Quantum Krylov Learning

AI summary

Overview

Research area: Quantum information science, specifically Hamiltonian learning for quantum sensing and boundary spectroscopy, using operator Krylov subspace (Lanczos/Stieltjes) methods.

Technical level: Advanced. The paper assumes familiarity with Liouville space, Krylov subspaces, Lanczos recursion, spectral measures, and Lieb–Robinson bounds, though the physical setup it addresses is easy to state.

Scope (one sentence): The paper introduces Quantum Krylov Learning (QKL), a protocol that reconstructs model-independent spectral information about an unknown Hamiltonian from a single-temperature probe autocorrelation, using only probe state preparation, free evolution, and probe readout.

What This Paper Is About

Normally, learning the Hamiltonian of a natural quantum system requires either proposing a microscopic model and testing it, or having programmable-device-level control: arbitrary state preparation, controlled gates, and multi-basis readout. In many realistic settings, such as quantum sensing, the system of interest is inaccessible and can only be touched through a controllable probe degree of freedom.

The paper asks what can still be learned when the only allowed operations are preparing the probe, letting the whole system evolve freely under the unknown Hamiltonian, and measuring the probe. The authors show that if the inaccessible part of the system is at or near infinite temperature, the probe autocorrelation alone is enough to reconstruct the operator Krylov Jacobi matrix, a compact object that encodes the dynamical modes visible to the probe.

Key Contributions

1. A restricted-access, oracle-free learning framework (QKL). The protocol requires no symmetry assumption, no parametric model of the Hamiltonian, and no controlled-unitary gates beyond free evolution and single-probe measurement. It needs measurements at only a single fixed bath temperature, rather than the temperature scan used in quantum probe tomography.

2. Exact reconstruction theorem. When the probe autocorrelation is sampled faster than twice the highest Bohr frequency (Nyquist) over a sufficiently long window, the operator Krylov Jacobi matrix is reconstructed exactly from probe-only data (Theorem 3, Sec. 4). This is verified numerically as an exact equivalence between oracle and oracle-free constructions (Fig. 5).

3. Operator Krylov structure, Lieb–Robinson bound, and transport formula. The Lanczos process for the Liouvillian in the Hilbert–Schmidt inner product yields a real symmetric tridiagonal Jacobi matrix whose diagonal entries vanish identically for any Hamiltonian, without symmetry assumptions (Theorem 1). The associated Krylov chain obeys a Lieb–Robinson bound with velocity 2 (Theorem 2), and a transport-time formula τ_op(m) = Σ_{n=1}^{m−1} 1/β_op_n that serves as a data-driven criterion for the required observation window.

4. End-to-end sample complexity guarantee. Under generic conditions on spectral gaps and spectral weights, a number of measurement shots polynomial in 1/ε suffices for spectral precision ε, with explicit polynomial degrees (Theorem 4, Sec. 6). The bound avoids the exponential cost of resolving every many-body spectral feature by fixing the spectral resolution in advance.

Main Findings

• Autocorrelation is a characteristic function. At infinite temperature, the probe autocorrelation C_α(t) = Tr[σ_p^α(t) σ_p^α]/D equals the characteristic function ∫ e^{iωt} dμ_α(ω) of the operator Krylov spectral measure μ_α, which is supported on the Bohr frequencies (energy differences, not energies) and is an even, positive measure with total weight one.

• Two-state measurement suffices. The infinite-temperature autocorrelation is obtained as a single difference of two probe signals, (1/2)[s_α(t|ρ₊) − s_α(t|ρ₋)], using probe eigenstate preparation and probe-only readout, with no control or readout of the bath.

• Vanishing diagonal is unconditional. The Jacobi diagonal coefficients α_op_n vanish for every n and every Hamiltonian because Krylov vectors alternate in adjoint parity and the trace is cyclic. This originates from the Hermiticity of the probe seed.

• No dark sector relative to the Jacobi matrix. A state-dependent signal s_α(t) can miss transitions that carry zero weight for the chosen initial state; the autocorrelation C_α(t) misses a transition only when its weight p_ν vanishes, which is exactly when it does not enter J_op_m either. A dark sector still exists relative to H itself, consisting of transitions with ⟨E_j|σ_p^α|E_k⟩ = 0.

