Answers are grounded in this paper and cite their sources.
High-precision calibration, diagnostics, and closed-loop control of Noisy Intermediate-Scale Quantum (NISQ) pro-cessors require identifying unknown many-body interaction Hamiltonians H(t), parasitic crosstalk couplings, andopen-system dephasing rates γ(t) from scarce, noisy time-series measurements. While recent advances such asSparse Identification of Quantum Hamiltonian Dynamics (SIQHDy)[1] and Block Tensor Train Com-pressed Sensing Quantum State Tomography (Block-TT-CS QST)[2] address aspects of sparse circuitlearning and state compression independently, existing methods remain limited by: (i) static point-estimationassumptions that cannot track time-varying drift, regime quenches, or non-Markovian 1/f α flicker and RandomTelegraph Noise (RTN); (ii) the absence of posterior uncertainty quantification (credible intervals and ContinuousRanked Probability Scores, CRPS); (iii) problem-dependent manual tuning of sparse thresholding hyperparametersλ; and (iv) exponential O(4N ) memory and computational scaling when constructing full density matrices andPauli operator bases. In this work, we establish a four-layer unified framework—TT-SIQHDy-HNK / MAD-CGC-ES-MDA—backed by five mathematical propositions and validated across seven comprehensive numericalbenchmarks. First, we parameterize mixed quantum states on the Block-TT manifold ˆρTT = A •(N −1) AH toguarantee positive semidefiniteness (PSD) and unit trace by construction, and introduce a matrix-free prefix/suffixtransfer-matrix engine combined with symbolic SU (2) Pauli commutator algebra that evaluates the Liouville sen-sitivity Jacobian ΞTT in O((N + M )R3TT) operations without ever instantiating 2N × 2N matrices (§2.1). Thisachieves state fidelity F = 0.9995 at sampling ratio M/4N = 0.55 and accelerates 12-qubit (412 = 16, 777, 216Pauli space) Jacobian assembly by 1.9 × 107× (611.0 ms and 0.20 MB memory vs. 4.0 PB full-matrix storage; Ex-periments 1 & 4). Second, we integrate parameter-free Median Absolute Deviation thresholding (MAD-ISTA)[6]into SIQHDy, achieving exact support recovery (F1 = 1.000, zero false positives) on both a 5-spin transverse-field Ising model (L = 105 candidates, 25% compressed measurements, 63.3% error reduction) and a 2D 2 × 4(8-qubit) superconducting plaquette lattice with hidden diagonal ZZ crosstalk (L = 276 candidates, 0.29% com-pressed measurements M = 192/65, 536, 79.2% error reduction; Experiments 1 & 4). Third, we couple the identifiedactive subspace to a localized skewness/kurtosis-adaptive hybrid LETKF-NETF (HNK) filter[3] and a statisti-cal t-taper Correlation-based Gaspari–Cohn (CGC) ES-MDA smoother[4, 5] (§2.2). In online drift/quenchtracking (Experiment 2), TT-SIQHDy-HNK reduces RMSE by 65.1% over static SIQHDy and improves CRPS by18.9% over LETKF with 93.0% credible coverage. Under partial observability where only 2 of 5 spins are accessible(N ′ = 6/1024 Paulis; Experiment 3), MAD-Sparse CGC-ES-MDA reduces parameter error by 57.7% and CRPS by81.4%. Under non-Markovian 1/f α drift, TLS telegraph jumps, and unknown asymmetric SPAM readout biases(Experiment 5), state-augmented adaptive-inflation Aug-HNK reduces RMSE by 48.4%, CRPS by 51.3%, and blindlyidentifies 3-axis SPAM bias to 0.0161 L2 error. Fourth, closed-loop submodular D-optimal Fisher active observableselection and symmetric Strang-BCH risk-averse Pareto gate control (§2.3, Experiment 6) boost Shannon informa-tion gain by +135.3% and reduce 3-qubit entangling gate infidelity 1 − Fgate by 42.5% over passive linear control(71.4% over uncompensated gates). Finally, we prove a rigorous Block-TT singular-value truncation error boundfor the Liouville Jacobian (Proposition 5) that supplies a truncation-aware EnKF representation error covarianceRrep ∝ ϵ2TT, and conduct an extensive statistical and information-theoretic validation (Experiment 7): our estima-tor attains 98.4% asymptotic efficiency relative to the active Classical Cramér–Rao Lower Bound (O(N −1/2shots ) slope−0.493), demonstrates statistically significant gains across all six leave-one-out ablations over Nseeds = 25 randomseeds (Wilcoxon signed-rank p ≤ 4.4 × 10−5), maps the exact compressed sensing and (Ne, σ) stability phase tran-sitions, and achieves flat Talagrand rank histograms (KS = 0.046) with 2.07% Expected Calibration Error (ECE)across all credible quantiles α ∈ [0.05, 0.99].
No takes yet. Share an insight, caveat, or question.
Kengo Imai (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: