Compile-once block encodings for masked similarity-transformed effective Hamiltonians

1. Problem Statement

Quantum simulation of electronic structure often requires solving families of closely related problems, such as geometry scans, active-space expansions, or adaptive similarity transformations (e.g., Schrieffer–Wolff downfolding).

• The Bottleneck: Current quantum simulation workflows typically treat each problem instance as unique. Changes in molecular geometry, basis sets, active spaces, or truncation thresholds usually trigger a complete structural recompilation of the quantum circuit. This includes rebuilding multiplexing trees, routing layers, and data-loading structures (e.g., qROM layouts).
• The Consequence: This recompilation overhead dominates wall-clock costs and obscures the separation between the fixed circuit topology (algebraic structure) and the numerical data (coefficients, masks). It hinders the scalability of adaptive workflows like Generator Coordinate Methods (GCIM) or Quantum Subspace Expansion (QSE), where the working subspace grows iteratively.

2. Methodology: The COMPOSER Architecture

The authors introduce COMPOSER (Compile-Once Modular Parametric Oracle for Similarity-Encoded Effective Reduction), an execution architecture designed to decouple circuit topology from numerical data.

A. Core Concept: Compile-Once, Dial-Many

• Compile-Once: The logical two-qubit circuit topology (selector trees, adaptor wiring, and QSP scaffolds) is synthesized once for a fixed orbital pool and qubit register.
• Dial-Many: Updates to geometry, active spaces, or truncation masks are absorbed entirely by re-dialing single-qubit rotation angles and streaming classical coefficients. No structural changes to the two-qubit fabric are required.

B. Mathematical Foundations

1. Rank-One Factorizations:

   • The molecular Hamiltonian ($\hat{H}$) and anti-Hermitian similarity generators ($\hat{\sigma}$) are decomposed into linear combinations of rank-one operators.
   • Hamiltonian: Using pivoted Cholesky decomposition and eigen-decomposition, $\hat{H}$ is expressed as a sum of rank-one bilinear operators ($\hat{a}^\dagger_u \hat{a}_v$) and projected quadratic operators ($\hat{O}^2$). The number of terms scales nearly linearly with system size ($O(N)$) for fixed thresholds.
   • Generators: Coupled-cluster doubles generators ($\hat{T}_2$) are similarly factorized into rank-one pair-excitation ladders using singular value decomposition (SVD).

2. Deterministic Ladder Circuits:

   • The authors construct deterministic, number-conserving ladder unitaries ($U_u$) based on Givens rotations and fermionic swap networks.
   • These ladders prepare specific one- or two-electron states from a pivot injection or act as basis rotations within a fixed particle-number sector.
   • Crucially, the topology of these ladders is fixed; only the rotation angles depend on the specific coefficient vectors ($u, v$).

3. Block Encoding & QSP:

   • Adaptors: Each rank-one term is wrapped in a single-ancilla block encoding using a fixed PREP-SELECT-PREP† template.
   • Similarity Sandwich: The effective Hamiltonian is constructed as $H_{\text{eff}} = P e^{-\hat{\sigma}} \hat{H} e^{\hat{\sigma}} P$.
   • Masking: A classical mask $M^{(m)}$ selects which rank-one terms are active. Switching masks only updates the state-preparation amplitudes in the selector register, leaving the SELECT table and QSP phases untouched.
   • Exponentiation: The exponential of the generator $e^{\hat{\sigma}}$ is implemented via Quantum Signal Processing (QSP). The polynomial degree is determined by the operator norm, not the number of terms, allowing a single QSP ladder to handle various masks.

3. Key Contributions

• Architectural Shift: Moves from instance-specific circuit generation to a topology-invariant execution model. The two-qubit connectivity and routing are compiled once, while instance dependence is isolated to single-qubit parameters.
• Unified Operator Language: Demonstrates that both Hamiltonians and similarity generators can be treated within a unified framework of rank-one ladders, enabling shared adaptor banks.
• Mask-Aware Effective Hamiltonians: Introduces a method to construct similarity-sandwiched effective Hamiltonians where the mask (defining the active space or truncation) is a classical parameter that does not alter the quantum circuit structure.
• Resource Efficiency:
  • Ancilla: Requires only a constant number of signal ancillas and a selector register of width $\lceil \log_2 \ell \rceil$.
  • Depth: The depth scales as $O(d \cdot \ell_{\sigma} \cdot D_{\max} + \text{depth}(W_{\text{sel}}(\hat{H})))$, where $d$ is the QSP degree. Crucially, the Hamiltonian block encoding is invoked only once, avoiding multiplication by $d$.
• Stability Diagnostics: Provides metrics (e.g., weighted Area Under the Curve of subspace overlap) to validate that low-rank excitation manifolds derived from inexpensive methods (MP2) are stable enough to serve as proxies for expensive methods (CCSD) across geometry changes.

4. Numerical Results

The authors validated the architectural assumptions using a benchmark set of molecules (H$_2$ to C$_6$H$_6$) across various basis sets (STO-3G, 6-31G, cc-pVDZ).

• Linear Scaling: The number of rank-one terms ($\ell_H$) required to represent the Hamiltonian exhibits quasi-linear growth with the number of qubits (system size), confirming that a fixed selector width can cover a wide range of molecular complexities.
• Geometry Scan Invariance: In a stress test on H$_2$O (varying bond lengths), the number of required ladder terms varied only slightly (e.g., 94–97 for 6-31G). This confirms that the selector width can be fixed at the maximum value, and geometry changes only require updating single-qubit angles.
• Subspace Stability: The overlap between MP2-derived and CCSD-derived rank-one excitation subspaces remained high (median > 0.85) across different molecules and basis sets. This supports the use of MP2 to guide mask selection and truncation without needing to recompile the circuit topology.
• End-to-End Demonstration: A toy example of a non-orthogonal subspace solve demonstrated that sweeping a continuous generator coordinate (dialing a parameter) could recover exact ground-state energies without changing the circuit topology.

5. Significance and Outlook

• Practical Feasibility: COMPOSER addresses a critical bottleneck in practical quantum chemistry: the cost of recompilation. By freezing the two-qubit fabric, it enables efficient adaptive workflows (like active-space growth or geometry optimization) that were previously computationally prohibitive due to circuit regeneration overhead.
• NISQ and Fault-Tolerant Relevance:
  • For NISQ, it simplifies error mitigation and calibration by reusing the same circuit structure.
  • For Fault-Tolerant Quantum Computing (FTQC), it decouples the logical topology from the synthesis of single-qubit rotations, allowing for predictable resource scaling even as parameters change.
• Future Directions: The framework opens avenues for automated hardware-aware scheduling, systematic error-budget allocation, and extensions to lattice models and open-system dynamics. It establishes a "hardware-native" block-encoding framework that aligns operator algebraic structure with fixed circuit topologies.

In summary, COMPOSER represents a paradigm shift from optimizing individual circuits to optimizing a fixed execution stack that can be dynamically tuned for a family of related quantum chemistry problems, significantly reducing the overhead of adaptive quantum simulation.