Finding useful variational quantum circuits requires more than representational capacity: training must also have access to directions that improve the objective. We study how discrete CNOT routing shapes both capacity and local accessibility. For circuits built from CNOT and arbitrary one-qubit gates, we separate binary routing matrices from continuous rotations to relate global capacity to local tangent accessibility. An exact finite-field formula for the operator Schmidt rank of a CNOT block and cumulative multiblock bounds yield capacity constraints and a Bell-pair fidelity obstruction. Locally, routing propagates generators into a routing-dependent tangent space, so all parameter gradients can vanish even when the ambient objective gradient is nonzero; frame bounds distinguish loss of gradient coverage from poor conditioning. A routing-conditioned identity insertion preserves the current unitary, state, loss, and existing gradients while adding new tangent directions. Candidate-only and incremental projection scores provide local criteria for ranking insertions. Exact-statevector tests show strong short-horizon score-descent correlations on controlled Bell tasks and effective equal-cost selection on connected Heisenberg chains up to twelve qubits, with complementary TFIM selector controls. A broader TFIM routing ensemble exhibits routing-dependent near-stagnation. Together, these results connect representational capacity, local accessibility, and same-point repair through a common discrete routing description. An architecture need not remain fixed during training: task-dependent geometric information can guide the opening of new descent directions while preserving the learned circuit at insertion. This provides a concrete local approach to a broader problem: how useful quantum circuits can be discovered from the information a task supplies.
Comments
55 pages, 8 figures. Code and data: https://github.com/wang-yu-quantum/VQC-routing-geometry
Finding useful variational quantum circuits requires more than representational capacity: training must also have access to directions that improve the objective. We study how discrete CNOT routing shapes both capacity and local accessibility. For circuits built from CNOT and arbitrary one-qubit gates, we separate binary routing matrices from continuous rotations to relate global capacity to local tangent accessibility. An exact finite-field formula for the operator Schmidt rank of a CNOT block and cumulative multiblock bounds yield capacity constraints and a Bell-pair fidelity obstruction. Locally, routing propagates generators into a routing-dependent tangent space, so all parameter gradients can vanish even when the ambient objective gradient is nonzero; frame bounds distinguish loss of gradient coverage from poor conditioning. A routing-conditioned identity insertion preserves the current unitary, state, loss, and existing gradients while adding new tangent directions. Candidate-only and incremental projection scores provide local criteria for ranking insertions. Exact-statevector tests show strong short-horizon score-descent correlations on controlled Bell tasks and effective equal-cost selection on connected Heisenberg chains up to twelve qubits, with complementary TFIM selector controls. A broader TFIM routing ensemble exhibits routing-dependent near-stagnation. Together, these results connect representational capacity, local accessibility, and same-point repair through a common discrete routing description. An architecture need not remain fixed during training: task-dependent geometric information can guide the opening of new descent directions while preserving the learned circuit at insertion. This provides a concrete local approach to a broader problem: how useful quantum circuits can be discovered from the information a task supplies.
Variational quantum algorithms (VQAs) are commonly developed under the assumption that a parameterized circuit corresponds to a single, well-defined physical realization. In practice, this assumption is not satisfied. Differences in compilation, gate ordering, and hardware-level perturbations lead to multiple admissible circuit realizations, each associated with a distinct objective over the same parameter space. The resulting optimization problem is therefore not defined by a single loss function, but by a family of configuration-dependent objectives whose aggregate behaviour must be considered. This study formulates optimization explicitly at the level of such configuration families. A consistent extension of natural-gradient methods is developed in which both the descent direction and the local geometric structure are obtained through averaging across configurations. The resulting update preserves the structure of quantum natural gradient methods while ensuring alignment with the family-defined objective. The approach is evaluated using exact statevector simulations on transverse-field Ising and Heisenberg Hamiltonians under controlled variability. Across all tested regimes, the method produces lower average loss, substantially reduced dispersion across configurations, and improved worst-case performance relative to both configuration-specific optimization and gradient-averaged baselines. These results show that when the objective depends on a set of admissible circuit realizations, optimization must be constructed at the same level to obtain stable and reliable solutions.
Variational Quantum Algorithms (VQAs) are a leading paradigm for near-term quantum computing, yet their training suffers from sensitivity to circuit depth, initialization, and landscape pathologies such as barren plateaus. We study \emph{progressive depth training} (PDT) -- a layerwise curriculum that trains a shallow circuit before appending new layers -- and identify a fundamental obstacle: fixed entangling gates (CNOTs) in hardware-efficient ansätze cause \emph{initialization shock}, an energy spike when new layers are added. We propose \emph{identity-paired progressive depth training} (IP-PDT), which appends forward/inverse block pairs -- each consisting of a standard rotation+CNOT block followed by its reverse -- that compose to the identity at initialization. Because the adjacent CNOT rings cancel, the effective circuit retains only \textit{a single entangling layer} surrounded by \textit{overparameterized local rotations}. We prove a simple \textit{Reachable Set Saturation Theorem}: under this construction the variational manifold expands exactly once (when post-entangler rotations are first introduced) and then \emph{saturates}; all subsequent depth increases provide pure overparameterization of single-qubit unitaries. Despite this saturation, progressive addition of rotation parameters can continue to improve optimization outcomes -- a phenomenon we term \emph{trainability beyond expressibility}. We formalize IP-PDT as a continuation method on nested manifolds, prove monotone energy guarantees under an acceptance rule, and connect energy error to ground-state fidelity through spectral-gap inequalities. A detailed resource analysis shows that IP-PDT achieves lower total gate cost than both baselines by eliminating most CNOT gates.
