Dissipative quantum neural networks were first proposed in 2020 as a direct quantum analogue of the classical feedforward networks that power modern AI. Instead of artificial neurons passing numbers to one another, the quantum version uses quantum systems, or qudits, arranged in input, hidden and output layers, with trainable unitary interactions playing the role of weights and biases. There is a catch, however. Unlike their classical cousins, these networks contain no nonlinear activation functions, the ingredient that gives classical networks their celebrated ability to approximate virtually any continuous function. The Vienna team argues that rather than forcing nonlinearity into the picture, the right move is to define a quantum-specific notion of universality, one based not on functions over data points but on quantum channels, the completely positive and trace-preserving maps that describe every physically allowed transformation of a quantum state.

With that goal in mind, the researchers introduce what they call extended DQNNs. In the conventional design, each layer of the network is initialized in a fixed fiducial state and perceptrons are unitary gates connecting adjacent layers. The team’s key observation is that those fiducial initializations are actually unnecessary bookkeeping: the entire scheme can be reformulated using isometries, the mathematical objects that map a smaller Hilbert space into a larger one. This is not merely a cosmetic change. A unitary perceptron acting on d-dimensional qudits carries d to the fourth power free parameters, while the corresponding isometry carries only d squared times two d minus one; for d equal to four, the parameter count is cut by more than half. Because fewer knobs must be tuned, training becomes computationally lighter, and the isometry viewpoint makes the network’s task conceptually transparent: it is learning the Stinespring representation of a target quantum channel, the canonical way physicists decompose any quantum process into a coupling to an environment.

To make the reformulation practical, the authors derive a one-to-one parametrization of isometries built on an earlier composite parametrization of the unitary group. Every allowed isometry can be written as an ordered product of simple exponential factors, each controlled by a single real parameter, with no redundancy in the counting. That matters because optimization algorithms thrive on well-behaved parameter spaces. With a redundant parametrization, gradient descent can waste effort moving along directions that change nothing; with this scheme, every parameter genuinely reshapes the channel the network implements. The researchers also develop two distinct training protocols. The first, random state training, resembles classical stochastic gradient descent: a set of input quantum states is sampled, the network’s outputs are compared to target outputs, and the isometry parameters are nudged downhill. The second, Choi training, exploits the fact that every quantum channel has a unique Choi state, obtained by sending half of a maximally entangled pair through the network. Comparing Choi states directly removes any dependence on how training states were sampled, allowing cleaner, more objective statements about whether a network can actually learn a given channel.

The architectural upgrade follows from a simple counting argument. A minimal conventional network with one input, one hidden and one output neuron cannot realize a general quantum channel, because the Stinespring isometry of an arbitrary channel cannot in general be factored into the product of the two perceptrons available. The fix is to add ancilla layers: each perceptron now writes not only to the next layer but also to an ancillary neuron that is subsequently traced out, mimicking the environmental degree of freedom in open-system dynamics. With this extension, a network of just three neurons, one input, one ancilla and one output, becomes quantum channel universal, capable of implementing any physically valid transformation from input to output, provided the ancilla and output dimensions match. Every building block of larger extended networks inherits this power, and when such perceptrons act in parallel on multipartite states, combined with intermediate measurements on ancillas, the architecture naturally realizes LOCC channels, the operations central to the theory of entanglement as a resource.

Architecture alone does not decide whether a network trains well; the choice of objective function is equally consequential, and this is where the study delivers its most actionable result. Because the geometry of quantum state space is not unique, there is no single obvious measure of how close the network’s output state is to the target state. The authors survey a menagerie of candidates: the Hilbert-Schmidt distance, the trace distance with its operational meaning as the optimal error probability for distinguishing two states, the Uhlmann-Jozsa fidelity and the Bures distance rooted in a natural Riemannian geometry of quantum states, generalized p-fidelities, the quantum Chernoff bound governing state discrimination, and the quantum relative entropy from asymmetric hypothesis testing. Each carries distinct theoretical virtues, and some, such as the trace distance and relative entropy, satisfy the data-processing inequality that makes them well-behaved under any physical evolution. But theoretical elegance and experimental convenience do not always coincide.

The numerical experiments zoom in on a minimal extended DQNN of three qubits trained with the ADAM optimizer over one thousand iterations, with the diamond distance, a worst-case measure of the difference between two quantum channels, used as an independent benchmark of success. Under Choi training on one hundred randomly sampled target channels, the distance-based objective functions outperformed all fidelities, with the Bures distance and the Hilbert-Schmidt distance leading the pack, reaching mean diamond distances of 3.43 times ten to the minus four and 4.55 times ten to the minus four respectively, errors small enough to count as essentially perfect learning at this scale. Strikingly, the two hypothesis-testing-inspired quantities fared worst, with the quantum relative entropy leaving the network at a mean diamond distance of 0.386. Under random state training, convergence was faster but plateaued around five times ten to the minus two for nearly all objectives, roughly two orders of magnitude worse than the Choi-trained results, a gap the authors interpret as an early sign of barren plateaus, the phenomenon in which gradients vanish exponentially with Hilbert space dimension.

