Quantum State Preparation with the QNN-based SRBB Algorithm

Technical Summary: Quantum State Preparation with the QNN-based SRBB Algorithm

Problem Statement
Quantum State Preparation (QSP) is a fundamental subroutine in numerous quantum algorithms, including Quantum Machine Learning (QML), linear equation solvers, and Hamiltonian simulation. The objective is to construct a unitary operator $U$ such that $U|0\rangle^{\otimes n} = |\psi\rangle$, where $|\psi\rangle$ is a desired $n$-qubit state. While exact QSP algorithms exist (e.g., using Uniformly Controlled Rotations), they often suffer from exponential depth and gate counts ($O(2^n)$), making them infeasible for near-term hardware. Furthermore, many existing approaches rely on ancillary qubits to achieve better space-time trade-offs, which increases hardware requirements. The challenge lies in developing an approximate QSP method that minimizes resources (specifically CNOT count and circuit depth) while maintaining high accuracy, without relying on ancillary qubits.

Methodology
The authors propose a novel variational quantum algorithm structured on the Standard Recursive Block Basis (SRBB), a hierarchical construction for the matrix algebra of the $SU(2^n)$ group. The methodology leverages the topological and algebraic properties of Lie groups to design a Variational Quantum Circuit (VQC).

1. Theoretical Framework (SRBB and Lie Algebras):

   • The algorithm is grounded in the decomposition of the unitary group $U(2^n)$ into a local product of $U(1)$ and $SU(2^n)$. The SRBB provides a recursive basis for the Lie algebra $\mathfrak{su}(2^n)$.
   • Crucially, the authors restrict the variational ansatz to the diagonal sub-algebra (the $Z$-factor) of the SRBB. While the full SRBB includes off-diagonal components ($\Psi$ and $\Phi$ factors) necessary for general unitary synthesis, the diagonal component alone is sufficient to parameterize diagonal unitary matrices.
   • By utilizing the identity $R_y(\gamma) = SH \cdot R_z(\gamma) \cdot HS^\dagger$, the authors map the traditional "ladder" structure of QSP (composed of Uniformly Controlled Gates or UCGs) onto a circuit composed entirely of $S$, $H$, and $R_z$ gates. This allows the diagonal SRBB $Z$-factor to approximate the moduli of the target state amplitudes, while a separate block handles the phases.

2. Algorithm Architecture:

   • The proposed Quantum Neural Network (QNN) operates in two distinct training stages to avoid optimization conflicts:
     • Modulus Learning: The first part of the circuit approximates the real amplitudes (moduli) of the target state using the diagonal SRBB $Z$-factor.
     • Phase Learning: The second part approximates the complex phases of the target state.
   • The circuit depth is optimized by exploiting the recursive structure of the SRBB $Z$-factor, which utilizes a Gray Code pattern to minimize CNOT gates. The resulting depth scales as $D(n) \approx 6 \cdot 2^n - n^2/2 + 7n/2 - 3$, which is competitive with asymptotically optimal bounds achieved by other methods but without ancillary qubits.

3. Optimization and Loss Functions:

   • The algorithm is tested using two optimizers: Adam (used with Fidelity and Trace Distance loss functions) and Nelder-Mead (used with Frobenius Norm).
   • The training minimizes the distance between the state produced by the VQC and the target state (or the ideal unitary matrix).

Key Contributions
The paper makes four primary contributions:

1. Novel Algorithm: It proposes a specific use case of the SRBB hierarchy for approximate QSP, framing it as a variational problem.
2. Depth Reduction: It implements the traditional QSP ladder structure using only the CNOT-optimized diagonal SRBB sub-algebra. This reduces the circuit depth and CNOT count by an exponential factor compared to using the full algebra, approaching asymptotically optimal bounds without ancillary qubits.
3. Comprehensive Evaluation: The performance is assessed across random states, sparse states, and uniform probability distributions for systems ranging from 2 to 8 qubits in simulation.
4. Real-Device Validation: The algorithm is tested on real quantum hardware (IQM Deneb, IQM Garnet, Rigetti Ankaa-3, and IonQ Forte1) to evaluate performance in noisy environments.

Results

• Simulation:

  • For up to 4 qubits, the algorithm achieves high accuracy with errors ranging from $10^{-13}$ to $10^{-15}$ using the Nelder-Mead optimizer.
  • The Adam optimizer shows longer training times for small qubit counts but does not scale exponentially in time, maintaining high accuracy for up to 4 qubits.
  • Performance degrades as qubit count increases beyond 4; for 8 qubits, errors rise to approximately $10^{-3}$.
  • The network successfully reproduces specific states like Bell and GHZ states with high fidelity in simulation.

• Real Hardware:

  • On real devices (2–5 qubits), the algorithm achieves good results for small qubit counts, though performance varies significantly based on the target state.
  • For specific states like uniform superpositions and Bell states, the Hellinger distance is relatively low (e.g., $\sim 0.01$ to $0.12$ for 2 qubits).
  • For sparse states or more complex distributions, the error increases significantly (e.g., Hellinger distance $\sim 0.74$ for a 4-qubit sparse state on IQM Deneb).
  • Increasing the number of shots (from 1024 to 8192) on Rigetti devices did not yield meaningful improvements in performance, suggesting that noise and circuit depth are the limiting factors rather than shot noise.

Significance and Claims
The authors claim that this work provides a new scheme for approximate quantum state preparation within a variational framework that explicitly links variational parameters to the geometry of unitary groups. The primary significance lies in the ability to reduce circuit resources (CNOTs and depth) by exploiting the diagonal sub-algebra of the SRBB, thereby adhering to the "relaxation principle" of minimizing resources while maintaining accuracy.

The paper modestly acknowledges current limitations:

• Scalability: The algorithm faces challenges with increasing numbers of qubits, where the optimizer struggles to converge effectively, likely due to the complexity of the parametric landscape (potential Barren Plateaus).
• Hardware Noise: On real devices, the error is not fixed but varies significantly depending on the specific target state, indicating sensitivity to noise and state complexity.
• Future Work: The authors suggest that future research should compare this framework against established QSP algorithms and encoding techniques (like Amplitude Encoding), optimize rotation gates further, and analyze the optimization landscape to mitigate convergence issues in larger systems.

The work does not claim to solve the QSP problem for arbitrary large-scale systems immediately but demonstrates a viable, resource-efficient approach for small-to-moderate qubit counts that bridges Lie algebra theory with practical variational quantum circuits.