Quantum Extreme Learning Machine (QELM)

• Quantum Extreme Learning Machine (QELM) is a quantum analogue of classical ELM that encodes inputs into quantum states processed by a fixed reservoir, with only the output layer optimized.
• It employs diverse physical platforms—from spin networks to photonic circuits—to map complex quantum dynamics into classical features for tasks like state-property estimation and channel characterization.
• QELMs offer efficient training by confining learning to a single linear readout while enabling extensions for temporal memory and higher-order functional mapping in both classical and quantum applications.

Quantum Extreme Learning Machine (QELM) denotes the quantum analogue of a classical extreme learning machine: an input is encoded into a quantum state, processed by a fixed quantum reservoir or quantum channel, measured to produce a classical feature vector, and mapped to the target by a trainable linear readout. Across the literature, the defining restriction is that only the output layer is optimized—typically by Moore–Penrose or ridge regression—while the reservoir dynamics, internal couplings, and measurement settings remain fixed. This architecture has been used for both classical-data tasks and genuinely quantum tasks such as state-property estimation, tomography, and channel characterization (Mujal et al., 2021, Innocenti et al., 2022).

1. Defining architecture and relation to quantum reservoir computing

A QELM inherits the three-layer organization of a classical extreme learning machine: input encoding, a fixed hidden layer, and a trainable linear readout. In the quantum version, the hidden layer is implemented by a physical quantum substrate such as a spin network, a photonic interferometer, a Gaussian or boson-sampling circuit, or a shallow digital quantum circuit. The output of that substrate is not another quantum state passed to a trainable quantum block; rather, it is a classical vector of measurement outcomes or expectation values that feeds a linear regressor or classifier (Mujal et al., 2021, Joly et al., 16 May 2025).

A recurrent distinction in the literature is the contrast between QELMs and quantum reservoir computers (QRCs). QELMs are described as memoryless in the sense that each output is ordinarily produced from a single injection of the input state, followed by one fixed evolution-and-measurement cycle. QRCs, by contrast, exploit temporal correlations and internal recurrence. Several later works extend QELMs toward temporal tasks by concatenating delayed inputs, by using multiple injections of the same state, or by distributed schemes that emulate higher-order functionals without training the reservoir itself (Innocenti et al., 2022, Kawanabe et al., 25 Feb 2026).

The concrete realization of the reservoir varies considerably. Examples in the literature include disordered transverse-field Ising systems with Hamiltonian

$$H_{\rm res}=\sum_{i<j}J_{ij}\sigma_i^x\sigma_j^x + h\sum_i\sigma_i^z,$$

nearest-neighbor XX chains,

$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$

continuous-variable Gaussian interferometers, Gaussian boson-sampling devices, multimode fibers, and photonic frequency-bin networks. In all of these cases, the common design principle is the same: fixed quantum dynamics supply a nonlinear or high-dimensional feature map, and classical training is confined to a single linear layer (Assil et al., 17 Mar 2026, Lorenzis et al., 8 Sep 2025, Maier et al., 15 Oct 2025).

2. Mathematical formulation and the effective-measurement viewpoint

A general QELM can be written in terms of a channel $\Lambda$, a measurement $\{\mu_b\}$, and a readout matrix $W$. For an input state $\rho$, the measured features are

$p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$

In the Heisenberg picture this becomes

$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$

so the entire QELM can be interpreted as an effective POVM acting directly on the input. The readout is linear,

$y = W\,p(\rho),$

and the least-squares solution for training data $P$ and targets $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$0 is

$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$1

or its ridge-regularized variant. This formalism makes explicit that a single-injection QELM can reproduce expectation values of an observable $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$2 iff

$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$3

so exact retrievability is determined by the span of the effective measurement operators (Innocenti et al., 2022).

A complementary analysis uses the Pauli-transfer-matrix (PTM) representation. For an $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$4-qubit system with Pauli basis $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$5, the encoded state can be expanded as

$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$6

The reservoir channel acts linearly on $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$7 through its PTM, while the encoding determines which nonlinear classical features are available in the first place. If $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$8 denotes the measurement-selection matrix, then

$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$9

This yields an interpretive shift: QELM design can be posed as a decoding problem in which the reservoir and measurement must expose task-relevant components of $\Lambda$0 to a linear regressor (Gross et al., 20 Feb 2026).