The team then stress-tested the architecture on the Werner channel, a highly symmetric family of quantum processes whose Choi state is the famous Werner state of quantum foundations lore, and which interpolates between entanglement-breaking and strongly non-classical behavior. Learning speed tracked the channel’s parameter: the more positive the parameter, and hence the more entanglement-breaking the channel, the faster the network converged, with all cases between minus 0.7 and 1 reaching diamond distances below 6.7 times ten to the minus three within five hundred iterations. The lone exception was the completely depolarizing channel, where the network struggled through roughly 350 optimization steps before the error suddenly collapsed, hinting at subtle structure in the cost landscape even for a channel that outputs pure noise. The authors suggest this correlation between trainability and channel properties opens a rich vein for future research, including the question of which objective functions are inherently less susceptible to barren plateaus.

What elevates the work beyond elegant theory is its experimental bottom line. The Hilbert-Schmidt distance is readily measurable on a quantum computer using the SWAP test, and because it involves only quantities at most quadratic in the quantum states, it can even be estimated through shadow tomography via randomized measurements rather than full state tomography. That combination of top-tier performance and hardware accessibility leads the authors to recommend it as the default objective function for Choi-trained dissipative quantum networks. As quantum processors graduate from demonstrations to genuinely useful machine learning devices, decisions like this, which cost function to log, which architecture to fabricate, will shape what is practically achievable. By showing that a three-neuron quantum network can learn any quantum channel, and by identifying the cheapest reliable way to teach it, the Vienna group has handed the field a sharper blueprint for building learning machines out of qubits.

Subject of Research: Architecture and cost function optimization for dissipative quantum neural networks

Article Title: The impact of architecture and cost function on dissipative quantum neural networks

Article References: Sutter, T. C., Popp, C., & Hiesmayr, B. C. (2026). The impact of architecture and cost function on dissipative quantum neural networks. Quantum Machine Intelligence, 8(2), Article 115. https://doi.org/10.1007/s42484-026-00432-2

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00432-2

Keywords: quantum machine learning, dissipative quantum neural networks, quantum channels, isometries, Hilbert-Schmidt distance, barren plateaus, quantum computing, Werner channel, Choi state, Stinespring representation, universality, quantum information

At the heart of the framework lies a distinction between the dynamical Lie algebra of a circuit and the instantaneous tangent space realized at a particular parameter configuration. The dynamical Lie algebra, generated by commutators of the circuit’s Hamiltonian generators, captures everything the model could ever reach as parameters vary freely. The tangent space, by contrast, is the set of directions actually available at a given point in parameter space. The authors show that these two objects can diverge in consequential ways: a direction that is relevant to the learning task may be fully present in the dynamical algebra yet entirely absent from the instantaneous tangent space at the configuration where training begins. In such a case, no amount of algebraic expressibility guarantees that gradients will flow toward the task-relevant subspace.

The second key move is a decomposition of the accessible first-order signal into two separate quantities: its total amount and its density across independent directions. This separation matters because a model can possess a large aggregate signal that is spread thinly over many orthogonal directions, or a modest signal concentrated in a few directions that align well with the task. The researchers derive an exact geometric characterization of both quantities and establish restriction and invariance properties that govern how they behave when circuits are extended, restricted, or reparametrized. These results give practitioners precise tools for asking not just whether a circuit can represent a task, but whether the task’s structure is visible to a gradient-based optimizer at first order.

A particularly striking theoretical finding concerns the conditions under which algebraic compatibility acquires a direct gradient-variance interpretation. The authors prove that such an interpretation holds only under isotropic directional sampling, that is, when gradient directions are sampled uniformly over the sphere of accessible directions. Real circuits, however, do not sample isotropically. For actual parametrized circuits, an additional tangent-frame geometry is required to connect the accessible signal to the parameter gradients that an optimizer actually measures. This means that analyses relying on isotropic assumptions can misrepresent the trainability of concrete ansätze, and the paper provides the corrected geometric machinery to avoid such errors.

The technical engine behind these results is an exact representation of the pointwise gradient. For a circuit built from exponentials of local Hamiltonians, the derivative of the cost function with respect to any parameter can be written as a trace involving the commutator of an effective, parameter-dependent generator with the quantum state or the observable. The authors derive this representation carefully in both an output frame, where the state evolves and the observable is fixed, and an input frame, where the state is fixed and the observable is transformed. Crucially, they demonstrate that the two frames are exactly equivalent, and that mixing elements from the two, for example combining an input-frame generator with an untransformed state-observable pair, is generally invalid. This pointwise discipline underpins the entire framework and prevents common algebraic shortcuts from producing wrong conclusions.

Numerical experiments validate the theoretical decomposition and deliver some of the paper’s most consequential messages. The simulations confirm that task-relevant directions can indeed live in the dynamical algebra while being missing from the instantaneous tangent space, and that two ansätze with identical algebraic dimension can exhibit markedly different alignment with a given task. The experiments also reveal that real tangent frames are strongly anisotropic, reinforcing the warning against isotropic reasoning. Together, these results suggest that expressibility measures based solely on Lie algebra dimension, which have become popular in the barren plateau literature, are insufficient predictors of whether a specific circuit will train well on a specific problem.

