Quantum-Classical Truncated Newton

• The paper presents a hybrid method that leverages quantum linear system solvers with classical truncated Newton techniques to efficiently target key curvature modes in high-dimensional problems.
• The approach employs quantum phase estimation, block-encoding, and ancilla rotations to solve linear systems, thereby improving convergence in deep learning and quantum chemistry applications.
• The truncation strategy, underpinned by random matrix theory, filters noise-dominated eigenvalues to ensure robust error control and enhanced computational efficiency.

A quantum-classical truncated Newton method is a hybrid optimization strategy that integrates quantum linear system solvers with classical truncated or saddle-free Newton frameworks to accelerate large-scale nonlinear optimization, particularly in machine learning, quantum chemistry, and scientific computing. This approach leverages the polynomial or exponential scaling advantages of quantum subroutines while relying on classical components for data preparation, oracle construction, and certain update steps. The method is designed to escape saddles and correct for crucial curvature modes, exploiting the empirical spectral structure—such as the Marchenko-Pastur distribution—of Hessian matrices in high-dimensional problems (Wossnig et al., 2017, Li et al., 2024, Feldmann et al., 2022, Buxadé et al., 24 Mar 2026).

1. Mathematical Foundations and Classical Truncated Newton Methods

At the core, quantum-classical truncated Newton algorithms address optimization problems of the form

$\min_{\theta \in \mathbb{R}^n} f(\theta),$

where $f$ is nonconvex and twice-differentiable, typical in deep neural network training or SCF orbital optimization. The classical (full) Newton’s method updates parameters via

$$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$

where $H$ is the Hessian. For high-dimensional settings ($n \gg 10^6$), explicitly forming or inverting $H$ is prohibitive ($O(n^2)$ memory, $O(n^3)$ computation), and the presence of small or negative eigenvalues leads to ill-conditioning and saddle-point attraction.

The truncated Newton framework operates by spectrally decomposing $H$,

$H = \sum_{i=1}^n \lambda_i v_i v_i^T,$

replacing negative eigenvalues with their magnitudes (saddle-free Newton), and truncating the pseudo-inverse to ignore directions with $f$0:

$f$1

Thus, the update is performed in the subspace of large curvature modes, sidestepping noise-dominated directions and reducing ill-conditioning (Wossnig et al., 2017, Feldmann et al., 2022).

2. Quantum Linear System Solvers and Hybrid Integration

Quantum acceleration is achieved by mapping the Newton or truncated Newton step to a quantum linear system solve (QLSS). For a system $f$2 or $f$3, a quantum register is prepared in the state $f$4. The quantum subroutine (typically a variant of the HHL algorithm) performs these major steps:

• Hamiltonian simulation via block-encoding: Implementing $f$5, scalable as $f$6, where $f$7 is the oracle sparsity.
• Quantum phase estimation (QPE): Decomposes $f$8 in $f$9’s eigenbasis and estimates each eigenvalue.
• Ancilla rotation and post-selection: Applies $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$0 or $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$1 in the quantum state, with rotation angles or counting/comparison-based gadgets to avoid explicit arithmetic.
• Amplitude amplification: Increases probability of success from $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$2 to $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$3, with $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$4 the effective condition number.

After uncomputing ancillas and measurement registers, $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$5 is obtained. Full classical read-out of $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$6 requires $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$7 repetitions, which currently limits speedups to polynomial, but if passed to subsequent quantum steps, this bottleneck can be circumvented (Wossnig et al., 2017, Li et al., 2024, Buxadé et al., 24 Mar 2026).

3. Random Matrix Structure, Truncation, and Practical Error Control

The truncation strategy is rigorously justified by random matrix theory, particularly the Marchenko-Pastur theorem. Empirical studies of deep learning Hessians show their spectra comprise a "bulk" near zero (interpreted as noise) and a small number $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$8 of informative outlier eigenvalues. The truncation threshold $$\Delta\theta = -[H(\theta_t)]^{-1} \nabla f(\theta_t)$$9 is set just above the MP bulk edge so that inversion is restricted to signal-dominated directions (Wossnig et al., 2017). In quantum chemistry SCF contexts, density-difference subspace expansion and projected Hessian actions via the Augmented Roothaan-Hall (ARH) method ensure that the Hessian’s action is approximated in a lower-dimensional but information-rich subspace, yielding stability and computational efficiency (Feldmann et al., 2022).

Error control in the quantum-classical scheme involves balancing the quantum-solve precision $H$0 (e.g., $H$1 for $H$2 ancilla qubits in QPE), the Newton residual at each outer iteration, and the inexact Newton forcing terms (Dembo-Eisenstat-Steihaug criteria). Provided the quantum-induced inexactness remains within the forcing term, local superlinear or quadratic convergence is retained (Buxadé et al., 24 Mar 2026).

