⚛️ quantum physics

From Projected Subspaces to Full-Space Implementations: Representation-Equivalence Audits for Open-Shell ADAPT-VQE

This paper establishes a representation-equivalence audit framework demonstrating that variational states optimized in symmetry-projected spaces often fail to correspond to their full-space implementations due to representation mismatches, necessitating spin-preserving generalized constructions or circuit-level validation to ensure fidelity and energy accuracy in open-shell ADAPT-VQE.

Original authors: Lucia Malíčková, Petr Klenovský

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lucia Malíčková, Petr Klenovský

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to simulate the behavior of atoms and molecules on a quantum computer, scientists face a fundamental challenge: how to describe the complex dance of electrons without overwhelming the machine. One popular strategy involves building a mathematical model of the molecule in a simplified, reduced space where the most important interactions are kept, while less critical details are temporarily set aside. This approach, known as a projected subspace method, allows researchers to find a low-energy state for the system much faster. The hope has been that once this efficient solution is found in the simplified world, it can be directly translated back into the full, complex reality of the molecule to be run on actual quantum hardware. The assumption is that the shortcut taken in the math does not change the physical outcome.

However, a new study reveals that this translation step is far more treacherous than previously thought. The researchers found that a solution found in a simplified, symmetry-projected space is not automatically the same as the solution one would get by running the exact same instructions in the full, unprojected space. In fact, for certain types of chemical systems, taking the parameters from the simplified calculation and applying them directly to the full system results in a catastrophic failure, producing a state that is physically wrong and chemically useless. The study does not discard the idea of using simplified spaces; instead, it establishes a strict new rule: before any result from a reduced-space calculation is trusted as a full-space solution, it must pass a specific audit to prove that the two spaces are truly equivalent.

The team, led by Lucia Malíčková and Petr Klenovský, tested this idea using a model of a copper ion, a common component in biological systems and industrial catalysts. They first ran a standard adaptive algorithm in a reduced space that preserved the correct spin properties of the electrons. This calculation worked perfectly, finding a state with an energy error of just 1.57 milli-Hartrees, a unit of energy used in quantum chemistry. This result was considered a success within the simplified framework. The researchers then took the exact same sequence of mathematical operations and the exact same numerical values for the steps, but applied them to the full, unprojected space of the molecule. The result was a disaster. The energy error skyrocketed to 464.73 milli-Hartrees, and the resulting state had almost no resemblance to the correct one, with a fidelity of only 0.07. In simple terms, the instructions that worked in the simplified room led the system completely astray when applied to the full building.

The investigation showed that this failure was not due to a single bad step but was a systemic issue. Out of 323 possible types of electron movements used in the calculation, 274 of them violated a specific mathematical condition required to keep the system within the correct spin state when moved between spaces. When the researchers tried to fix the problem by re-optimizing the numbers in the full space while forcing the system to stay close to the correct spin state, the error improved significantly, dropping to 8.48 milli-Hartrees. This proved that the original failure was primarily a problem of transferring parameters from one mathematical representation to another, rather than the sequence of steps being inherently incapable of solving the problem. However, even with this re-optimization, the full-space sequence could not quite reach the high precision of the original reduced-space result, indicating that the two approaches are fundamentally different ways of constructing the solution.

To confirm that this was a general issue and not just a quirk of the copper model, the team tested a nitrogen monoxide radical. Here, the effect was less severe but still present. The reduced-space calculation found a good solution, but applying the same numbers to the full space resulted in a much larger error, though the system remained closer to the correct state than in the copper case. This confirmed that the mechanism of failure exists across different molecules, but its severity depends on the specific system and the order of the steps taken. The researchers also demonstrated that if one builds the full-space calculation from the start using a pool of operators that are guaranteed to preserve the spin symmetry, the reduced-space and full-space results match perfectly. This "positive control" showed that the problem is solvable, but it requires a different construction method than simply translating a reduced-space result.

The study also addressed the practical side of running these simulations on real quantum hardware. Even if the mathematical state is correct, turning it into a physical circuit involves approximations, such as breaking down complex operations into simpler gates. The team showed that by carefully selecting which parts of the circuit needed more precision and which could be simplified, they could reduce the number of quantum gates required by nearly 70 percent while still maintaining the accuracy needed to pass a strict validation test. This highlights that the process of building the circuit and the process of measuring the final energy are separate layers of validation, each with its own potential for error.

Ultimately, this work serves as a crucial validation framework rather than a new algorithm for solving chemical problems. It warns researchers that a low energy score in a reduced-space calculation does not guarantee that the corresponding full-space circuit will produce the same state. The authors conclude that whenever a reduced-space optimization is connected to a full-space implementation, the optimized state, the parent sequence of operations, the compiled circuit, and the measurement protocol must be validated as distinct objects. The low energy of the simplified model alone is not enough to certify that the full-space circuit implements the same physical reality. This finding ensures that future simulations of complex molecules, from copper proteins to new materials, are built on a foundation of verified equivalence rather than unproven assumptions.

Technical Summary: Representation-Equivalence Audits for Open-Shell ADAPT-VQE

Problem Statement
The paper addresses a critical, often overlooked ambiguity in Variational Quantum Eigensolvers (VQE), specifically within the ADAPT-VQE framework for open-shell systems. When an adaptive ansatz is optimized within a symmetry-projected subspace (e.g., a specific spin sector like a doublet), the resulting state is defined by a sequence of projected operators. A common operational assumption is that the corresponding "parent" sequence of unprojected fermionic generators, when exponentiated in the full Hilbert space with the same amplitudes, will yield the identical variational state.

