Quantum error mitigation using energy sampling and extrapolation enhanced Clifford data regression

This paper enhances Clifford Data Regression for error mitigation in quantum chemistry simulations by introducing Energy Sampling to optimize training circuit selection and Non-Clifford Extrapolation to better model noise evolution, both of which outperform the original method in noisy VQE experiments.

The Gist

Imagine you are trying to bake a perfect chocolate cake (the perfect answer) in a kitchen that is currently under construction. The ovens are flickering, the measuring cups are slightly bent, and the flour is a bit damp. This is the current state of quantum computers: they are powerful, but they are "noisy" and prone to making mistakes.

This paper is about a new set of techniques to help us get a delicious cake out of this messy kitchen, even before we can build the perfect, noise-free ovens of the future.

Here is the story of how the authors fixed the recipe, explained simply.

The Problem: The "Noisy" Kitchen

The authors are using a method called VQE (Variational Quantum Eigensolver) to calculate the energy of a molecule (specifically, a tiny molecule called H4). Think of this as trying to find the exact temperature needed to bake the cake.

Because the quantum computer is noisy, the temperature it reads is wrong. If you just trust the machine, your cake will be burnt or raw. We need a way to correct the reading without buying a new, perfect oven (which is too expensive and hard to build right now).

The Old Solution: The "Clifford Data Regression" (CDR)

The authors started with an existing trick called CDR. Here is how it works, using a metaphor:

Imagine you have a broken thermometer that always reads 5 degrees too high. To fix it, you don't throw it away. Instead, you:

1. Take a bunch of water samples where you know the exact temperature (because you can calculate them perfectly on a regular computer).
2. Measure those same samples with your broken thermometer.
3. You draw a line on a graph connecting the "True Temp" to the "Broken Temp."
4. Now, when you measure a new, unknown sample with the broken thermometer, you use that line to guess what the real temperature must be.

In quantum terms, the "samples" are simple circuits made of "Clifford gates" (which are easy for regular computers to simulate). The "broken thermometer" is the noisy quantum computer.

The New Improvements

The authors realized the old method had some flaws. Sometimes the line you draw isn't quite right, or you picked the wrong water samples. They proposed two upgrades:

1. Energy Sampling (ES): "Picking the Best Candidates"

The Analogy:
Imagine you are trying to guess the height of a giant mountain. You ask 100 people to guess.

• Old Method: You ask 100 random people. Some are guessing a hill, some are guessing a skyscraper. You average their answers.
• New Method (Energy Sampling): You ask 1,000 people, but you only keep the answers from the 100 people who guessed the lowest heights (closest to the ground). Why? Because you know the mountain is tall, but you want to focus on the people who are thinking about the "base" of the problem, not the clouds.

In the Paper:
The authors simulate thousands of "near-perfect" circuits on a regular computer. Instead of using all of them to train their correction model, they filter them. They only pick the ones that have the lowest energy (the most promising candidates).

• Result: By training the model only on the "best" data, the correction becomes much sharper. It's like training a student only with the best examples, rather than random ones.

2. Non-Clifford Extrapolation (NCE): "Teaching the Model to Evolve"

The Analogy:
Imagine you are trying to predict how a car will drive at 100 mph.

• Old Method: You test the car at 10 mph, 15 mph, and 20 mph. You draw a line and guess what happens at 100 mph. But maybe the car behaves totally differently at high speeds! Your guess might be way off.
• New Method (NCE): You tell the computer, "Don't just look at the speed. Look at how the speed changes." You give the model data from 10, 15, 20, 25, and 30 mph. You teach it the pattern of how the car behaves as it speeds up. Then, you ask it to predict 100 mph based on that pattern.

In the Paper:
The "speed" in this analogy is the number of complex, hard-to-simulate parts in the circuit (called Non-Clifford parameters).

• The old method only looked at circuits with a tiny amount of complexity.
• The new method (NCE) looks at circuits with varying amounts of complexity (1, 2, 3, 4... complex parts). It teaches the model how the noise behaves as the circuit gets more complex.
• Result: The model learns the "shape" of the error. When it finally tries to predict the result for the full, complex circuit, it doesn't just guess; it extrapolates based on a trend it has learned.

The Results

The authors tested these ideas on a simulated noisy quantum computer (IBM's "Torino" model).

• The "Bias" (Picking the right data): Simply picking the training data that looks most like the answer (lowest energy) made a huge difference.
• The "Extrapolation" (Learning the trend): The NCE method was the most powerful. It allowed them to predict the correct energy much more accurately than the old method, even when the circuit was very complex.

The Bottom Line

This paper is like a chef saying: "We can't fix the broken oven yet, but if we choose our test ingredients more carefully (Energy Sampling) and teach our correction formula how the ingredients change over time (Non-Clifford Extrapolation), we can bake a much better cake today."

It shows that we don't need perfect quantum computers to get useful results; we just need smarter ways to clean up the noise.

Technical Summary

1. Problem Statement

Quantum computers in the Noisy Intermediate-Scale Quantum (NISQ) era suffer from significant noise and decoherence, which severely limits the accuracy of computational outcomes, particularly for high-precision applications like quantum chemistry. While full quantum error correction (QEC) is currently infeasible due to resource overhead, Quantum Error Mitigation (QEM) offers a practical alternative.