Perhaps the most surprising empirical result concerns the origin of gradient decay in scaling experiments. The prevailing narrative in the field attributes vanishing gradients to systematic suppression, in which all directions become uniformly flat as circuits grow. The Salamanca team’s data tell a more nuanced story: pooled gradient decay is explained by causal-cone dilution, a mechanism in which the fraction of parameters whose light cones overlap the task-relevant degrees of freedom shrinks, rather than by uniform suppression of comparable active directions. In other words, many directions may remain individually healthy, but the task-relevant signal becomes diluted across an expanding parameter space. This reframing has direct implications for initialization strategies and ansatz design, since interventions that preserve causal-cone overlap may outperform those that merely boost overall expressibility.

The framework also delivers a clean verdict on circuit extensions. When idle extensions, additional gates or qubits that do not interact with the task-relevant structure, are appended to a circuit, the active compatibility and the gradient statistics remain unchanged. This invariance property provides a practical diagnostic: adding structure to a variational model is harmless to first-order trainability exactly when that structure is dynamically idle with respect to the task. Conversely, extensions that entangle new degrees of freedom into the causal cones of task-relevant generators can dilute the signal even when they increase the algebraic dimension of the model, a trade-off that the framework makes explicit and quantifiable.

Importantly, the authors are careful to position their contribution relative to the barren plateau literature rather than against it. The framework isolates a structural component of first-order trainability, but the paper explicitly cautions that compatibility itself should not be identified with a barren-plateau criterion. Barren plateaus concern the scaling of gradient variances with system size under specific distributions, whereas the new quantities characterize task-resolved signal availability and its conversion into gradients at a structural level. The two perspectives are complementary: Lie-algebraic theories of barren plateaus, such as earlier work connecting dynamical Lie algebras to gradient vanishing, address one slice of the trainability problem, while the task-resolved decomposition addresses another. By keeping the two separate, the new framework avoids overclaiming and clarifies exactly what each kind of analysis can and cannot predict.

The broader significance of this work lies in its potential to guide the engineering of variational quantum models with predictable training behavior. As the field of quantum machine learning matures from proof-of-principle demonstrations toward applications in chemistry, optimization, and data classification, the ability to diagnose, before running an optimizer, whether a chosen ansatz exposes its task-relevant structure to gradients becomes increasingly valuable. The Salamanca framework supplies exactly such a diagnostic, grounded in exact gradient representations, restriction and invariance theorems, and numerically validated decompositions. It suggests that the path to trainable quantum models runs not through maximizing expressibility alone, but through aligning the algebraic reach of a circuit, the geometry of its tangent frame, and the causal structure of the task. For a field wrestling with landscapes that vanish into flatness, that alignment principle may prove to be one of the most actionable ideas to emerge from the Lie-algebraic program so far.

Subject of Research: Lie-algebraic analysis of trainability and gradient geometry in variational quantum machine learning models

Article Title: Task-resolved algebraic compatibility and gradient geometry in variational quantum models

Article References: Román-Gallego, J.-Á., & Casado, R. (2026). Task-resolved algebraic compatibility and gradient geometry in variational quantum models. Quantum Machine Intelligence, 8(2), Article 112. https://doi.org/10.1007/s42484-026-00450-0

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00450-0

Keywords: variational quantum algorithms, quantum machine learning, trainability, dynamical Lie algebra, tangent space, barren plateaus, gradient variance, causal cone, ansatz design, quantum circuits, optimization, parametrized quantum circuits

The stakes are considerable. Theories that extend the Standard Model, such as supersymmetry and the left-right symmetric model, predict the existence of a heavier cousin of the W boson, dubbed the W-prime boson. If such a particle exists, it would be far too short-lived to observe directly, but it would decay promptly into a top quark and a bottom quark. Catching that decay means identifying highly energetic, or boosted, top-quark jets amid an overwhelming background of jets produced by quantum chromodynamics, the theory of the strong interaction. Previous searches, including early work by the D0 Collaboration in 1995 and a 2017 analysis by the CMS Collaboration, have pushed the lower bound on the mass of a right-handed W-prime boson to 2.4 teraelectronvolts at 95 percent confidence, but the hunt continues at the proposed High Luminosity LHC.

The difficulty lies in the geometry of boosted decays. A top quark is unstable and decays into a b-quark and a W-boson, and when the top quark carries enormous transverse momentum, the decay products are squeezed into a narrow angular cone. Physicists visualize this by building a two-dimensional histogram, called a jet image, in which pixel intensities represent the energy fractions of the detected particles. In a boosted top decay, the subjets from the b-quark and the W decay products can overlap so completely that the resulting image looks strikingly similar to a background QCD jet. Classical convolutional neural networks can achieve accuracies between 90 and 94 percent on such tasks, but the benchmark studies show this often comes at the price of trainable parameter counts ranging from one thousand to 1.46 million, as in the state-of-the-art ResNeXt architecture, with the associated computational cost and overfitting risks.

