⚡ electrical engineering

Statistical Signal Processing for Quantum Error Mitigation

This paper proposes a statistical signal processing framework for quantum error mitigation that combines a filtering stage to remove depolarizing noise with an expectation-maximization algorithm to derive maximum likelihood estimates of noiseless outputs, demonstrating its effectiveness and scalability on both simulated and synthetic NISQ data.

Original authors: Kausthubh Chandramouli, Kelly Mae Allen, Christopher Mori, Dror Baron, Mário A. T. Figueiredo

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kausthubh Chandramouli, Kelly Mae Allen, Christopher Mori, Dror Baron, Mário A. T. Figueiredo

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 current era of quantum computing, machines are powerful but imperfect. They operate in a phase known as the noisy intermediate-scale era, where the devices are large enough to attempt complex calculations but small enough that they cannot yet fully correct their own mistakes. The core challenge is that these machines are incredibly sensitive to their environment. When a quantum circuit runs, the delicate information it holds is easily scrambled by random interference, much like a whisper lost in a crowded room. To get useful results, scientists must use a technique called quantum error mitigation. This is not about fixing the machine itself, but rather about using classical computers to clean up the messy data after the quantum machine has finished its work. The goal is to look at the noisy, corrupted output and figure out what the clean, correct answer was supposed to be.

A team of researchers at North Carolina State University and Instituto Superior Técnico has developed a new way to perform this cleanup, treating the problem as a statistical signal processing task. Instead of trying to predict how the machine will fail, they built a model that assumes the noise behaves in two specific ways. First, they assume that deep circuits often produce results that look like random static, where every possible answer appears with equal likelihood. Second, they account for simple mistakes where a single bit of information flips from zero to one or vice versa during the reading process. By separating these two types of errors, the researchers created a two-step method to recover the true signal.

The first step of their method acts as a filter. Imagine a room filled with people shouting answers to a question. If most of the shouting is just random noise, it is impossible to hear the real answers. The researchers' algorithm scans the thousands of measurements taken by the quantum machine and identifies the patterns that look like this random static. It then discards those unhelpful measurements, leaving behind only the data that carries a hint of structure. This process strips away the overwhelming background noise, making the remaining data much easier to analyze.

Once the random noise is removed, the team applies a second step using a mathematical technique known as expectation-maximization. This is an iterative process where the computer makes an educated guess about the correct answers, checks how well that guess fits the remaining data, and then refines the guess. It repeats this cycle over and over, slowly honing in on the most likely set of true solutions. Unlike some other methods that assume there is only one correct answer, this approach is designed to handle situations where a quantum algorithm might have several valid outcomes. It also does not need to know in advance how many correct answers exist; it figures that out as part of the process.

The researchers tested this approach using simulations and real data from an IBM quantum processor. They ran experiments on systems with up to fourteen qubits, which are the basic units of quantum information. In these tests, the method proved highly effective, recovering the correct answers with very few errors. When compared to other existing statistical techniques for cleaning up quantum data, their method performed better, achieving near-perfect accuracy in many cases. The team also tested the limits of their approach by generating synthetic data for a much larger system with 128 qubits. Even in this simulated environment with heavy noise, the algorithm successfully identified the correct solutions, suggesting that the method could scale to the larger machines of the future.

The study highlights that the key to success was the specific way the noise was modeled. By acknowledging that deep circuits often produce a uniform, random distribution of errors, the researchers could filter out the worst interference before trying to solve the puzzle. They found that with enough measurements, the algorithm could determine the correct number of solutions and identify them with high precision. However, the researchers are careful to note that their work is not a final solution for all quantum problems. Their method assumes that the errors are symmetric and does not yet account for more complex interactions between qubits that occur in real hardware. Additionally, the current implementation relies on a specific type of noise model that works well for certain circuits but may need adjustment for others.

Despite these limitations, the results offer a promising path forward. The team demonstrated that by applying principles from classical signal processing to quantum data, it is possible to extract reliable information from very noisy machines. Their work suggests that we do not necessarily need to wait for perfect, fault-tolerant quantum computers to get useful results. Instead, with the right statistical tools, we can make the noisy machines we have today much more capable. The findings indicate that principled statistical methods can provide scalable and interpretable solutions for error mitigation, offering a practical way to improve the reliability of quantum computing in the near term.

Technical Summary: Statistical Signal Processing for Quantum Error Mitigation

Problem Formulation
In the Noisy Intermediate-Scale Quantum (NISQ) era, quantum circuits suffer from noise that corrupts measurement outcomes, necessitating Quantum Error Mitigation (QEM) to recover reliable results. This paper addresses the challenge of estimating the most likely noiseless output states from noisy quantum measurements. The authors model the noise environment as a combination of two primary components:

Depolarizing Noise: A worst-case scenario where the quantum state is replaced by a maximally mixed state with probability $p$ . This results in measurement outcomes that are uniformly distributed across all possible bit strings, carrying little exploitable structure.
Readout Bit-Flip Errors: A classical noise process where measured bits flip (0 to 1 or 1 to 0) with probability $\epsilon$ .