The paper focuses on Clifford Data Regression (CDR), a learning-based QEM method. CDR mitigates errors by training a regression model on "near-Clifford" circuits (circuits that are classically simulatable) to predict the noise-free expectation values of target circuits containing non-Clifford gates. However, the authors identify limitations in the standard CDR approach:

• Training Data Selection: Standard CDR often uses random sampling of near-Clifford circuits, which may not optimally represent the target state's energy landscape.
• Extrapolation Limitations: Standard CDR typically trains on a single fixed number of non-Clifford gates ($k$) and attempts to predict the result for the full circuit ($k=n$). This ignores the evolving relationship between noisy and ideal values as the circuit complexity increases.
• Accuracy: Even with optimized parameters, standard CDR struggles to achieve milliHartree precision for molecular ground state energies.

2. Methodology

The authors investigated these limitations using the Variational Quantum Eigensolver (VQE) with a tiled Unitary Product State (tUPS) Ansatz to simulate the ground state energy of the H4 molecule (4 electrons in 4 orbitals, 8 qubits). Simulations were performed using the IBM Torino noise model (via Qiskit's FakeTorino backend).

The study introduced two specific enhancements to the CDR framework:

A. Energy Sampling (ES)

• Concept: Instead of randomly selecting near-Clifford circuits for the training set, the authors propose classically simulating a large pool of candidates ($M$) and selecting only the $N$ lowest-energy samples (where $N < M$) to form the training dataset.
• Rationale: Since the ground state energy is variationally bound, biasing the training data toward lower energies ensures the regression model samples the region of the Hilbert space closest to the true ground state, improving the relevance of the training data without increasing the quantum hardware cost (since the quantum device only runs the $N$ selected circuits).

B. Non-Clifford Extrapolation (NCE)

• Concept: The authors propose a regression model that treats the number of non-Clifford parameters ($k$) as an additional input feature, rather than fixing $k$ to a single value.
• Implementation: The model is trained on data collected across a range of small $k$ values (e.g., $k=1$ to $k=6$). The regression function is extended to a multivariate polynomial:
  $$f_{NCE}(X_{noisy}, k) = a_1(X_{noisy})^2 + a_2k^2 + a_3kX_{noisy} + a_4X_{noisy} + a_5k + a_6$$
• Rationale: This allows the model to learn the trajectory of how the relationship between noisy and ideal values evolves as the circuit approaches the optimal non-Clifford configuration. The model then extrapolates this learned trend to the target case where $k=n$ (the full circuit).

C. Regression Models

The study compared Linear and Quadratic regression models for both the traditional and enhanced methods.

3. Key Contributions

1. Systematic Hyperparameter Analysis: A comprehensive benchmark of traditional CDR on the H4 system, demonstrating that:
   • Biasing (setting non-Clifford phases to values closest to the target) significantly outperforms zeroing (setting phases to 0).
   • Training Set Size: Convergence is rapid (around $N \approx 50$ samples), with diminishing returns beyond this point.
   • Non-Clifford Count ($k$): Increasing $k$ improves accuracy but incurs exponential classical simulation costs, necessitating a trade-off.
2. Proposal of Energy Sampling (ES): Demonstrated that filtering training data by energy significantly reduces absolute errors, particularly for small training set sizes.
3. Proposal of Non-Clifford Extrapolation (NCE): Introduced a multi-$k$ training strategy that captures the non-linear evolution of noise effects, enabling more accurate extrapolation to the full circuit depth.
4. Resource Scaling Analysis: Clarified that while NCE requires more classical simulation (due to multiple $k$ values), the quantum overhead remains linear in the number of training circuits, similar to standard CDR.

4. Results

• Traditional CDR Baseline:
  • Biasing reduced absolute errors by ~0.05 Ha (2-layer) and ~0.3 Ha (3-layer) compared to unbiased methods.
  • Quadratic models offered negligible improvement over linear models.
  • Even with optimal parameters, errors remained above 0.1 Ha.
• Energy Sampling (ES) Performance:
  • ES significantly outperformed traditional CDR, especially when $N$ was small (halving the error in some cases).
  • Increasing the classical pool size ($M$) while keeping the quantum training set size ($N$) fixed further improved accuracy, as it allowed for better selection of low-energy candidates.
• Non-Clifford Extrapolation (NCE) Performance:
  • 2-Layer Case: NCE outperformed both traditional and ES CDR, reducing absolute error by ~0.13 Ha compared to traditional CDR. It achieved near-zero error for $k \geq 6$.
  • 3-Layer Case: NCE outperformed traditional CDR but showed inconsistent results compared to ES CDR (slightly higher errors in some regimes).
  • Trade-off: NCE requires a larger total number of training samples (across multiple $k$ values) to converge, leading to higher classical computational costs and slower convergence with respect to sample count compared to ES.

5. Significance

This work advances the state-of-the-art in learning-based error mitigation for quantum chemistry:

• Practicality: It offers a pathway to higher accuracy on current NISQ devices without requiring additional qubits or full error correction.
• Strategic Insight: It highlights that the quality of training data (via Energy Sampling) and the structure of the learning model (via Non-Clifford Extrapolation) are as critical as the sheer volume of data.
• Scalability: The findings suggest that Energy Sampling is the most cost-effective immediate improvement for reducing quantum hardware costs, while NCE offers a systematic route to higher precision if classical simulation resources are available.
• Future Directions: The paper sets the stage for combining ES and NCE and exploring more advanced regression models to further push the boundaries of quantum utility in molecular simulations.