The new work, led by Hala Elhag of the Deutsches Elektronen-Synchrotron and Humboldt-Universität zu Berlin, together with colleagues at institutions including Northeastern University, the Cyprus Institute, the University of Tokyo and the Physikalisch-Technische Bundesanstalt, tackles this bottleneck with a quantum convolutional neural network, or QCNN. First introduced by Cong, Choi and Lukin in 2019, the QCNN replaces the classical filters and pooling operations of a conventional CNN with parametrized quantum circuits. Because it relies on shallow circuits, the architecture is naturally suited to today’s noisy intermediate-scale quantum devices, and theoretical work has shown that its shallow depth also protects it from the barren plateau problem that cripples the trainability of many deeper quantum models.

Before any quantum processing can happen, the classical jet images must be encoded into quantum states, and the choice of encoding turns out to matter enormously. The team tested four schemes: tensor product encoding, also known as angle encoding; one- and two-layer versions of a hardware-efficient encoding; and a classically hard embedding. Each encoded state is then passed through alternating convolutional and pooling layers built from two-qubit unitary gates. The researchers compared two convolutional circuits, the SO(4) circuit, which performs real-valued transformations with six trainable parameters per two-qubit block, and the more general SU(4) circuit, a universal two-qubit gate set with fifteen trainable parameters. The layers repeat until only a single qubit remains, whose Pauli-Z expectation value serves as the network’s prediction.

The dataset came from the JetNet library, a Python package built on PyTorch that provides open Monte Carlo datasets for machine learning in high-energy physics. The researchers used the TopTagging dataset, containing hadronic top jets as signal and QCD jets as background. Each jet’s particles carry four-momentum components, and the team preprocessed the data using a Lorentz boost fixed at a gamma factor of ten, followed by a Gram-Schmidt orthonormalization that constructs the image coordinates from the three highest-momentum constituents. This transformation makes jet images look similar regardless of the original top-quark boost, easing the learning problem. Principal component analysis then compressed the 28-by-28 pixel images down to just four pixels, preserving 50.54 percent of the dataset’s variance, so that each pixel could be loaded onto a single qubit.

Running everything on a noiseless classical simulator using the PennyLane framework, with the classical CNNs built in TensorFlow, the team varied the loss function, encoding and batch size while carefully matching parameter counts between quantum and classical models. The SO(4) QCNN with 30 parameters faced a CNN with 33; the SU(4) QCNN with 48 parameters faced a CNN with 51. Across 50 averaged training runs of 30 epochs each, the quantum model outperformed its classical counterpart in most configurations, particularly with the SO(4) circuit, the simpler one-layer hardware-efficient encoding, and small batch sizes. Accuracy declined as batch size grew from 16 to 128, and the QCNNs also converged faster in the early epochs. The researchers attribute the poor showing of the more complex encodings to dataset-induced barren plateaus, a trainability pathology linked to data encoding complexity.

The most striking result came from dimensional expressivity analysis, a technique that identifies redundant parameters in a parametric quantum circuit by checking whether the derivative of the circuit’s output with respect to each parameter can be written as a linear combination of the others. Applying this analysis to the SU(4) circuit revealed that 17 of its 48 parameters were redundant, leaving a minimal, maximally expressive circuit of just 31 parameters. Trained with mean-square-error loss and the one-layer hardware-efficient encoding over 1000 runs, this trimmed quantum circuit achieved higher accuracy than a comparable 33-parameter CNN. When the classical side was beefed up with additional dense layers, pushing the CNN to 73 and then 91 parameters, the 31-parameter quantum circuit still beat the 73-parameter model and matched the performance of the largest classical network.

The authors are careful about what these results do and do not prove. Because everything ran on a noiseless simulator, the effects of hardware noise, finite-shot sampling and circuit depth remain untested, and they explicitly caution that the findings should not be read as a demonstration of practical quantum advantage. They also note that QCNNs processing classical data have been shown to be effectively classically simulable, which paradoxically is an asset: it allows researchers to scale up simulations to many more qubits and work directly with full-resolution jet images, bypassing the lossy PCA step. Accuracy comparisons with other published top-taggers are likewise not straightforward, since results depend heavily on dataset and preprocessing choices, and real experiments must classify single jets rather than composite images.

Even with those caveats, the study points to a concrete path forward. Reducing parameter counts matters enormously for eventual deployment on noisy quantum hardware, where every additional gate invites error, and the expressivity-optimized circuit shows that leaner quantum models need not sacrifice accuracy. The team also highlights equivariant quantum neural networks, which build the rotational symmetries of the data directly into the architecture, as a natural next step. And beyond classical jet images, quantum machine learning may eventually prove most powerful on genuinely quantum data, such as the spin entanglement among final-state particles that the ATLAS Collaboration recently observed with top quarks. For now, the message is that in the demanding arena of particle physics classification, carefully designed quantum circuits are already competitive, and the design principles that get them there may matter more than raw qubit counts.

