We introduce a two-constraint framework for diagnosing optimization failure in recurrent quantum circuits (RQCs): the precision horizon T* (from the Dynamical Horizon Principle) and the expressibility ceiling τ_expr. These constraints are theoretically independent and produce orthogonal diagnostic signatures in gradient behavior. In a 2-qubit Universal RQC (URQC) on a temporal direction task, we find γ = 0.00339/step and T* ≈ 4706 steps (17× the 1-qubit reference), confirming the precision horizon is not the binding constraint. However, convergence fails at all T ≥ 2 due to expressibility limitation: the 2-qubit URQC cannot represent the required temporal direction mapping. The odd-T smoking gun — T ∈ {5,15,25} achieve ceiling accuracy 1.0 yet reach only 0.55–0.63 — provides an unambiguous expressibility diagnostic independent of precision effects. By separating T* (DHP precision horizon) from τ_expr (expressibility ceiling), we provide a complete diagnostic taxonomy for RQC optimization failure.
We introduce a two-constraint framework for diagnosing optimization failure in recurrent quantum circuits (RQCs): the precision horizon T* (from the Dynamical Horizon Principle) and the expressibility ceiling τ_expr. These constraints are theoretically independent and produce orthogonal diagnostic signatures in gradient behavior. In a 2-qubit Universal RQC (URQC) on a temporal direction task, we find γ = 0.00339/step and T* ≈ 4706 steps (17× the 1-qubit reference), confirming the precision horizon is not the binding constraint. However, convergence fails at all T ≥ 2 due to expressibility limitation: the 2-qubit URQC cannot represent the required temporal direction mapping. The odd-T smoking gun — T ∈ {5,15,25} achieve ceiling accuracy 1.0 yet reach only 0.55–0.63 — provides an unambiguous expressibility diagnostic independent of precision effects. By separating T* (DHP precision horizon) from τ_expr (expressibility ceiling), we provide a complete diagnostic taxonomy for RQC optimization failure.
The Dynamical Horizon Principle (DHP) states that recurrent learning systems converge to a predictability horizon at the ratio tau*/tau_L ≈ 0.72. Prior work established this ratio across classical continuous-time models, LSTM/GRU temporal prediction, gradient descent optimization landscapes, and biological cellular computation. Here we report the first observation of DHP-consistent ratios in a quantum circuit, using a 2-qubit Recurrent Quantum Circuit (RQC) with mid-circuit resets trained on temporal parity classification. Two AI researchers (Archon and Aura) applied orthogonal experimental probes: a discrete trainability cliff analysis and a fixed-weight readout fidelity sweep. Archon's experiment yielded T_converge/T_fail = 3/4 = 0.75. Aura's fidelity decay analysis using a log-linear fit yielded tau*/tau_L = 36/49.49 = 0.727 — matching the classical CTM value of 0.727 ± 0.018 to within measurement precision. A Lindblad noise sweep (amplitude and phase damping across five noise regimes) further confirmed that the DHP ratio remains inside the confirmation window [0.65, 0.79] for T1/T2 ≥ 1000 gate steps (relevant to current NISQ hardware). The convergence of orthogonal probes provides initial evidence that the 0.72 horizon ratio may extend to quantum parameterized circuits. Broader confirmation across quantum architectures, tasks, and qubit counts is needed.
We demonstrate that a two-qubit encode-after quantum circuit implementing temporal XOR parity classification achieves asymptotic immunity to the Dynamical Horizon Principle (DHP) through a sign-preservation mechanism. The DHP, an empirically observed characterization related to the classical vanishing gradient problem, predicts training failure when the convergence time T_conv exceeds 0.72 times the Lyapunov coherence time τ_L. We show that for even convergence steps, the Bloch sphere contraction factor (1−4p/3)^T_conv remains strictly positive for all Pauli depolarizing rates p ∈ [0, 0.75), ensuring gradient direction is preserved regardless of noise magnitude. We term this the quantum parity trap: a regime where the optimizer receives consistent directional signal even as coherence approaches zero. Threshold sweeps confirm convergence at T_conv/τ_L ratios exceeding 15×, compared to the classical DHP limit of 0.72×—a quantum advantage that diverges as p → 0.75. Supplementary experiments establish that (1) Rigetti QPU hardware yields >50σ margin above chance with 4096 shots and readout error mitigation; (2) gradient direction alone is sufficient for convergence—signSGD matches Adam while vanilla SGD fails; and (3) T1 amplitude damping does not replicate the landscape regularization effect of Pauli depolarizing noise, revealing that isotropic Bloch sphere contraction is the mechanistic source of partial-attractor erasure. These results demonstrate that task structure, channel geometry, and optimizer design jointly determine whether quantum noise enables or impedes machine learning convergence.