The goal is to recover a set of $K$ unknown valid noiseless binary vectors ( $X = \{x_1, \dots, x_K\}$ ) and their associated mixing weights ( $\alpha$ ) from a dataset of $S$ noisy shots ( $Y$ ), assuming the underlying errors follow the depolarizing and bit-flip model.

Methodology
The proposed approach treats QEM as a statistical signal processing problem, utilizing a two-stage pipeline:

Filtering Stage (Depolarization Removal):
Before applying statistical estimation, the authors employ a density-based filtering algorithm (Algorithm 1) to discard shots consistent with depolarizing noise. The method counts the frequency of every observed bit string and its 1-Hamming distance neighbors. By establishing a threshold $T$ based on the expected frequency of uniform noise ( $\lambda = S/2^n$ ), the algorithm suppresses low-support entries. This step effectively removes the "background" of uniformly distributed noise by discarding the shots that resemble it, leaving a dataset dominated by bit-flip errors, which are more amenable to statistical modeling.
Expectation-Maximization (EM) Algorithm:
On the filtered data, the authors apply an Expectation-Maximization algorithm to compute the Maximum Likelihood (ML) estimate of the noiseless outputs.
- Model: The problem is framed as a finite mixture model where latent variables indicate which of the $K$ valid solutions generated a specific noisy observation.
- Likelihood: The likelihood function accounts for independent bit-flip probabilities across qubits.
- Optimization: Since the number of solutions $K$ is unknown and the likelihood function is non-convex, the authors use a modified log-likelihood function based on the Minimum Message Length (MML) criterion. This allows the algorithm to perform "component annihilation," effectively reducing the number of active components ( $K$ ) during the M-step if they do not contribute significantly to the data fit.
- Procedure: The algorithm iterates between an E-step (calculating the posterior probability of each shot belonging to a specific solution) and an M-step (updating the solution vectors, mixing weights, and bit-flip probabilities).

Key Contributions

Statistical Framework: The paper introduces a principled statistical signal processing approach to QEM that is agnostic to specific circuit details, implicitly estimating the density function of the output state.
Hybrid Noise Handling: Unlike methods that focus solely on readout errors or require specific circuit structures, this method addresses both depolarizing gate errors (via filtering) and bit-flip readout errors (via EM).
Scalability: The authors assert that an ideal QEM system must have computational complexity polynomial in the number of qubits to scale to large systems; the proposed method is designed to address this requirement, and experimental results demonstrate its effectiveness on larger qubit counts using synthetic data.
Adaptation of EM: The work adapts the EM algorithm for binary output distributions and integrates MML for model selection (estimating $K$ ), allowing the system to handle multiple valid outputs rather than assuming a single ground truth.

Experimental Results
The authors evaluated the method using Qiskit simulations, data from the IBM Brisbane QPU, and synthetically generated data for larger systems:

Small-Scale Performance: On simulated circuits with $n \in \{10, 12, 14\}$ qubits and depths up to 800, the method achieved a Bit Error Rate (BER) of 0.000 to 0.007 when the number of solutions $K$ was correctly estimated. The probability of incorrectly estimating $K$ ( $P_{Kerror}$ ) was 0.83% across all configurations and converged to 0 with sufficient shots (10,000).
Comparison with State-of-the-Art: On IBM Brisbane hardware, the proposed EM method outperformed existing statistical QEM techniques (HAMMER, M3, and Q-BEEP) in terms of Hellinger fidelity across various benchmark circuits (Bernstein-Vazirani, GHZ, W state, adder) ranging from 3 to 14 qubits.
Large-Scale Scalability: Using synthetically generated data for $n=128$ qubits with $K \in \{2, 4, 8\}$ , the algorithm achieved a BER of 0 in all trials, suggesting the method scales effectively to larger qubit counts under the assumed noise model.

Significance and Claims
The paper claims that principled statistical methods can offer scalable and interpretable solutions for QEM in realistic NISQ settings. The authors assert that their approach is particularly effective because it:

Does not rely on classical simulation of the full circuit (unlike Clifford data regression), making it feasible for larger qubit counts.
Handles scenarios with multiple valid outputs, unlike majority-vote decoding which assumes a single solution.
Natively addresses the challenge of depolarizing noise, which often renders 80–90% of shots in deep circuits appear random, by filtering out the shots that resemble this noise.

Limitations and Future Work
The authors acknowledge several limitations:

The method assumes symmetric bit-flip probabilities, whereas real hardware may exhibit asymmetry.
It does not model complex readout errors like qubit-to-qubit crosstalk.
The assumption of a finite number of solutions ( $K$ ) limits its direct application to variational algorithms (which may have continuous or non-finite solution spaces).
The initialization of the EM algorithm heavily influences the final ML estimate, and the heuristic filtering strategy warrants further rigorous analysis.

The work concludes that while the current implementation is suboptimal in certain aspects, the EM-based framework provides a promising foundation for scalable quantum error mitigation.

Technical Summary: Statistical Signal Processing for Quantum Error Mitigation

More like this