Subject of Research: Quantum convolutional neural networks for top-quark jet image classification in high-energy physics

Article Title: Quantum convolutional neural networks for jet images classification

Article References: Elhag, H., Hartung, T., Jansen, K., Nagano, L., Pirina, G., & Di Tucci, A. (2026). Quantum convolutional neural networks for jet images classification. Quantum Machine Intelligence, 8(2), Article 109. https://doi.org/10.1007/s42484-026-00457-7

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00457-7

Keywords: quantum machine learning, quantum convolutional neural networks, top-quark tagging, jet images, high-energy physics, Large Hadron Collider, dimensional expressivity analysis, principal component analysis, data encoding, barren plateaus, beyond the Standard Model, classical simulation

The research team, led by Christopher P. Fulton of the United States Air Force Test Pilot School alongside Irene Tsapara of National University and Lawrence V. Fulton of Boston College, designed a controlled comparison in which real-valued, quaternion-valued, and quantum classification heads all operated on identical frozen feature representations. By holding the upstream representation fixed across every model family, the study isolates the contribution of the classifier itself, stripping away the confounds that have plagued earlier comparisons between quantum and classical approaches. The benchmarks spanned MNIST, FashionMNIST, and CIFAR-10, with the CIFAR-10 experiments run under two distinct feature regimes — a learned 16-dimensional bottleneck and frozen ImageNet-pretrained ResNet18 features — to separate architectural effects from representation quality.

The core idea behind the comparison is mathematically elegant. Quaternion-valued neural networks and variational quantum circuits both derive their local transformations from SU(2) geometry, the group of two-by-two unitary matrices with determinant one that describes rotations in three-dimensional space and, in quantum mechanics, the evolution of single qubits. A quaternion encodes such a rotation with four real numbers, and layers built from quaternion multiplications can be interpreted as chains of these same local rotations that quantum gates perform. If shallow quantum circuits draw their expressive power from this shared geometry, the reasoning goes, then classical quaternion networks should be able to replicate that power without the overhead of state-vector simulation or real quantum hardware.

That is precisely what the experiments found. On MNIST and FashionMNIST, quaternion networks achieved near-equivalence with real-valued multilayer perceptrons, while product-state variational quantum circuits — circuits whose qubits remain unentangled throughout — exhibited both lower accuracy and substantially higher computational cost. The gap was not marginal. Post-hoc statistical analysis of the five-seed MNIST evaluation using a Friedman test provided strong evidence of a non-random model ordering, with a chi-squared statistic of 12.796 and a p-value of 0.0051. Wilcoxon signed-rank tests comparing QuatNet against every quantum model yielded effect sizes exceeding five, a magnitude that signals an overwhelming practical difference rather than statistical noise.

On the harder CIFAR-10 benchmark, quaternion networks retained 94 to 97 percent of real-valued performance across both feature regimes and remained remarkably stable when the feature dimensionality was increased thirty-two-fold. That stability matters for practitioners: a classifier that degrades gracefully as representations grow is far more deployable than one that requires careful retuning at every scale. The product-state quantum circuits, by contrast, underperformed the quaternion classifiers across every benchmark tested, suggesting that whatever advantage shallow SU(2) rotations confer, they can be captured just as effectively — and far more cheaply — by classical quaternion arithmetic.

Perhaps the most provocative finding concerns entanglement, the resource most often cited as the source of quantum advantage. In this study, entanglement provided only modest gains on the grayscale datasets, MNIST and FashionMNIST, and the effect actually reversed under pretrained CNN features, where the entangling circuit suffered a 9.25 percentage-point degradation relative to the product-state circuit. For a field that has long treated entanglement as a proxy for expressive power, the result is a pointed reminder that more quantum structure does not automatically translate into better learning, particularly when the underlying data carries no intrinsic quantum signature.

The team also examined whether quantum-inspired optimization could rescue the circuits’ performance. Fubini–Study and quantum Fisher information natural-gradient methods — sophisticated techniques that exploit the geometry of quantum state space to steer training — did improve geometric alignment, but they did not materially improve short-horizon loss reduction relative to the standard Adam optimizer. In other words, even when the quantum models were trained the ‘right’ way, according to their own native geometry, they failed to close the accuracy gap with their classical quaternion competitors.

For the FashionMNIST and CIFAR-10 evaluations, where only three random seeds were used, the authors relied on large effect sizes — all exceeding 2.0 — as the primary inferential statistic, a pragmatic choice that acknowledges the expense of quantum circuit simulation while still quantifying the magnitude of the differences observed. The pattern held across all datasets and all feature regimes: quaternion networks matched or approached real-valued baselines, and quantum circuits lagged behind both.

The authors are careful to bound their conclusions. The findings apply specifically to shallow, measurement-limited variational circuits operating on classical image-classification tasks without intrinsic quantum structure. They do not rule out quantum advantage on problems with genuine quantum data, deeper circuits beyond current hardware capabilities, or feature spaces engineered to encode quantum correlations. But within the regime studied, the message is clear: shared local SU(2) geometry and shallow entanglement are not sufficient to confer practical quantum advantage on classical vision tasks.

