Expressibility, Noise, and Error Mitigation in VQE Ansatz Selection

This study demonstrates that while standard expressibility predicts Variational Quantum Eigensolver performance under ideal and zero-noise extrapolation conditions, it fails to do so under realistic noise or probabilistic error cancellation, necessitating the use of computationally efficient topology metrics like gate count to forecast error mitigation outcomes.

The Gist

Imagine you are trying to bake the perfect chocolate cake (finding the lowest energy state of a molecule) using a very new, slightly glitchy oven (a near-term quantum computer). You have a recipe book full of different ways to mix the ingredients (these are called "ansatz circuits"). The big question is: How do you pick the best recipe before you even turn the oven on?

For a while, scientists thought they had a magic ruler called Expressibility. This ruler measured how "flexible" a recipe was—essentially, how many different cake textures the recipe could theoretically create if the oven were perfect. The idea was: "The more flexible the recipe, the better the cake."

However, this paper (by Annis, Kassem, and Coleman) is like a group of bakers who decided to test this ruler in a real, noisy kitchen. They found that the ruler doesn't work the way we hoped, and they tested some "fix-it" tools to see if they could make the ruler work again.

Here is what they discovered, broken down into simple concepts:

1. The "Magic Ruler" Breaks in a Noisy Kitchen

In a perfect world (an ideal simulation), the "Expressibility" ruler was okay at predicting which recipes would work. But in a real, noisy kitchen (simulating a real quantum computer with errors), the ruler stopped making sense.

• The Analogy: Imagine you have two recipes. Recipe A is very complex and flexible (high expressibility), while Recipe B is simple. In a perfect kitchen, Recipe A wins. But in a noisy kitchen where the oven flickers and the mixer shakes, the complex Recipe A falls apart, and the simple Recipe B actually tastes better.
• The Finding: The "best" recipe in a perfect world often becomes the "worst" recipe in a noisy world. The ranking of recipes completely scrambles.

2. The "Fix-It" Tools Didn't Work as Hoped

Scientists have developed two main tools to fix the noise in the kitchen:

• ZNE (Zero-Noise Extrapolation): This is like baking the cake at 100%, then 150%, then 200% heat, and mathematically guessing what the cake would taste like at 0% heat (perfect conditions).
• PEC (Probabilistic Error Cancellation): This is like adding a special "anti-noise" ingredient to the batter to cancel out the oven's glitches. It requires a lot of extra math and baking attempts.

The Results:

• ZNE was hit-or-miss: It helped some recipes (about 4 out of 12 for the Hydrogen molecule) but made others worse. It didn't magically fix the "Expressibility" ruler to make it work again.
• PEC was a disaster: For almost every recipe they tried, adding the "anti-noise" ingredient made the cake taste worse. It increased the error significantly. The only time it helped was for a recipe that was already failing so badly it couldn't bake a cake at all; the extra math somehow helped it find a path to a decent cake, but that's a rare exception.

3. The "Simple Count" is Better than the "Magic Ruler"

Since the complex "Expressibility" ruler failed, the authors looked for simpler ways to predict which recipes would work. They found that counting the number of two-ingredient mixings (two-qubit gates) was a surprisingly good predictor.

• The Analogy: Instead of measuring the theoretical flexibility of the recipe, they just counted how many times the mixer had to spin. They found that the more times the mixer spins, the more likely the noise will ruin the cake.
• The Finding: For the "fix-it" tool PEC, simply counting the mixings was almost perfect at predicting failure. If a recipe had too many mixings, PEC would break it. For the other conditions, the simple count was just as good as, or better than, the complex ruler.

4. The "Noisy Ruler" is Too Expensive to Use

The authors tried to create a new version of the ruler that accounts for the noise (called "Noisy Expressibility").

• The Finding: This new ruler did predict the results very well for the small molecules they tested. However, calculating this ruler requires simulating the entire kitchen with all the noise, which is like trying to solve a puzzle that gets exponentially harder with every extra ingredient.
• The Catch: For larger molecules (like Lithium Hydride), calculating this "Noisy Ruler" is so computationally expensive that it's practically impossible to do. It's like trying to calculate the perfect cake recipe for a banquet by simulating every single crumb of flour; you run out of time and computer power before you finish.

5. The "Flexibility" Trap at Large Scales

Finally, they looked at bigger molecules (12 to 14 qubits). They found that as recipes get bigger, they all start looking the same on the "Expressibility" ruler.

• The Analogy: Imagine you have a tiny Lego set and a giant Lego castle. The ruler says the castle is "infinitely flexible." But because the castle is so huge and complex, it's impossible to build it without it falling apart. The ruler loses its ability to tell the difference between a "good" big castle and a "bad" big castle because they all look equally "flexible" on paper.
• The Finding: The ruler stops being useful for large systems because it can't distinguish between good and bad designs anymore.

The Bottom Line for the Baker

If you are trying to pick a recipe (ansatz) for a noisy quantum computer:

1. Don't rely on the "Expressibility" ruler alone; it often lies in noisy conditions.
2. Don't expect the "Fix-It" tools (especially PEC) to save a bad recipe; they often make things worse.
3. Do count your mixings: The simplest way to predict success is to look for recipes with fewer complex steps (fewer two-qubit gates).
4. Keep it simple: For now, the best strategy is to filter out the overly complex recipes first, then use the standard ruler to pick the best of the remaining simple ones.

The paper concludes that there is no single "magic metric" that works for everything. The best approach is a practical, tiered strategy: first, avoid circuits that are too complex for the noisy hardware, and then use simple, cheap-to-calculate metrics to make the final choice.

