⚛️ quantum physics

Certification failure in variational-principle observables

This paper demonstrates that in variational quantum algorithms, a ground-state energy estimate converged to within $\epsilon$ of the true value does not guarantee accurate observable predictions, as the error in observables can be significantly amplified—up to $\sqrt{2\epsilon\chi_O}$ —due to non-linear error amplification, rendering energy convergence an insufficient certification for observable correctness.

Original authors: Ganga Singh Manchanda

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ganga Singh Manchanda

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 understand the building blocks of matter, scientists rely on powerful mathematical tools to predict how atoms and molecules behave. At the heart of many of these tools is a simple, comforting idea: if you can find a state of a system that has the lowest possible energy, you have found the true state of that system. This principle, known as the variational principle, acts as a compass for researchers. It suggests that if a computer algorithm can calculate an energy value that is very close to the true ground energy, the resulting picture of the molecule must also be very close to reality. For decades, this assumption has been the bedrock of chemical simulation, allowing scientists to trust their calculations of properties like magnetism or chemical reactivity simply because the energy numbers looked good.

However, a new study challenges this long-held confidence, revealing a hidden trap in how we judge the quality of these simulations. The research, conducted by Ganga Singh Manchanda at the University of Oxford, demonstrates that a simulation can produce a nearly perfect energy reading while simultaneously generating a completely wrong picture of the molecule's physical properties. The study shows that energy is not always a reliable witness; it is possible for a calculation to be "certified" as successful based on its energy alone, even when the observable features of the system are wildly inaccurate. This discovery forces a rethinking of how scientists validate their models, suggesting that in many complex situations, a low energy error does not guarantee a correct state.

To understand why this happens, one must look at how energy and other properties respond to changes in a system. In the world of quantum mechanics, energy is a very forgiving metric. It is a "second-order" quantity, meaning that small mistakes in the description of a system's state result in even smaller errors in the calculated energy. Other properties, such as the magnetic orientation of a molecule or its total spin, are "first-order" quantities. This means that the same small mistake in the state description can cause a much larger, direct error in these properties. The paper formalizes a mechanism the author calls "non-linear error amplification." Essentially, a tiny, almost invisible error in the energy calculation can be magnified into a massive error in the observable property, especially when the system is in a delicate state where two different configurations are very close in energy.

The researchers tested this theory using three different types of advanced simulation algorithms on three distinct physical systems. In the first case, they simulated a chain of atoms known as a transverse-field Ising chain. They lowered a magnetic field to push the system into a state where the atoms align. When using a standard, flexible algorithm, the simulation found an energy that was very close to the true value. However, the algorithm failed to correctly predict the direction of the magnetism. Because the energy landscape was so flat near the correct state, the algorithm could settle on a solution that was slightly off in its internal structure but still had the right energy. This led to a magnetization value of 0.98, whereas the true ground state has zero magnetization, despite the energy looking perfect.

In a second experiment, the team modeled two short chains of hydrogen atoms pulled apart from each other. As the chains moved further away, the energy gap between their lowest state and the next possible state became incredibly small. The researchers compared two versions of a popular simulation method called density-matrix renormalization group. One version was allowed to explore all possible states, while the other was restricted to only look at states with a specific symmetry. The unrestricted version found an energy that was chemically accurate, yet it predicted that the two chains had a total spin of one, indicating a broken symmetry. The restricted version, which respected the underlying symmetry of the problem, correctly predicted a spin of zero. In this case, the algorithm failed to find the correct singlet state even though a lower-energy singlet solution existed at that separation; only when the chains were pulled even further did the broken-symmetry state become the true lowest-energy state, at which point even a comparison of the two calculations would choose it, yet it remained physically incorrect for the intended singlet state.

