# Lee-Yang-Gapped Quantum Neural Layer

- ID: 706
- Canonical URL: https://synthcore.org/idea/706/lee-yang-gapped-quantum-neural-layer
- API JSON: https://synthcore.org/api/idea/706.json
- API Markdown: https://synthcore.org/api/idea/706.md
- Verification status: unverified
- Source: [arXiv:2607.10765](https://arxiv.org/abs/2607.10765)
- Category: training
- Solves: stability, sample-efficiency
- ML areas: optimizer, training, regularization
- Math tags: spectral-theory, linear-algebra, dynamical-systems
- Ratings: usefulness 5/10; difficulty 6/10; novelty 9/10

## Idea description

Build a variational quantum neural network whose trainable 2-qubit Hamiltonian is projected into the Lee-Yang coupling cone and augmented by a uniform field term -h sum_i Z_i. The theorem certifies a nondegenerate ground state and a gap at least h/4, enabling imaginary-time state-preparation layers with predictable exponential suppression of excited-state error.

## Mathematical statement

For qubits i,j, the Hamiltonian is H=sum_{ij} H_{ij}, with H_{ij}=-w^z_{ij}Z_iZ_j+w^x_{ij}X_iX_j+w^y_{ij}Y_iY_j+w^{xy}_{ij}X_iY_j+w^{yx}_{ij}Y_iX_j. Here X_i,Y_i,Z_i are Pauli operators and the real coefficients w^z_{ij},w^x_{ij},w^y_{ij},w^{xy}_{ij},w^{yx}_{ij} satisfy w^z_{ij} >= (1/2)*sqrt((w^x_{ij}-w^y_{ij})^2+(w^{xy}_{ij}+w^{yx}_{ij})^2) + (1/2)*sqrt((w^x_{ij}+w^y_{ij})^2+(w^{xy}_{ij}-w^{yx}_{ij})^2). Define H_h=H-h*sum_i Z_i with h>0. The paper proves that its ground state is nondegenerate and that the spectral gap Delta=E_1-E_0 satisfies Delta>=h/4, independently of system size and coupling strengths. For imaginary-time filtering exp(-tau H_h), the excited-state-to-ground-state amplitude ratio is multiplied by at most exp(-tau Delta) <= exp(-tau h/4). If the initial state has nonzero ground-state overlap, tau >= (4/h)*log(1/epsilon) makes this multiplicative ratio at most epsilon.

## Key formulas

- $$H_{ij}=-w^{z}_{ij}Z_iZ_j+w^{x}_{ij}X_iX_j+w^{y}_{ij}Y_iY_j+w^{xy}_{ij}X_iY_j+w^{yx}_{ij}Y_iX_j$$
- $$w^z_{ij}\geq\frac{1}{2}\sqrt{(w^x_{ij}-w^y_{ij})^{2}+(w^{xy}_{ij}+w^{yx}_{ij})^{2}}+\frac{1}{2}\sqrt{(w^x_{ij}+w^y_{ij})^{2}+(w^{xy}_{ij}-w^{yx}_{ij})^{2}}$$
- $$H_h=H-h\sum_{i=1}^{n}Z_i,\qquad E_1-E_0=\Delta\geq\frac{h}{4}$$
- $$\frac{\|P_{\perp}e^{-\tau H_h}|\psi\rangle\|}{\|P_0e^{-\tau H_h}|\psi\rangle\|}\leq e^{-\tau\Delta}\frac{\|P_{\perp}|\psi\rangle\|}{\|P_0|\psi\rangle\|}\leq e^{-\tau h/4}\frac{\|P_{\perp}|\psi\rangle\|}{\|P_0|\psi\rangle\|}$$

## Implementation notes

1. Integration point: use this Hamiltonian as the energy generator in a small variational quantum neural network, quantum Boltzmann model, or hybrid QNN. Replace unconstrained pairwise Hamiltonian parameters with coefficients for every active edge (i,j), and use H_h as the training objective whose ground state is represented by the circuit. Implement imaginary-time filtering with a variational quantum imaginary-time algorithm or a short product-formula approximation to exp(-delta_tau H_h). 2. Pseudocode: initialize raw parameters theta for each edge; decode theta into wx, wy, wxy, wyx and raw_wz; compute r_minus=sqrt((wx-wy)^2+(wxy+wyx)^2) and r_plus=sqrt((wx+wy)^2+(wxy-wyx)^2); set wz=0.5*(r_minus+r_plus)+softplus(raw_wz)+margin; construct H=sum_edges Hij-h*sum_i Zi; estimate the current energy and force/metric matrices; solve the variational imaginary-time update, apply the circuit parameters, and repeat for K steps. Choose delta_tau and K so tau=K*delta_tau is near 4*log(1/epsilon)/h. 3. The coupling cone and certified lower bound Delta>=h/4 come from the paper. Ground-state overlap, Trotter error, metric conditioning, shot noise, and the actual finite-system gap are estimated empirically. 4. First experiment: use 8-16 qubits with random graph edges and random couplings, comparing an unconstrained QNN Hamiltonian, a constrained Hamiltonian without field, and the Lee-Yang-constrained model with h in {0.2,0.5,1.0}. Use dense diagonalization to obtain exact gaps and ground states, and train all models with equal circuit depth and measurement budget. Track energy error, fidelity, excited-state weight, gradient variance, and the number of imaginary-time steps. Success means a larger observed minimum gap, excited-state suppression consistent with exp(-tau*h/4), and lower final energy error or fewer steps at equal measurement cost.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.