST while reducing transmitted feature size through pooling (Yun et al., 2022).

Because of this terminological multiplicity, "Quantum Signal Learning" is most informative when reserved for the property-learning framework of classical-field sensing and adjacent measurement-driven inference tasks, rather than for the unrelated but entrenched meanings attached to the same acronym in condensed matter, quantum dynamics, and distributed learning.