Abstract Variational quantum circuits are a central paradigm of quantum machine learning and near-term quantum computing, yet their practical optimization is often limited by strongly suppressed parameter gradients. Dynamical Lie algebras provide a rigorous description of the infinitesimal directions accessible to a variational architecture, but dynamical accessibility alone does not determine the gradient signal exposed by an actual circuit parametrization. In this work, we develop a Lie-theoretic description of parameter-gradient structure that separates four successive mechanisms: dynamical accessibility, observable visibility, state-dependent activation, and local parameter realization. The optimization observable determines which accessible directions can influence the objective and how strongly they couple to it, while the quantum state selects which of these response modes are locally activated. This information is then pulled back through the circuit parametrization to recover the actual parameter gradients. The resulting construction yields an exact modal decomposition of the accessible gradient signal, a hierarchy of structural bounds, and a natural description of state averaging. An analytical single-qubit example and controlled numerical experiments validate the successive levels of the hierarchy and show that accessible Lie dimension alone is insufficient to characterize the parameter-gradient scale. The results support a mechanism-based view of variational trainability in which parameter gradients emerge from the interaction between accessible dynamics, observable coupling, quantum-state activation, and circuit parametrization.
The performance of variational quantum algorithms depends in general on the structure of the parametrized quantum circuit, but the most common ansätze are typically based on local couplings. Motivated by the extended connectivity available with neutral atoms and trapped ions, we examine when structured long-range connectivity provides a useful resource, focusing on sparse power-of-two (PWR2) coupling graphs. Using dynamical Lie-algebra analysis, approximate unitary-design diagnostics, and finite-depth measures of expressibility and entanglement, we examine how these geometries enlarge the accessible operator space. This enlarged space alone is not sufficient to ensure trainability of the parameterized circuit for given target problems, and we explore performance across example problems with and without long-range coupling, identifying where sparse coupling graphs are or are not likely to provide an advantage. We also introduce a variational scheme that maps hierarchical long-range Hamiltonians to geometrically local ones that can be optimized with short-range circuits. Together, these results identify circuit geometry and qubit reconfigurability as task-dependent resources for variational algorithms, relevant to ongoing developments in quantum hardware with long-range connectivity.
Quantum operations on pure states can be fully represented by unitary matrices. Variational quantum circuits, also known as quantum neural networks, embed data and trainable parameters into gate-based operations and optimize the parameters via gradient descent. The high cost of training and low fidelity of current quantum devices, however, restricts much of quantum machine learning to classical simulation. For few-qubit problems with large datasets, training the matrix elements directly, as is done with weight matrices in classical neural networks, can be faster than decomposing data and parameters into gates. We propose a method that trains matrices directly while maintaining unitarity through a single regularization term added to the loss function. A second training step, circuit alignment, then recovers a gate-based architecture from the resulting soft-unitary. On a five-qubit supervised classification task with 1000 datapoints, this two-step process produces a trained variational circuit in under four minutes, compared to over two hours for direct circuit training, while achieving lower binary cross-entropy loss. In a second experiment, soft-unitaries are embedded in a hybrid quantum-classical network for a reinforcement learning cartpole task, where the hybrid agent outperforms a purely classical baseline of comparable size.
Quantum machine learning (QML) has become a promising area for real world applications of quantum computers, but near-term methods and their scalability are still important research topics. In this context, we analyze the trainability and controllability of specific Hamming weight preserving variational quantum circuits (VQCs). These circuits use qubit gates that preserve subspaces of the Hilbert space, spanned by basis states with fixed Hamming weight k . In this work, we first design and prove the feasibility of new heuristic data loaders, performing quantum amplitude encoding of (nk) -dimensional vectors by training an n -qubit quantum circuit. These data loaders are obtained using controllability arguments, by checking the Quantum Fisher Information Matrix (QFIM)'s rank. Second, we provide a theoretical justification for the fact that the rank of the QFIM of any VQC state is almost-everywhere constant, which is of separate interest. Lastly, we analyze the trainability of Hamming weight preserving circuits, and show that the variance of the l2 cost function gradient is bounded according to the dimension (nk) of the subspace. This proves conditions of existence/lack of Barren Plateaus for these circuits, and highlights a setting where a recent conjecture on the link between controllability and trainability of variational quantum circuits does not apply.