The implications ripple well beyond the benchmarks. As research groups worldwide invest in hybrid quantum-classical pipelines, this study offers a disciplined template for asking whether the quantum component is pulling its weight — and a warning that classical models exploiting the same mathematical structure may be waiting in the wings. Quaternions, discovered by William Rowan Hamilton in 1843 and long confined to aerospace rotation sequences and computer graphics, may turn out to be the quiet classical workhorses that quantum machine learning must first outrun. On the evidence presented here, that race has not even begun.

Subject of Research: A controlled comparison of quaternion-valued neural networks and shallow variational quantum circuits on classical image classification benchmarks

Article Title: Classical (\textrm{SU}(2)) models match or exceed shallow variational quantum circuits on vision benchmarks

Article References: Fulton, C. P., Tsapara, I., & Fulton, L. V. (2026). Classical $$\textrm{SU}(2)$$ models match or exceed shallow variational quantum circuits on vision benchmarks. Quantum Machine Intelligence, 8(2), Article 93. https://doi.org/10.1007/s42484-026-00430-4

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00430-4

Keywords: quaternion neural networks, variational quantum circuits, SU(2) geometry, quantum machine learning, CIFAR-10, MNIST, FashionMNIST, entanglement, natural gradient, barren plateaus, hybrid quantum-classical computing, image classification

The review, titled From graphs to qubits: a critical review of quantum graph neural networks, surveys the emerging field of Quantum Graph Neural Networks, or QGNNs, architectures that fuse the relational power of graph neural networks with the principles of quantum computation. Quantum computers manipulate qubits, which unlike classical bits can exist in superpositions of zero and one, and can become entangled so that the state of one qubit cannot be described independently of another. An n-qubit register lives in a Hilbert space spanned by all 2-to-the-n possible bit strings, an exponentially large arena that quantum algorithms can, in principle, exploit. The authors argue that this richness could offer graph learning a fundamentally different feature map, one capable of encoding complex topological relationships in ways that are hard for classical methods to reach.

But the quantum path is constrained by reality. Today’s machines are Noisy Intermediate-Scale Quantum devices, a term coined by John Preskill to describe processors with limited qubit counts, shallow circuit depths and pervasive noise. The dominant pragmatic approach on such hardware is the variational quantum circuit, in which a parameterized quantum circuit encodes data, evolves it under trainable rotations and entangling gates, and is measured repeatedly, with a classical optimizer updating the parameters in an iterative loop. The choice of data encoding matters enormously: angle encoding maps each feature to a rotation angle and is hardware-friendly but requires operations proportional to the number of features, while amplitude encoding compresses a d-dimensional vector into only log d qubits, yet preparing an arbitrary amplitude-encoded state can still cost O(d) operations. The review stresses that qubit efficiency does not automatically translate into end-to-end speedup, because state preparation, measurement shots and classical optimization all consume the budget.

To bring order to a fragmented literature, the authors propose a three-way taxonomy. Fully Quantum GNNs perform every processing stage in the quantum domain, encoding graph structure directly into Hamiltonian dynamics; they are conceptually elegant but severely limited by noise and qubit scarcity. Hybrid Quantum-GNNs embed quantum operations inside the core learning mechanism itself, implementing message passing, aggregation or graph convolution through parameterized circuits, while classical layers handle the rest. Quantum-Assisted GNNs keep the graph network entirely classical and use quantum modules only externally, for preprocessing, feature transformation or downstream classification. The distinction, the authors emphasize, is functional rather than merely architectural: the key question is not whether a quantum circuit is present, but whether it participates in the graph-learning operation or merely assists it.

The field’s founding idea came in 2019, when Verdon and colleagues introduced QGNNs inspired by the Quantum Approximate Optimization Algorithm. Their general ansatz applies a sequence of parameterized Hamiltonian evolutions whose interaction topology mirrors the problem graph, with each node of the graph associated with a quantum subsystem. From this seed, the review traces several branches: quantum recurrent GNNs that tie parameters across time steps to model temporal dependencies, quantum convolutional GNNs that enforce permutation invariance and globally shared Hamiltonian parameters, quantum time-series convolutional models that use the Schrödinger equation to capture periodic temporal dynamics, and equivariant quantum graph circuits that preserve symmetry under node permutation. One notable construction, the Equivariantly Diagonalizable Unitary circuit, can approximate any real-valued function on bounded graphs and passes the 1-Weisfeiler-Lehman test, outperforming classical message-passing networks in expressive power, at least in theory.

