Quantum Path Kernel (QPK)

• Quantum Path Kernel (QPK) is a mathematical formalism that links quantum dynamics, path integrals, and high-dimensional kernels with rigorous error bounds.
• It is applied across quantum mechanics, optics, and quantum machine learning to encode propagators, feature maps, and state overlaps efficiently.
• QPK constructions enable practical implementations via Trotterised circuits and one-clean-qubit models, improving generalization in noisy and complex systems.

The Quantum Path Kernel (QPK) is a mathematical and algorithmic formalism linking quantum dynamics, path integrals, and high-dimensional kernel methods. QPK emerges in rigorous quantum field theory, quantum machine learning, quantum optics, and quantum signal processing as a unifying object for encoding the structure and statistical properties of quantum evolutions along parameter, state, or physical paths. The QPK provides a flexible kernel representation for propagators, feature maps, and functional overlaps in models ranging from supersymmetric quantum mechanics to quantum neural networks and quantum circuits. Multiple rigorous and physically grounded constructions are present in the literature, including kernelized Feynman time-slice propagators, coherent sum-of-paths kernels for optical systems, trajectory-averaged neural tangent kernels, and randomized unitary circuit overlaps for quantum path signatures.

1. Path Integral and Heat Kernel Construction

Within the context of $N=1$ supersymmetric quantum mechanics on a compact, even-dimensional Riemannian manifold $M$, the QPK is defined as the limit of Feynman time-sliced path integrals yielding the heat kernel of the Laplace–de Rham operator $A_{LdR}$ on differential forms. For time interval partition $P = (t_1,\dots,t_N)$ of $t$, the $N$-fold product of short-time approximate propagators $K(x,y,\psi_x,\psi_y;t)$ combines bosonic and Grassmann (fermionic) degrees of freedom. The fine-partition limit yields the QPK kernel: $$K_{QPK}(t; x, y) = \lim_{|P| \to 0} \prod_{i=1}^N K(x_{i-1}, x_i; \Delta t)$$ in the sense of *-product composition (Berezin integral in $\psi$ and Riemannian measure in $x$), which converges uniformly to the heat kernel for $M$0 as demonstrated in (Fine et al., 2012).

Key properties include:

• Semigroup Law: The *-product nearly satisfies a semigroup property, controlling error terms under time-slicing.
• Uniform Convergence: Precise operator-norm error bounds, $M$1 for any $M$2.
• Small-Time Asymptotics: Steepest-descent analysis of the short-time kernel yields a leading supertrace term equal to the Pfaffian of the curvature 2-form; this recovers the Gauss–Bonnet–Chern integrand.

This construction makes QPK a rigorous realization of the path integral kernel for supersymmetric diffusions, with the full structure of curvature couplings and geometric invariants encoded in its short-time expansion.

2. Quantum Path Kernels in Multi-Plane Optics and Quantum Path Computing

The QPK is realized within photonic quantum path computing (QPC) via multi-plane diffraction (MPD) architectures, wherein the quantum evolution is encoded by the coherent sum of amplitudes over exponentially large path ensembles traversing multiple diffraction planes. The path amplitude for a given detector coordinate $M$3 is: $M$4 where $M$5 are quadratic-phase propagators (Fresnel, fractional Fourier, or general linear canonical transforms), $M$6 are slit-mask functions selecting a particular path, and $M$7 is the source wavefunction (Gulbahar, 2019).

By treating the output coordinate as an input for learning, a normalized feature map $M$8 induces the kernel: $M$9 with the dimension of the feature space scaling exponentially with the number of planes and slits. This construction applies to both discrete (sum-over-paths) and continuous (path-integral) regimes and is positive-definite and Hermitian by construction. Modulations of the source mode—e.g., employing Hermite–Gaussian profiles—enrich the feature space, embedding high-order polynomial non-linearities. The resulting QPKs serve as kernel machines for quantum-inspired classification (e.g., SVMs), neural network layers, and non-trivial function approximation in optical hardware.

3. Quantum Path Kernel as a Generalized Quantum Neural Tangent Kernel

The QPK, as formulated in quantum machine learning, generalizes the quantum neural tangent kernel (QNTK) by integrating over the full parameter trajectory $A_{LdR}$0 of a quantum neural network (QNN) during training. For a QNN predictor function $A_{LdR}$1, the instantaneous QNTK is $A_{LdR}$2. The QPK is then: $A_{LdR}$3 or numerically as a time-averaged sum over epochs (Incudini et al., 2022).

This construction embeds the hierarchical, layer-wise dynamics of training—the divergence from lazy (linear) training—by accumulating the evolving tangent kernel structure as parameters traverse the optimization landscape. Empirical results on Gaussian XOR mixtures confirm that QPK-based kernels outperform QNTK-based kernels particularly under moderate to high noise, and better capture hierarchical feature learning effects mimicking deep classical networks.