Multiple-injection schemes enlarge the retrievable function class from linear observables to polynomial functionals. If the same state $\Lambda$1 is injected $\Lambda$2 times, the final effective measurement acts on $\Lambda$3, and polynomial targets such as purity become accessible. The corresponding symmetric-operator space has real dimension

$\Lambda$4

which also quantifies the resource growth required for arbitrary degree-$\Lambda$5 functionals (Innocenti et al., 2022).

3. Expressivity, scrambling, and fundamental limitations

A central theoretical result is that QELM predictions admit an exact Fourier decomposition whose frequencies are fixed by the encoding. For scalar input and encoding $\Lambda$6, the expectation of a measured observable can be written as

$\Lambda$7

where $\Lambda$8 are eigenvalues of the encoding generator. The reservoir and the measurement affect the coefficients $\Lambda$9, but not the set $\{\mu_b\}$0 itself. The resulting Fourier-expressivity obeys the bound

$\{\mu_b\}$1

so expressivity is jointly limited by the number of observables, the encoding-induced frequencies, and the size of the measured subsystem (Xiong et al., 2023).

This theoretical picture leads directly to the main no-go mechanisms. Four distinct sources of concentration were identified: randomness, hardware noise, entanglement, and global measurements. Each can drive the measured observables toward input-independent values with exponentially small fluctuations, turning the QELM into what the paper calls a useless input-agnostic oracle. In particular, highly random reservoirs drawn from a Haar-like ensemble are singled out as unscalable in this sense (Xiong et al., 2023).

At the same time, later work complicates any simple “less scrambling is always better” conclusion. For XX-chain QELMs used for image classification, accuracy remains low until a critical time $\{\mu_b\}$2, then rises sharply and saturates; the plateau matches the performance of random unitaries, and the critical time is nearly independent of the system size. Because this threshold corresponds only to nearest-neighbor propagation, the required depth is $\{\mu_b\}$3, which implies efficient classical simulation for the cases studied (Lorenzis et al., 8 Sep 2025). A related state-estimation study found that, for the Hamiltonian families examined, reconstruction remains efficient even beyond the scrambling time: the condition number of the training matrix saturates to $\{\mu_b\}$4, and the long-time reconstruction error reaches the same floor as Haar-random global unitaries, with $\{\mu_b\}$5 in the reported $\{\mu_b\}$6-qubit setting (Vetrano et al., 2024).

These results do not remove the concentration barrier; rather, they localize it. A plausible implication is that the relevant control parameter is not “scrambling” in isolation, but the joint design of encoding, reservoir family, measured observables, and shot budget. That interpretation is made explicit in the PTM framework, where optimization is cast as exposing decodable task features, not as maximizing generic quantum complexity (Gross et al., 20 Feb 2026).

4. Temporal memory, delayed embeddings, and distributed architectures

The original memoryless definition of a QELM is restrictive for temporal data and for non-Markovian quantum processes. One extension is the memory-enhanced QELM developed for parameter estimation in a tunable collision model. There, the input is a sequence of reduced system states

$\{\mu_b\}$7

while the reservoir is a disordered many-body system with fixed Hamiltonian. After injecting $\{\mu_b\}$8, local $\{\mu_b\}$9-expectations

$W$0

form the raw feature vector $W$1. Temporal extensions concatenate present and past outputs, for example

$W$2

or use a distant reference $W$3. In that study, temporal extensions consistently and significantly enhanced estimation accuracy relative to the baseline single-time protocol. Immediate-past memory reduced NMSE by up to $W$4–$W$5, and distant-reference memory yielded up to $W$6 reduction, whereas adding extra observables produced only marginal gains. The benefit became more pronounced as the dynamics became more strongly non-Markovian, indicating that environmental memory acts as a constructive learning resource (Assil et al., 17 Mar 2026).