The applications surveyed span strikingly diverse territory. In high-energy physics, hybrid quantum-classical networks have been applied to jet tagging and particle track reconstruction at the Large Hadron Collider, where the upcoming High-Luminosity upgrade demands faster processing of sparse, high-rate collision data. One quantum jet-discrimination architecture achieves a complexity of O(N) in the number of particles, a polynomial speedup over the O(N squared) scaling of classical models, alongside more stable multiclass training, though its raw accuracy remains comparable to classical baselines. In molecular chemistry and biology, QGNNs have predicted molecular energies, HOMO-LUMO gaps and perovskite formation energies; a nine-qubit model for water molecules exploits the geometry of the problem, while an ego-graph decomposition strategy achieved competitive graph classification results using only 1.68 percent of the parameters of its classical counterparts.

In complex systems, the picture is similarly mixed. A temporal-spatial quantum graph convolutional network for traffic congestion prediction, built on a Schrödinger-based temporal model, proved robust but did not beat classical predictors. In finance, a compact QGNN with six qubits and roughly 200 parameters reached 94.5 percent accuracy on credit card fraud detection against 92.4 percent for a classical GraphSAGE baseline, a modest but real gain. Perhaps the most striking result comes from Earth science: a quantum-assisted model for predicting the Oceanic Niño Index, which tracks El Niño, improved accuracy over state-of-the-art classical forecasts while cutting training time by an order of magnitude, converging in five epochs instead of fifty. The review also highlights quantum-native tasks, such as learning Ising Hamiltonian dynamics, preparing GHZ entangled states for quantum sensing, spectral clustering and graph isomorphism testing, where the correspondence between graph structure and quantum interactions is direct and the fit is most natural.

Crucially, the authors introduce a disciplined vocabulary that the field has often lacked. They reserve quantum advantage for cases where a quantum model demonstrably outperforms the best classical counterpart under a clearly specified computational model, accounting for the full pipeline including encoding, state preparation, circuit evaluations, measurement shots and classical preprocessing. Quantum utility describes practically relevant benefits, such as improved accuracy, reduced parameter counts or better trainability, that fall short of formal advantage. Quantum-inspired improvement covers classical methods that borrow quantum concepts without using quantum hardware. Judged by this standard, most current QGNN results demonstrate task-dependent quantum utility rather than definitive quantum advantage, and the review says so plainly.

The obstacles are formidable. Noise and decoherence degrade fragile quantum states, and correlated errors such as crosstalk and non-Markovian noise complicate optimization, introducing systematic bias into objective evaluations. Barren plateaus, regions of the cost-function landscape where gradient variance decays exponentially with qubit count, can stall training entirely, and the problem worsens with noise and with global cost functions. Scalability is perhaps the deepest concern: direct node-to-qubit encodings require at least O(|V|) qubits, edge-dependent interactions may demand O(|E|) entangling gates per layer, and dense graphs can push this to O(|V| squared), before hardware routing adds SWAP gates on connectivity-limited devices. The review also notes that the vast majority of published QGNN studies rely exclusively on classical simulation of quantum circuits, which cannot reproduce real hardware noise, and that initialization strategies for quantum parameters remain underexplored despite their demonstrated impact on convergence.

The authors’ conclusion is neither dismissive nor triumphant. QGNNs, they find, are viable and sometimes competitive alternatives to classical graph networks, particularly in parameter efficiency, training behavior and problem-specific complexity reduction, and they are most naturally suited to graph-structured quantum problems rather than generic large-scale classical graph learning. They call for hardware-aware ansatz design, efficient graph-to-circuit mappings, standardized benchmarks that report qubit counts, compiled circuit depth, shot counts and optimization costs, and greater use of noise-aware simulation and real-device experiments. They also point to QAOA-inspired designs, which encode graph structure directly into the circuit, and to extensions toward hypergraphs and simplicial complexes as promising directions. Until fault-tolerant quantum hardware arrives, the honest verdict is that quantum graph neural networks offer genuine, measurable utility today, while the decisive quantum advantage that would transform graph learning at scale remains an open and rigorously framed research question.

Subject of Research: A critical review of quantum graph neural networks, their architectures, applications, and the gap between quantum utility and proven quantum advantage.

Article Title: From graphs to qubits: a critical review of quantum graph neural networks

Article References: Ceschini, A., Mauro, F., De Falco, F., Sebastianelli, A., Verdone, A., Rosato, A., Le Saux, B., Panella, M., Gamba, P., & Ullo, S. L. (2026). From graphs to qubits: a critical review of quantum graph neural networks. Neural Computing and Applications, 38(17), Article 731. https://doi.org/10.1007/s00521-026-12428-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12428-x

Keywords: quantum computing, graph neural networks, quantum graph neural networks, variational quantum circuits, NISQ devices, barren plateaus, quantum machine learning, high-energy physics, molecular chemistry, fraud detection, El Nino prediction, quantum advantage

Researchers Ernesto Acosta and Carlos Cano of the University of Granada, together with Guillermo Botella of Complutense University of Madrid, have reformulated the parameter-optimization problem of variational quantum algorithms as a Quadratic Unconstrained Binary Optimization problem, the native language of quantum annealers. Their work, published as volume 8, article 101 of the journal, demonstrates that the same machines designed to find low-energy configurations of combinatorial optimization problems can also serve as scalable optimizers for gate-based quantum models, opening a hybrid pathway that combines gate-model quantum computing, quantum annealing, and classical control in a single training loop.