4. Algorithmic Workflow and Hybrid Scheduling

A typical quantum-classical truncated Newton iteration involves:

1. Classical evaluation: Compute oracles for the Hessian and gradient, regularize and sparsify $H$3 (with symmetry-aware pruning and addition of $H$4 for conditioning), and prepare data for quantum loading.
2. Quantum solve: Implement QLSS (modified HHL, block-encoding, or density-matrix exponentiation) to obtain an amplitude-encoded solution to $H$5 or $H$6.
3. Measurement and update: Recover solution components via quantum measurement or amplitude estimation; classically update parameters.
4. Scheduling: Runtime selection between classical and quantum solvers is based on estimated $H$7 (classical) and $H$8 (quantum) costs, incorporating problem size $H$9, condition number $n \gg 10^6$0, sparsity $n \gg 10^6$1, gate times $n \gg 10^6$2, and desired tolerance $n \gg 10^6$3. For well-conditioned, sparse regimes, the quantum step is favored; otherwise, classical solvers (e.g., CG, LU) are used (Li et al., 2024).

This scheduling paradigm realizes an adaptive hybrid optimization pipeline, dynamically leveraging quantum resources where speedup is viable.

5. Complexity and Scaling Analysis

The asymptotic cost per Newton iteration for key algorithms is:

Approach	Per-iteration Complexity	Speedup Regime
Classical saddle-free Newton	$n \gg 10^6$4 (truncated, low-rank), $n \gg 10^6$5 (full)	Baseline, highly intensive
Quantum-classical truncated Newton	$n \gg 10^6$6 (quantum) + $n \gg 10^6$7 (read-out)	Polynomial, limited by measurement
Q-Newton (hybrid scheduling)	$n \gg 10^6$8	Polynomial/exponential (sparsity, conditioning dependent)
Quantum Newton for PDEs	$n \gg 10^6$9 qubits and gate depth	Exponential for large $H$0, if full quantum chaining

Measurements dominate cost if full classical readout is required. In theory, exponential speedup is attainable if adjacent quantum steps can consume quantum outputs directly, bypassing measurement (Wossnig et al., 2017, Buxadé et al., 24 Mar 2026).

6. Applications in Scientific Computing and Machine Learning

Quantum-classical truncated Newton methods have been applied or proposed in several fields:

• Deep neural networks: Training can be accelerated by quantum subroutines that deliver second-order curvature information on mini-batches, with empirical improvements in training time over both SGD and classical Newton methods (Wossnig et al., 2017, Li et al., 2024).
• Quantum chemistry (ARH method): SCF orbital optimizations, both electronic and nuclear-electronic, benefit from subspace-projected truncated Newton iterations, enhancing convergence in strongly correlated and quantum-nuclear systems (Feldmann et al., 2022).
• Nonlinear PDEs: Quantum-accelerated inner solves for Newton iterations of discretized PDEs (Navier-Stokes, nonlinear Poisson) allow for polynomial or exponential scaling in very high dimensions, contingent on state preparation and measurement overheads (Buxadé et al., 24 Mar 2026).

In all domains, quantum acceleration hinges on exploiting intrinsic low-rank structure, spectral gaps, and efficient quantum data oracles.

7. Resource Estimates, Limitations, and Prospective Improvements

Resource requirements typically involve $H$1 qubits to encode solution states and $H$2 ancilla for quantum phase estimation. Gate depth scales as $H$3, with the main bottlenecks being quantum measurement overhead, current quantum gate times, and state preparation. For realistic applications (e.g., $H$4 unknowns in PDEs), $H$5 qubits are required but coherence time and circuit depth remain challenging (Buxadé et al., 24 Mar 2026).

Potential improvements include:

• Advanced preconditioning and regularization to reduce $H$6.
• Block-encoding constructions for structured or low-rank Hessians.
• Integration within automatic differentiation frameworks for quantum-aware end-to-end training.
• Quantum schedule modules that dynamically select between classical/quantum steps based on problem instance (Li et al., 2024).

This suggests that future quantum hardware advances—particularly reductions in gate time and improved state manipulation—will be critical in realizing the full potential of quantum-classical truncated Newton methods across scientific disciplines.

---

References:

• Wossnig et al., "Quantum-classical truncated Newton method for high-dimensional energy landscapes" (Wossnig et al., 2017)
• Feldmann, Baiardi, Reiher, "Second-order self-consistent field algorithms: from classical to quantum nuclei" (Feldmann et al., 2022)
• Mandelt Buxadé et al., "Solving Nonlinear Partial Differential Equations via a Hybrid Newton Method Using Quantum Linear System Solver" (Buxadé et al., 24 Mar 2026)
• Zhao et al., "Q-Newton: Hybrid Quantum-Classical Scheduling for Accelerating Neural Network Training with Newton's Gradient Descent" (Li et al., 2024)