Technical Summary

Technical Summary: Expressibility, Noise, and Error Mitigation in VQE Ansatz Selection

Problem Statement

The Variational Quantum Eigensolver (VQE) is a leading algorithm for near-term quantum chemistry, yet selecting an optimal ansatz circuit remains a significant challenge. While "expressibility"—a metric quantifying a circuit's ability to uniformly sample the Hilbert space—has been proposed as a guide for ansatz selection, recent work by Saib et al. demonstrated that it fails to predict VQE performance under realistic noise conditions for the hydrogen molecule ($H_2$). Furthermore, it remains unclear whether modern Quantum Error Mitigation (QEM) techniques, specifically Zero-Noise Extrapolation (ZNE) and Probabilistic Error Cancellation (PEC), can restore the predictive power of expressibility or if they introduce new dependencies between circuit properties and performance.

Methodology

The authors conducted a systematic investigation across four molecules: $H_2$ (4 qubits, 12 circuits), $H_3^+$ (6 qubits, 6 circuits), $LiH$ (12 qubits, 4 circuits), and $BeH_2$ (14 qubits, 2 circuits). The study utilized 24 hardware-efficient ansatz circuits with varying single-qubit rotations, two-qubit gate types (CNOT, CZ), and entanglement patterns.

Experiments were simulated using Qiskit Aer with three distinct noise models:

1. COMBINED: Depolarizing error combined with thermal relaxation (matching Saib et al.'s methodology).
2. DEPOL: Pure depolarizing error.
3. AMPLITUDE_DAMPING: Pure amplitude damping.

Four execution scenarios were evaluated for each circuit:

• Ideal: Noiseless simulation.
• Noisy: Simulation with noise, no mitigation.
• Noisy + ZNE: Post-processing using Richardson extrapolation with noise scaling factors $\{1.0, 1.5, 2.0\}$.
• Noisy + PEC: Post-processing using quasi-probability sampling with a depolarizing noise representation.

Performance was quantified by the energy error ($\Delta E = |E_{VQE} - E_{exact}|$). The study compared standard expressibility (computed via ideal statevector simulations) and noisy expressibility (computed via density matrix simulations) against zero-cost topology metrics, including two-qubit gate count, circuit depth, and expected infidelity.

Key Contributions and Results

1. Ranking Instability and Mitigation Efficacy

The study reproduces Saib et al.'s finding that circuit rankings scramble under noise. The Spearman rank correlation between ideal and noisy rankings was negligible ($\rho \approx -0.1$ for $H_2$).

• ZNE: Preserves noisy rankings ($\rho = +0.80$ for $H_2$) but offers inconsistent error reduction. It reduced error for only 4 of 12 $H_2$ circuits and 4 of 6 $H_3^+$ circuits.
• PEC: Actively reorders circuit rankings ($\rho = -0.22$ for $H_2$) and predominantly degrades performance. PEC increased error in 11 of 12 $H_2$ circuits and all 6 $H_3^+$ circuits. The degradation is driven by the exponential sampling overhead ($\gamma^{2N_g}$), not by noise model mismatch, as confirmed by ablation studies under pure depolarizing noise.

2. Predictive Power of Expressibility

• Standard Expressibility: Shows moderate-to-strong correlation with noisy and ZNE performance for $H_2$ ($r = +0.74$ and $r = +0.77$, respectively) across all noise models. However, it shows no significant predictive power for $H_3^+$ or larger molecules.
• Noisy Expressibility: Computed via density matrix simulations, this metric strongly predicts unmitigated performance for $H_3^+$ ($r = +0.91$). However, the authors argue this metric is computationally intractable at scale ($O(4^n)$) and theoretically conflates state-space coverage with noise-induced purity loss.
• Topology Metrics: Zero-cost metrics, particularly two-qubit gate count, provide comparable or superior predictive power for PEC degradation ($r = +0.96$ for $H_3^+$) and serve as effective proxies for noise susceptibility.

3. Scaling Limitations

In a preliminary scaling study for $LiH$ (12 qubits) and $BeH_2$ (14 qubits), standard expressibility values collapsed toward zero as circuits approached 2-designs. This "expressibility collapse" renders the metric non-discriminative at larger scales, as all circuits appear equally expressive despite varying significantly in VQE performance. Additionally, noisy simulations for these larger systems were deemed computationally intractable.

Significance and Claims

The paper claims that the hypothesis—that error mitigation restores expressibility as a universal ansatz selection metric—is unsupported. Instead, the authors establish that:

1. Mitigation is Condition-Dependent: QEM techniques do not universally improve performance; PEC often degrades it due to sampling overhead, while ZNE success depends heavily on the noise model and circuit structure.
2. Contextual Predictors: No single metric dominates across all conditions. Standard expressibility is a useful predictor for $H_2$ under noisy and ZNE conditions, but topology metrics (gate count) are superior for predicting PEC degradation.
3. Scalability Constraints: Expressibility loses discriminative power as qubit counts increase (approaching 2-designs), and noisy expressibility becomes computationally infeasible beyond $\sim10$ qubits.

The authors conclude that for practitioners, a tiered strategy is most effective: first filter circuits by gate count to eliminate noise-vulnerable candidates, then use standard expressibility to rank the remaining options. The study emphasizes that while noisy expressibility correlates with error, it often acts as a proxy for noise accumulation rather than a pure measure of representational capacity, making simple topology metrics a more practical choice for ansatz selection in the NISQ era.