• Lieb–Robinson velocity equals 2. The bound takes the form |(Q_m|e^{iJ_op_m t}|Q_n)| ≤ exp[−λ ρ_op(m,n) + Ω_op_{M,λ}|t|], with Ω_op_{M,λ}/λ approaching 2 as λ → 0. The velocity is 2 because each site of the Krylov chain couples in two directions.

• Round-trip versus one-way propagation. The break time defined from the round-trip quantity C_α − C_α^(m) scales as t*op(m) ≈ a τ_op(m) + b with slope a ≈ 1, because the signal must travel out to the truncation boundary and back at velocity 2. A one-way matrix element such as (Q{m−1}|e^{iJ_op_m t}|Q_0) instead becomes appreciable near τ_op(m)/2, corresponding to slope a ≈ 0.5.

• Finite temperature is a controlled perturbation. At β > 0 the bath weighting is no longer proportional to the identity, so the measured signal picks up unknown thermal weights on otherwise unchanged Bohr frequencies. Section 6.2 bounds this deviation and argues it is subdominant to finite-window and shot-noise effects for NV-center and donor-spin platforms; for the bitstring-randomization route the infinite-temperature condition holds exactly.

• Numerical benchmark on an XXZ spin ring with a transverse field. The fully data-driven pipeline reproduces the leading oracle Jacobi coefficients to within a few percent, the higher coefficients accessible within the present observation window to 10–34%, and the transport-time scale τ_op to 0.9–6.6%. The specific ring size used in this benchmark is not reported in the supplied text.

• Comparison to prior state-Krylov work. Reference [30] develops the state Krylov analog, proving a Lieb–Robinson bound with velocity v_LR = 2 and the transport formula t*(m) ≈ τ_m = Σ_{n<m} 1/β_n with slope a ≈ 1, confirmed numerically for N = 6–12. The operator chain here is of the same mathematical type, so the same argument transfers with β_n replaced by β_op_n.

Methodology in Plain English

The physical picture. Split the world into a controllable probe and an inaccessible bath. The probe can be prepared, left to evolve under the unknown total Hamiltonian, and measured. The bath cannot be initialized, controlled, or read. The one assumption about the bath is that it is at or near infinite temperature, meaning the thermal energy scale k_B T is much larger than the bath's characteristic internal energy scale ΔE.

Step 1: Measure a correlation. Prepare the probe in two opposite eigenstates of a chosen Pauli operator, let the system evolve freely for a set of discrete times, and measure the probe. Half the difference of the two resulting signals equals the infinite-temperature autocorrelation. This works because at infinite temperature every bath configuration carries equal weight, so the bath state drops out and only Hamiltonian dynamics remain.

Step 2: Extract frequencies and weights. Treat the measured time series as a sum of oscillating terms. A matrix-pencil (Prony-type) estimator, built on a generalized eigenvalue problem for a Hankel matrix, recovers the oscillation frequencies and their weights. A non-negative least-squares step enforces the physical requirement that the weights be positive.

Step 3: Build the Jacobi matrix. Feed the recovered frequency-weight pairs into the Stieltjes/Lanczos recursion, exactly as one would with a known Hamiltonian. The output is the operator Krylov Jacobi matrix: a one-dimensional tight-binding chain in operator space whose hopping amplitudes β_op_n describe which dynamical modes the probe can actually see.

Step 4: Know when to stop measuring. Because the hopping coefficients are now known from data, the transport-time formula τ_op(m) = Σ 1/β_op_n tells the experimenter, without any model input, how long an observation window is needed for a given Krylov depth.

Why it can be efficient. A generic many-body system has exponentially many distinct spectral contributions with shrinking gaps and weights. Instead of resolving all of them, the protocol fixes the spectral resolution in advance. At fixed resolution the number of resolvable spectral features stays bounded independently of system size, and the leading Jacobi coefficients and τ_op depend only on low moments of the spectral measure, which coarse-grained reconstruction already captures.

Why This Matters

Impact on research. The work removes three assumptions that have limited Hamiltonian learning under restricted access: symmetry (such as excitation-number conservation in the Burgarth–Maruyama–Nori protocol), a known parametric form (as in quantum probe tomography), and controlled operations beyond free evolution. It reframes Hamiltonian learning as spectral reconstruction of a Jacobi matrix rather than parameter fitting, and it supplies both a data-driven stopping criterion and an end-to-end shot-count guarantee, which are often missing from protocol proposals.