The third case involved a molecule of lithium fluoride, where the researchers simulated the effect of an electric field on the bond between the atoms. They used a method called Krylov quantum diagonalization to track how the molecule's dipole moment changed as the bond stretched. When the simulation was run under conditions where the energy gap between two states closed, the algorithm produced a dipole moment that was off by as much as 22 debye units, even though the energy error was small enough to be considered chemically accurate. In contrast, when they adjusted the simulation to keep the energy gap open, the same algorithm produced the correct dipole moment with high precision. This confirmed that the failure was not due to a flaw in the algorithm itself, but rather a fundamental limitation of using energy as the sole measure of success when the energy levels are crowded together.

The study concludes that the reliability of a simulation depends heavily on the specific property being measured and the nature of the system's energy levels. If a system has a large gap between its lowest energy state and the next one, energy is a good guide. But if that gap is small, or if the property being measured is highly sensitive to the mixing of these states, energy becomes a poor proxy. The researchers show that for a simulation to be trusted, one must look beyond the energy number and consider the "susceptibility" of the property in question—a measure of how easily the property changes in response to the system's internal shifts.

This work does not suggest that these powerful simulation tools are useless, but rather that they require a new kind of caution. Scientists can no longer assume that a well-converged energy automatically certifies the correctness of a molecule's behavior. Instead, they must verify that the specific properties they care about are robust against the small uncertainties that inevitably exist in any calculation. The findings serve as a crucial reminder that in the complex dance of quantum particles, the lowest energy does not always tell the whole story, and sometimes, the most dangerous errors are the ones that hide behind a perfect score.

Technical Summary: Certification Failure in Variational-Principle Observables

Problem Statement
Variational algorithms, widely used in both classical and quantum computing to approximate ground states, operate under a fundamental assumption: that a state with an energy converged to within a small resolution $\epsilon$ of the true ground-state energy $E_0$ necessarily implies that the state itself is a good approximation, and thus that its observable expectation values are also well-converged. This paper challenges that assumption, demonstrating that a well-converged energy does not "certify" the correctness of observable expectation values. The author argues that in systems with quasi-degenerate low-lying spectra or small spectral gaps, the error in an observable can be orders of magnitude larger than the energy error. This phenomenon, termed "non-linear error amplification," arises because the variational energy is a second-order function of the state error, whereas most observable expectation values are first-order functions. Consequently, energy alone cannot guarantee the fidelity of an observable, particularly in regimes of physical interest such as bond breaking, phase transitions, and open-shell molecular systems.

Methodology and Theoretical Framework
The paper formalizes the conditions under which certification fails by defining the certification radius $\delta_O(M_\epsilon)$ . This is defined as the supremum of the error in an observable $O$ 's expectation value over the set of all density matrices $M_\epsilon$ that satisfy the energy constraint $E_0 \le \text{Tr}(H\rho) \le E_0 + \epsilon$ .

The author derives rigorous upper and lower bounds for this radius using spectral decomposition and linear response theory, improving upon existing weak bounds that depend solely on the operator norm $\|O\|_\infty$ :

Decomposition of Error: The error in the observable is decomposed into two terms: a "population" term (scaling linearly with $\epsilon$ ) and a "coherence" term (scaling with $\sqrt{\epsilon}$ ).
The Role of Susceptibility: The dominant term for small $\epsilon$ is the coherence term, which depends on the static susceptibility $\chi_O$ of the observable. The author shows that to leading order, the certification radius scales as $\sqrt{2\epsilon\chi_O}$ .
Theorems 1 and 2:
- Theorem 1 provides an upper bound on the certification radius, showing it is governed by the spectral gap, low-lying spectral gaps, and transition matrix elements (encoded in $\chi_O$ ).
- Theorem 2 provides a matching lower bound, proving that the radius is indeed proportional to $\sqrt{2\epsilon\chi_O}$ when the susceptibility is non-zero.
Symmetry Considerations: The paper analyzes how symmetry can restore certifiability. If an observable is a conserved quantity, or if the ansatz is confined to a specific symmetry sector (e.g., parity or spin), the coherence term may vanish or the effective gap may increase, thereby reducing the certification radius.

Numerical Studies
To demonstrate the universality of this failure mechanism, the author applies their framework to three distinct systems and algorithms:

Variational Quantum Eigensolver (VQE) on the Transverse-Field Ising Chain:
- System: An open chain of $L=12$ sites. The transverse field $h$ is lowered to close the gap between the ground state and the first excited state (a parity doublet).
- Observation: A hardware-efficient ansatz (not conserving parity) converges to a broken-symmetry state with high energy accuracy but large magnetization error ( $|M| \approx 0.98$ ) when the gap is small. Conversely, a parity-preserving ansatz yields $M=0$ with zero certification radius due to symmetry protection.
- Mechanism: Failure driven by the coherence term due to the vanishing parity gap.
Density-Matrix Renormalization Group (DMRG) on Weakly Coupled Hydrogen Chains:
- System: Two $H_5$ chains separated by distance $d$ . As $d$ increases, the singlet-triplet gap closes exponentially.
- Observation: Standard $S_z$ -conserving DMRG (which allows triplet mixing) fails to capture the total spin $\langle S^2 \rangle$ , settling on a broken-symmetry state with $\langle S^2 \rangle = 1$ despite chemical accuracy in energy. Spin-adapted ( $SU(2)$ ) DMRG remains in the singlet sector and correctly yields $\langle S^2 \rangle = 0$ .
- Mechanism: Failure driven by the population term (linear in $\epsilon$ ) because the observable commutes with the Hamiltonian, but the effective gap seen by the ansatz is the small singlet-triplet splitting.
Krylov Quantum Diagonalisation (KQD) on Lithium Fluoride (LiF):
- System: LiF at the ionic-covalent avoided crossing. An electric field is used to tune the gap between the two lowest states.
- Observation: On the resonance line where the gap closes, the dipole moment $\mu$ becomes uncertifiable; a state with high fidelity (0.995) can have a dipole error of 1.7 Debye. On a detuned line where the gap is held open, the dipole is accurate.
- Mechanism: Failure driven by the coherence term due to the large transition dipole moment between the near-degenerate states.

Key Contributions and Results

Formalization of Certification Failure: The paper provides rigorous mathematical bounds that tighten previous results by incorporating the static susceptibility and transition matrix elements, explicitly deriving the $\sqrt{2\epsilon\chi_O}$ scaling.
Identification of Non-Linear Amplification: It identifies that small energy errors can be amplified non-linearly into large observable errors, a mechanism dependent on the static susceptibility and the structure of the low-lying spectrum.
Symmetry as a Remedy: The work clarifies how symmetry constraints (parity, spin sectors) can act as a "protection" mechanism, effectively increasing the relevant spectral gap or eliminating coherence terms, thereby restoring certifiability.
Asymptotic Implications: The author discusses the asymptotic limit, suggesting that for algorithms where the effective error $\epsilon_{eff}$ does not shrink faster than the closing spectral gap (e.g., $\epsilon_{eff} \lesssim L^{-z}$ ), certification failure is inevitable in the thermodynamic limit.

Significance and Claims
The paper claims that the "certifiability" of variational results is not a universal property of the algorithm or the system's energy, but is intrinsically tied to the specific observable being measured and its susceptibility. The author asserts that achieving "chemical accuracy" in energy is insufficient to guarantee accurate predictions for observables in systems with small spectral gaps.

The significance lies in shifting the burden of validation: rather than relying on energy convergence as a proxy for state quality, practitioners must explicitly consider the spectral gap and the observable's susceptibility. The paper concludes that without additional constraints (such as symmetry adaptation) or specific knowledge of the gap structure, a well-converged energy benchmark cannot guarantee the correctness of observable expectation values. This challenges the standard practice in molecular simulation where energy convergence is often taken as the sole indicator of a successful calculation.

Technical Summary: Certification Failure in Variational-Principle Observables

More like this