4. Quantum Path Signatures and Randomized Circuit Feature Maps

In the framework of quantum path signatures, the QPK arises as the overlap of randomized path-development unitary circuits constructed from a piecewise-linear path $A_{LdR}$4 and randomly sampled sparse Pauli-string "gauge fields" $A_{LdR}$5: $A_{LdR}$6 with each $A_{LdR}$7 for Rademacher random variables $A_{LdR}$8 and Pauli-strings $A_{LdR}$9. The quantum feature map is the pure state $P = (t_1,\dots,t_N)$0, and the QPK is: $P = (t_1,\dots,t_N)$1 or, equivalently, $P = (t_1,\dots,t_N)$2 for the associated mixed-state quantum signatures (Crew et al., 7 Aug 2025).

The circuit is implemented via Trotterised Pauli rotations, and the one-clean-qubit model (DQC1) provides a logarithmic-qubit, polynomial-time algorithm for empirical QPK estimation. Loop equations for the kernel connect the quantum and classical signature kernel worlds, with the quantum variant admitting higher expressivity and sample efficiency under quantum protocols.

5. Error Bounds, Convergence, and Complexity

Error control and convergence properties are proven in several QPK frameworks:

• In the path-integral QPK (supersymmetric quantum mechanics), operator-norm error bounds between the time-sliced approximation $P = (t_1,\dots,t_N)$3 and the limiting $P = (t_1,\dots,t_N)$4 scale as $P = (t_1,\dots,t_N)$5, and the limit exists uniformly for all positive $P = (t_1,\dots,t_N)$6 (Fine et al., 2012).
• For randomized circuit formulations, Trotterization error, sparse Pauli approximation error, and sampling error can each be made less than $P = (t_1,\dots,t_N)$7 with total gate count and memory usage polynomial in $P = (t_1,\dots,t_N)$8 and $P = (t_1,\dots,t_N)$9, enabling exponential space advantages over classical counterparts (Crew et al., 7 Aug 2025).
• In the QNN/QPK construction, complexity for assembling the path kernel matrix across $t$0 epochs is $t$1 in memory or $t$2 for on-the-fly computation, where $t$3 is the number of data points and $t$4 is the number of parameters (Incudini et al., 2022).

These results demonstrate that the QPK is not only mathematically well defined under appropriate limits, but can be numerically implemented on both classical and quantum architectures with quantifiable bounds.

6. Applications and Expressivity

QPKs are central in multiple application domains:

• Quantum Machine Learning: Kernel SVMs, quantum neural networks with QPK layers, and hierarchical feature extractors, yielding improved generalization under non-trivial, noise-corrupted datasets (Incudini et al., 2022, Gulbahar, 2019).
• Quantum Signal Processing and Control: QPKs enable systematic mapping of signal histories to high-dimensional quantum feature spaces, facilitating temporal correlations and entanglement exploitation (Gulbahar, 2019).
• Geometric Analysis and Topological Invariants: The path integral QPK construction yields the Gauss–Bonnet–Chern theorem via steepest-descent of the supertrace, with the short-time limit extracting topological Pfaffians (Fine et al., 2012).
• Quantum Advantage and Complexity: One-clean-qubit circuits for QPK estimation achieve exponential space and (for certain path classes) time efficiency compared to classical matrix-model or PDE approaches, with richer feature spaces achieved by randomized Pauli ensembles (Crew et al., 7 Aug 2025).

Universal-approximation theorems imply that QPKs, via suitably constructed ensembles, can approximate arbitrary function classes in compact domains, especially in optics-based or circuit-based realizations. The construction encourages tailored feature design—for instance, using Hermite–Gaussian sources or custom quantum circuit architectures—for domain-specific learning or simulation tasks.

7. Open Problems and Future Directions

Prominent research directions concerning QPKs include:

• Efficient Trajectory Sampling: Strategies for selecting influential epochs (e.g., Hessian-based event detection) in QPK computation for deep QNNs (Incudini et al., 2022).
• Spectral Analysis of QPKs: Analytical characterization of reproducing kernel Hilbert spaces induced by QPKs, including connections to representation theory for quantum circuits and path signatures.
• Extensions to Dissipative and Non-Unitary Dynamics: Generalization of QPK constructions to include dissipative QNNs, partial measurements, and non-Hermitian path developments (Incudini et al., 2022).
• Comparisons of Quantum and Classical Signature Kernels: Quantitative analysis of expressivity and sample complexity differentials under corresponding matrix-model and quantum circuit implementations (Crew et al., 7 Aug 2025).

A plausible implication is that QPK methodology may serve as a central mathematical bridge for unifying quantum learning theory, high-dimensional signal processing, and rigorous quantum analysis, particularly as near-term quantum devices mature and the need for expressive, efficient, and theoretically grounded kernels becomes more acute.