The core insight behind the method is elegant. Variational quantum algorithms depend on a Hamiltonian expressed in terms of trainable rotation angles, and the exponential structure of the circuit’s unitary operator can be transformed, through a heuristic substitution that replaces imaginary phases with real-valued angles, into sums of real exponential terms. Those terms map naturally onto the quadratic binary cost functions that quantum annealers minimize. In other words, the mathematical structure of the ansatz itself provides the bridge: instead of estimating gradients with hundreds of circuit shots, the team encodes candidate parameter regions directly into QUBO form and asks the annealer to find the configuration with the lowest energy.

Because continuous rotation angles must be represented with discrete binary variables, the researchers developed a recursive refinement strategy that progressively narrows the search space. The parameter range is partitioned into segments, validation points are sampled within each partition, and the QUBO problem is solved to identify the most promising regions. The angle range is then rescaled — in their worked example, halved at each level — and centered on the best solution found so far, and the process repeats. This coarse-to-fine scheme approximates high-quality continuous solutions with a small number of discrete QUBO solves, trading resolution for tractability in a controlled way.

A key strength of the framework is its adaptability. The method exposes a rich set of configurable metaheuristic parameters, including the number of training levels, the number of partitions per angle, and the number of validation points per partition. These knobs allow practitioners to tune the balance between solution quality and computational cost according to the problem at hand and the computational resources available. The authors also integrate an adaptive metaheuristic optimization scheme rather than a fixed search rule, making the training procedure generalizable to arbitrary Hamiltonians rather than being tied to a specific circuit family or ansatz design.

To test the approach, the team benchmarked their adiabatic training scheme on publicly available classification datasets, including the Iris, Heart Disease, and Diabetes datasets from the UCI machine learning repository. The experimental evaluations show that the method achieves accuracy comparable to, and in some configurations better than, established classical and evolutionary optimizers, while significantly reducing computational overhead. The authors’ hyperparameter exploration reveals that the highest performance efficiency — accuracy gained per unit of training time — tends to occur at lower numbers of levels and partitions, meaning that modest configurations already capture most of the benefit before costs escalate.

The implications extend beyond raw benchmark numbers. Gradient-based training on near-term quantum hardware is plagued by noisy gradient estimates, since every partial derivative must be inferred from finite numbers of noisy circuit measurements. Meanwhile, barren plateaus, whose origins have been linked both to deep random circuits and to noise accumulation, can render gradients exponentially vanishing and classical optimizers effectively blind. A gradient-free, annealer-driven optimizer sidesteps both pathologies: it never estimates a derivative, and its search is driven by the energy landscape of the QUBO encoding rather than by local slope information that may not exist in any meaningful sense.

The study builds on a growing body of work connecting annealing-style hardware to machine learning. Previous research has shown that QUBO formulations can be used to train classical machine learning models, that Ising machines can serve as training engines for standard neural networks, and that universal adiabatic quantum computers can be harnessed for neural network training. An earlier preprint by the same team had already explored adiabatic training for variational quantum algorithms; the new journal publication consolidates that line of inquiry into a general, configurable framework with systematic experimental validation, code released through a public repository, and full experimental detail across multiple datasets and hyperparameter regimes.

What emerges is a compelling vision of near-term quantum computing as a genuinely hybrid enterprise. In the architecture the Granada and Madrid researchers describe, a gate-based quantum processor executes the variational circuit, a quantum annealer solves the training problem at each iteration, and a classical system orchestrates the recursive refinement and manages the workflow. Each platform contributes what it does best, and none has to shoulder the full burden alone. As annealing hardware scales and QUBO encodings grow more expressive, this division of labor could become a practical route to training quantum models that classical optimizers struggle to handle.

Challenges certainly remain. The heuristic substitution that enables real-valued QUBO encoding sacrifices strict unitarity, and the discretization of continuous parameters means solution precision depends on the depth of the refinement schedule. The authors’ own data show that pushing training levels and partitions higher does not always pay off, underscoring the need for careful configuration. Still, the demonstration that quantum annealers can train variational quantum algorithms with competitive quality and lower overhead marks a meaningful step. For a field hunting for any advantage in the noisy intermediate-scale era, recruiting one quantum machine to train another may prove to be one of the more inventive entries in the playbook.

Subject of Research: A QUBO-based method that uses quantum annealers to train variational quantum algorithms without gradients.

Article Title: QUBO-based training for VQAs on quantum annealers

Article References: Acosta, E., Botella, G., & Cano, C. (2026). QUBO-based training for VQAs on quantum annealers. Quantum Machine Intelligence, 8(2), Article 101. https://doi.org/10.1007/s42484-026-00441-1

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00441-1

Keywords: quantum computing, variational quantum algorithms, quantum annealing, QUBO, optimization, barren plateaus, quantum machine learning, hybrid quantum computing, metaheuristics, NISQ era, gradient-free training, Hamiltonians