A second route is the time-delayed QELM for classical time-series prediction. Instead of feeding past values sequentially into a recurrent quantum reservoir, the delay vector

$W$7

is encoded in parallel onto $W$8 qubits, followed by one shallow reservoir evolution and one measurement round per time step. In the reported NARMA10 experiments, this reduces the circuit count from $W$9 for the compared QRC protocol to $\rho$0 with depth independent of sequence length. On noiseless simulation, TD-QELM achieved $\rho$1 versus $\rho$2 for QRC, and the same architecture remained more stable on FakeKawasaki and on real ibm_kawasaki hardware (Kawanabe et al., 25 Feb 2026).

Distributed QELMs generalize the same principle spatially. For learning linear functions of quantum states, a standard three-layer QELM requires reservoir dimension at least as large as the input dimension,

$\rho$3

Spatial multiplexing replaces one large reservoir by $\rho$4 independent smaller units coupled in parallel to the same input; the resulting per-unit reservoir-size requirement scales approximately as $\rho$5, while the total number of measurements remains $\rho$6. For nonlinear targets, the literature contrasts multiple-injection architectures—whose resource requirement grows combinatorially with degree—with a distributed entanglement-based design in which interacting subsystems reconstruct higher-order nonlinearities with reduced per-unit resources. Benchmarks on concurrence and negativity estimation show monotonically decreasing NMSE as the number of reservoirs increases from $\rho$7 to $\rho$8 (Gili et al., 12 Feb 2026).

5. Experimental platforms and application domains

QELMs have been instantiated on digital superconducting devices, spin-network simulators, multimode fibers, Gaussian and boson-sampling photonics, orbital-angular-momentum ancillary platforms, and frequency-bin photonic processors. The application range spans both quantum-state property learning and conventional regression or classification tasks. The table summarizes representative results reported in the literature.

Domain	Platform or reservoir	Reported outcome
Werner-state entanglement estimation	5-qubit transverse-Ising reservoir	$\rho$9 noiseless; $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$0 at $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$1; $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$2 near critical $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$3 for high noise (Assil et al., 3 Nov 2025)
Photonic entanglement witnessing	OAM ancilla, single-setting informationally complete measurement	test MSE $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$4, certification accuracy $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$5 (Zia et al., 25 Feb 2025)
Classically trained quantum inference	Frequency-bin biphotons, stimulated/spontaneous correspondence	entanglement witnessing $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$6, Hamiltonian-learning fidelity $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$7, training reduced from $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$8 h to $p_b(\rho)=\mathrm{Tr}[\mu_b\,\Lambda(\rho)].$9 h (Brusaschi et al., 20 Mar 2026)
Molecular PES and force fields	IBM superconducting hardware	IBM_BRISBANE energy RMSE $$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$0 Ha for LiH, $$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$1 for H$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$2O, $$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$3 for HCONH$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$4 (Monaco et al., 2024)
Industrial elevator QoS prediction	Deterministic HE encoder + Ising reservoir	average MSE $$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$5–$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$6 s$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$7 vs. $$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$8–$$p_b(\rho)=\mathrm{Tr}[\tilde\mu_b\,\rho],\qquad \tilde\mu_b=\Lambda^\dagger(\mu_b),$$9 s$y = W\,p(\rho),$0 for classical tree/SVM baselines (Wang et al., 2024)
Collider-data selection	Continuous-variable photonic Gaussian QELM	top-jet accuracy $y = W\,p(\rho),$1, Higgs accuracy $y = W\,p(\rho),$2 at $y = W\,p(\rho),$3 (Maier et al., 15 Oct 2025)

Several additional examples broaden the hardware picture. In the exoplanet-retrieval study, nine disjoint $y = W\,p(\rho),$4-qubit reservoirs were run on IBM Fez with $y = W\,p(\rho),$5 shots per input; the hardware results matched the finite-statistics simulation closely, with $y = W\,p(\rho),$6 accuracy for the planetary radius $y = W\,p(\rho),$7, $y = W\,p(\rho),$8 for CH$y = W\,p(\rho),$9, $P$0 for CO$P$1, $P$2 for H$P$3O, $P$4 for $P$5, and $P$6 for $P$7 (Vetrano et al., 3 Sep 2025). In the multimode-fiber photonic platform, indistinguishable-photon QELMs were linked to an increased effective rank of the feature matrix, and simulations for a 5-class FashionMNIST task showed accuracy increasing from $P$8 to $P$9 for indistinguishable photons at $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$00, versus $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$01 for distinguishable photons (Joly et al., 16 May 2025).