The authors demonstrate that this assumption is mathematically false unless the parent generators strictly preserve the target subspace. For an orthogonal projector $P$ onto the target space and its complement $Q = I - P$ , exact equivalence for all amplitudes requires the condition $QAP = 0$ (where $A$ is the anti-Hermitian generator). If this condition is violated, the reduced-space evolution $U e^{\theta A_D} U^\dagger$ is not identical to the compressed full-space evolution $P e^{\theta A} P$ . Consequently, a low energy obtained in the projected space does not guarantee that the corresponding full-space circuit implements the same state, nor does it certify the resource estimates derived from the parent sequence.

Methodology
The authors formulate a "representation-equivalence audit" to detect and quantify this mismatch. The core methodology involves:

Theoretical Audit: Establishing that $QAP = 0$ is the necessary and sufficient condition for the lifted reduced-space unitary to equal the compressed full-space evolution for arbitrary amplitudes. They derive finite-amplitude bounds showing that leakage scales linearly with amplitude ( $\theta$ ) while the mismatch in the target sector scales quadratically.
Benchmarking:
- Primary Benchmark: A canonical Type-1 Cu(II) model (CAS(15e,9o), 18 qubits) representing a "blue-copper" active site.
- Secondary Benchmarks: An NO radical (CAS(7e,6o), 12 qubits), an OH radical, and homogeneous Hubbard chains.
Comparative Experiments:
- Negative Control: Running ADAPT-VQE in the projected doublet space, then applying the exact same 121 selected parent generators and amplitudes in the full fixed- $M_S$ space.
- Reoptimization: Re-optimizing the amplitudes of the same fixed parent sequence in the full space under a hard constraint ( $p_D \ge 0.999$ ) to distinguish between parameter-transfer failure and intrinsic sequence failure.
- Positive Control: Constructing a "generalized spin-preserving" pool where generators are explicitly filtered to satisfy $[A, \hat{S}^2] \approx 0$ and $QAP \approx 0$ , ensuring representation faithfulness.
Circuit Validation: Synthesizing the ansätze into qubit circuits using Suzuki-Trotter factorization and Qiskit transpilation. The authors introduce a "circuit-equivalence gate" to ensure that the compiled circuit error (synthesis + transpilation) remains below a threshold (0.1 mEh) relative to the fermionic target.

Key Results

Catastrophic Representation Failure: In the Cu(II) benchmark, 274 of 323 bare Hartree–Fock excitations violate the invariance condition.
- The projected ADAPT sequence achieves an error of 1.5719 mEh.
- Applying the same 121 amplitudes to the unprojected parent sequence in full space yields a massive error of 464.730 mEh, with a fidelity of only 0.073 to the projected state and a quartet weight of 54.45%.
- Post-projecting this failed full-space state does not recover the target; the error remains ~465 mEh.
Parameter Transfer vs. Sequence Capability: Re-optimizing the 121 amplitudes in the full space (constrained to $p_D \ge 0.999$ ) reduces the error to 8.482 mEh and increases fidelity to 0.979. This indicates the catastrophic failure is primarily a representation/parameter-transfer failure (the projected amplitudes do not transfer to the full space), but the fixed-order sequence still cannot fully recover the 1.6 mEh projected result even after reoptimization.
System Dependence: The NO radical benchmark exhibits the same mechanism but less severely (24.5 mEh error vs. 1.4 mEh projected), with post-projection nearly repairing the state. This highlights that the severity of the failure is system and sequence-dependent.
Positive Control Success: Using a generalized spin-preserving pool (where generators are constructed to preserve the target sector), the authors achieve errors of 1.493 mEh (18 qubits) and 1.181 mEh (24 qubits) with unit doublet weight. In these cases, the full-space and reduced-space evolutions agree to machine precision.
Circuit Synthesis: For the 18-qubit positive control, a selectively refined Suzuki-2 schedule (varying repetition counts for specific factors) reduced the circuit depth and CX count by ~68% compared to a uniform schedule, while passing strict fidelity and energy-error gates.

Claims and Significance
The authors explicitly state that their contribution is not a new spin-adapted ADAPT algorithm, a new projector, or a claim of chemical accuracy for the specific molecular models tested. Instead, the significance lies in:

A Validation Framework: Providing a rigorous "representation-equivalence audit" to verify that a reduced-space optimization corresponds to a specific parent full-space implementation.
Decoupling Validation Layers: Demonstrating that reduced-space energy, parent fermionic sequences, compiled circuits, and measurement protocols are distinct objects. A low reduced-space energy alone does not certify that a quoted full-space circuit implements the same variational state.
Clarifying Limitations: Showing that "projection" in the optimization phase does not automatically imply that the unprojected parent sequence is a valid full-space implementation without an explicit equivalence check.

The paper concludes that while projected pools are legitimate reduced-space methods, associating them with full-space resource estimates or compiled circuits requires a separate verification step to ensure the generators preserve the target subspace. Without this check, the mapping from reduced-space parameters to full-space states is fundamentally ambiguous and can lead to catastrophic errors.

More like this