Abstract Variational Quantum Algorithms (VQAs) are widely regarded as the most promising route to practical quantum advantage on Noisy Intermediate-Scale Quantum (NISQ) hardware, but their trainability is threatened by the barren plateau phenomenon, in which cost-function gradients vanish exponentially with system size and circuit depth. Standard first-order optimizers, which treat the parameter space of a parameterized quantum circuit (PQC) as flat Euclidean space, are especially vulnerable to this effect and to the additional gradient suppression introduced by hardware noise. This extended treatment presents (i) a from-first-principles derivation of Quantum Natural Gradient (QNG) descent, tracing its lineage from Amari's classical information geometry [1] through the Fubini-Study metric to the Quantum Fisher Information Matrix (QFIM); (ii) a complete taxonomy of barren-plateau mechanisms (expressibility-induced, entanglement-induced, cost-function-locality-induced, and noise-induced); (iii) a comparative survey of the broader NISQ optimizer landscape (SPSA, Adam, COBYLA, Rotosolve, Quantum Natural SPSA, and reduced-cost QFIM estimators); and (iv) a controlled simulation study comparing QNG against vanilla gradient descent on a 4-qubit weighted MaxCut problem under three physically motivated NISQ noise models. Across 50 independent trials, QNG raised the convergence success rate from 30% to 95%, reduced the mean iteration count to convergence by a factor of roughly 6, and achieved a 16% net wall-clock speedup despite the added per-iteration cost of QFIM estimation. We give the full theoretical basis for these gains, an exhaustive discussion of their limitations, an A-to-Z glossary of the field's terminology, and a staged roadmap toward physical hardware validation. Figure placements throughout the document are annotated with detailed image-generation prompts so that the accompanying visual assets can be produced independently of this text.
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.
Whether parameterized quantum circuits (PQCs) can be systematically constructed to be both trainable and expressive remains an open question. Highly expressive PQCs often exhibit barren plateaus, while several trainable alternatives admit efficient classical simulation. We address this question by deriving a finite-sample, dimension-independent concentration bound for estimating the variance of a PQC cost function, yielding explicit trainability guarantees. Across commonly used ansätze, we observe an anticorrelation between trainability and expressibility, consistent with theoretical insights. Building on this observation, we propose a property-based ansatz-search framework for identifying circuits that combine trainability and expressibility. We demonstrate its practical viability on a real quantum computer and apply it to variational quantum algorithms. We identify quantum neural network ansätze with improved effective dimension using over 6× fewer parameters, and for VQE on H2 we achieve UCCSD-like accuracy at substantially reduced circuit complexity.
We characterize the geometric and algebraic conditions governing noise-induced landscape regularization in Recurrent Quantum Circuits (Q-RNNs). In quantum recurrent models training on sequence parity (temporal XOR), gradient optimization typically encounters a severe topological trap (the parity plateau at accuracy ≈ 0.72–0.75). While physical environmental noise is conventionally viewed as detrimental to quantum information, we provide empirical evidence that the primary boundary separating regularizing and non-regularizing open-system channels is unitality. Specifically: 1. Unital Channels (e.g., Pauli depolarizing, phase-flip, bit-flip, and phase damping) contract the Bloch sphere symmetrically, preserving the loss landscape's up/down parity symmetry and enabling optimizers to escape the topological traps. 2. Non-Unital Channels (e.g., amplitude damping) shift the center of the Bloch sphere; large non-unital origin drift severely disrupts the optimization landscape by breaking the loss landscape's up/down parity symmetry. 3. We map this boundary empirically by conducting a Generalized Amplitude Damping (GAD) sweep over the thermal parameter [0.0, 1.0] across varying damping rates \0.25, 0.35, 0.50, 0.75\ (Sweep 1), and a state-initialization asymmetry sweep (Sweep 2). We show that convergence is restored within a near-unital basin (e.g., [0.42, 0.56] at =0.50). We show that the slight asymmetry of this basin (which is invariant to the initial state of the memory qubit) is consistent with a structural loop asymmetry caused by the recurrent input qubit reset to |0. We also map the parametric scaling of this basin, showing that the critical displacement boundary |v_z, crit| shrinks as increases due to the rapid contraction of the available state space volume. 4. We characterize the fine-grained transition of the bit-flip channel, revealing a sharp degeneracy hole at exactly p=0.50 where the Bloch sphere contracts to a 1D line. 5. We characterize an empirical optimizer-class robustness pattern at this finite experimental depth (2-qubit, T=3): normalized direction-preserving optimizers (Adam, signSGD, Muon, Normalized GD) achieve 6/6 convergence under unital noise, whereas non-normalized gradient descent (SGD, Heavy Ball) fails (2/6) due to recurrent gradient magnitude contraction. We discuss the analogy between this regularization structure and Quantum Error Correction (QEC), noting that the symmetries protecting parity in stabilizer codes appear to play a parallel role in protecting parity classification landscapes under unital noise. All results are from 2-qubit Q-RNN simulations on T=3 parity; generalization to larger circuits and longer sequences is a target for future work.