These application studies collectively show that QELMs are not tied to a single notion of “reservoir.” In some settings the reservoir is a disordered many-body Hamiltonian; in others it is a photonic interferometer, a multimode fiber, a shallow gate-based circuit, or even an informationally complete ancillary measurement construction. What remains invariant is the fixed quantum feature map plus analytically trained classical readout (Mujal et al., 2021, Montesinos et al., 13 Jun 2026).

6. Noise, large-scale deployment, and open problems

NISQ performance is a recurrent fault line in the QELM literature. In three industrial software-testing case studies, adding realistic IBM noise models caused severe degradation: regression showed a median MSE increase of $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$02 when noise was present only during inference, and classification accuracy dropped by $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$03. Training under noise improved these figures, but not enough for practical deployment without mitigation. Zero-Noise Extrapolation reduced one regression drop from $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$04 to $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$05, while Q-LEAR yielded a more consistent $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$06 accuracy loss for classification, though it was less effective for regression (Muqeet et al., 2024).

At the same time, newer work shows that QELMs can be pushed to substantially larger scales when concentration and shot-noise issues are explicitly managed. A digital-superconducting implementation reported circuits on up to $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$07 qubits and more than $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$08 two-qubit gates, together with a multi-objective tuning strategy that monitored observable variability, reservoir complexity, and task performance. That study also introduced a local eigentask analysis for feature selection and found an operating regime identified at $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$09–$$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$10 qubits to be transferable to larger systems and to different tasks, including NARMA forecasting and Statlog Landsat classification (Dao et al., 13 Mar 2026).

Photonic studies frame the same issue differently: the relevant scaling parameter is often feature dimensionality rather than circuit depth. In the Gaussian boson-sampling QELM benchmark, photon-number sampling probabilities outperformed alternative feature families, and on MNIST the ideal test accuracy reached $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$11, remaining $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$12 under $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$13 feature noise and $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$14 under $$H = \frac12\sum_{i=1}^{N}\bigl(\sigma_x^{(i)}\sigma_x^{(i+1)}+\sigma_y^{(i)}\sigma_y^{(i+1)}\bigr),$$15 noise (Montesinos et al., 13 Jun 2026). In the XX-chain analysis, by contrast, the observation that optimal accuracy is achieved at constant evolution time implies efficient classical simulability for a broad class of tasks, and therefore no asymptotic quantum advantage in that regime (Lorenzis et al., 8 Sep 2025). This suggests that practical success and asymptotic separation are distinct questions in QELM research.

The main open directions stated across the literature are relatively consistent. They include optimal memory-depth selection as a function of non-Markovianity, multi-parameter simultaneous estimation, hybrid classical–quantum feature mixing, experimental implementations on NISQ processors, distributed architectures for higher-order nonlinear functionals, improved encoding and measurement design, QELM-tailored error mitigation, and a fuller theory of approximation power and information-processing capacity (Assil et al., 17 Mar 2026, Gili et al., 12 Feb 2026, Mujal et al., 2021). A parallel line of work adds interpretability goals: when the QELM is trained on trajectories of nonlinear dynamical systems, the learned model can be recast as a surrogate approximation of the underlying flow map, making the random-feature construction partially analyzable rather than purely black-box (Gross et al., 20 Feb 2026).

In that sense, QELM has developed into a family of fixed-dynamics, linear-readout models rather than a single canonical circuit. Its strongest established advantages are the avoidance of quantum-side nonconvex training, the flexibility of physical realization, and the ability to turn uncalibrated or partially characterized quantum dynamics into usable feature maps. Its strongest established limitations are the span constraint of the effective measurement, concentration phenomena at large scale, and the sensitivity of current hardware implementations to noise unless depth, observables, and hyperparameters are chosen with care (Innocenti et al., 2022, Xiong et al., 2023).