Real-world applications (as identified by the authors):

• Nitrogen-vacancy center sensors coupled to unknown nuclear-spin environments, where the sensor operates at room temperature and the bath's dipolar coupling scale is of order 1 kHz or less, far below the operating temperature.
• Donor-spin sensors, which typically operate at cryogenic temperatures for high-fidelity readout, and where the same separation of scales makes the infinite-temperature condition hold.
• Local probes of correlated materials, i.e., boundary spectroscopy of systems whose microscopic degrees of freedom cannot be individually addressed.
• Hole-spin qubits in Si and Ge nanostructures, which realize anisotropic, XXZ-like exchange through spin–orbit coupling (recently measured and controlled in a Si FinFET) and can be arranged in ring geometries, matching the paper's numerical benchmark.

Industry relevance. The protocol is directly aimed at sensing hardware and semiconductor spin-qubit platforms rather than gate-model processors. It requires no additional control hardware beyond state preparation, idle evolution, and readout, which lowers the experimental barrier. It also offers a route for characterizing unknown noise environments on programmable hardware by emulating the infinite-temperature bath through classical bitstring randomization.

Future Directions

• Extending beyond closed-system dynamics. The Jacobi (tridiagonal) structure relies on the Liouvillian being Hermitian, which holds for a closed system. For an open-system Liouvillian with dissipators the generator is generically non-Hermitian and the output becomes a Hessenberg rather than a Jacobi matrix; the consequences for the reconstruction pipeline are left open.

• Sharpening finite-temperature and finite-window error budgets. Section 6.2 bounds finite-temperature deviation and argues it is subdominant on NV and donor-spin platforms, but a full accounting across platforms and bath coupling strengths remains to be worked out.

• Improving higher-coefficient accuracy. The numerical benchmark reproduces leading Jacobi coefficients to within a few percent but higher coefficients only to 10–34% within the present observation window, leaving room for better estimators or longer effective windows.

• Hardware demonstrations. The authors identify Si and Ge hole-spin qubits in ring geometries and NV/donor-spin sensors as promising targets, but the supplied content reports no experimental implementation.

Target Audience

Researchers in quantum information and quantum sensing working on Hamiltonian learning, quantum probe tomography, or noise spectroscopy; condensed-matter and AMO physicists interested in the recursion method and Krylov subspace techniques applied to unknown Hamiltonians; experimentalists on NV-center, donor-spin, and semiconductor hole-spin platforms who need model-independent characterization with minimal control; and theorists interested in Lieb–Robinson bounds and operator-space transport in Krylov chains. Readers without a background in Liouville space and Lanczos recursion will find the technical sections demanding.

Authors’ abstract

The Hamiltonian of a natural quantum system is usually determined indirectly, by proposing microscopic models and comparing their predicted observables with experiment. A more direct approach is to extract information about the Hamiltonian from measurements on the system itself. In many settings, however, direct access to the full system is unavailable and measurements can be performed only through controllable probes, as in quantum sensing and boundary spectroscopy. Existing Hamiltonian learning protocols under such restricted access typically require a special symmetry, a known parametric form, or controlled operations beyond free evolution. We introduce Quantum Krylov Learning (QKL), a restricted-access framework that avoids these requirements. QKL uses single temperature probe autocorrelators and their operator Krylov structure to extract model-independent spectral information about the unknown system. Rather than fitting a parametrized Hamiltonian, it reconstructs the operator Krylov Jacobi matrix representing the dynamical modes accessible to the probe. The method assumes that the unobserved system is at, or near, infinite temperature. Under this condition, the Jacobi matrix can be reconstructed directly from the autocorrelation of a probe observable, without oracle access to the Hamiltonian. The required operations are free evolution under the unknown $H$, together with probe state preparation and measurement. The learning core therefore requires no symmetry, no parametric model, and no controlled-unitary gates. We show a Lieb--Robinson bound for the operator Krylov chain and derive a data-driven criterion for the required observation window. We also establish an end-to-end sample complexity guarantee: under generic conditions on the spectral gaps and weights, the number of measurement shots required to achieve spectral precision $\varepsilon$ is polynomial in $1/\varepsilon$.