⚛️ quantum physics

Learning error suppression strategies for dynamic quantum circuits

This paper introduces an empirical learning framework that optimizes dynamical decoupling sequences for dynamic quantum circuits, achieving a three-fold reduction in error rates and enabling high-fidelity implementations of complex algorithms like the quantum Fourier transform on up to 20 qubits.

Original authors: Christopher Tong, Liran Shirizly, Edward H. Chen, Derek S. Wang, Bibek Pokharel

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christopher Tong, Liran Shirizly, Edward H. Chen, Derek S. Wang, Bibek Pokharel

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

The Big Picture: The "Noisy Kitchen" Problem

Imagine you are trying to bake a very delicate, multi-layered cake (a quantum algorithm) in a chaotic, noisy kitchen.

In a standard quantum computer, the "chefs" (qubits) work together in a smooth, continuous dance. But in Dynamic Quantum Circuits, the process is more like a high-stakes cooking show where the chefs have to stop, taste the food (measure the qubit), and then immediately shout instructions to the other chefs on what to do next based on that taste (feedforward).

The Problem:
Every time a chef stops to taste the food, the kitchen gets messy.

The Noise: The act of tasting creates a loud noise (measurement) that startles the other chefs who are just waiting (idle qubits).
The Confusion: The instructions shouted out (feedforward) take time to travel, and during that wait, the other chefs might start wobbling or making mistakes because they are sensitive to the noise.
The Result: By the time the cake is finished, it's a bit burnt or lopsided. The "error rate" is too high to make a perfect cake.

The Old Solution: The "One-Size-Fits-All" Noise Canceller

Scientists have known for a while that you can use Dynamical Decoupling (DD) to fix this. Think of DD as a pair of noise-canceling headphones for the chefs. You play a specific rhythm of pulses (beeps) to the idle chefs to keep them steady and ignore the noise.

However, the old method was like giving everyone the exact same pair of headphones regardless of where they are standing in the kitchen.

If Chef A is next to the loud oven, they need heavy-duty headphones.
If Chef B is far away, they need light ones.
If Chef C is next to a specific noisy machine, they need a different rhythm entirely.

The old "theoretical" headphones were designed based on a perfect, ideal kitchen map. But real kitchens are messy, and the noise comes from different directions depending on who is tasting the food at that exact moment. The old headphones didn't work well enough.

The New Solution: "Learning" the Perfect Rhythm

This paper introduces a new approach: Empirical Learning. Instead of guessing the perfect noise-canceling rhythm, the scientists let the computer learn it by trial and error.

Here is how they did it, step-by-step:

1. Breaking the Kitchen into Zones (Motifs)

Instead of trying to fix the whole kitchen at once, they broke the problem down into small, manageable zones.

They realized that the noise isn't random; it has a pattern. If Chef 1 tastes the food, Chef 2 gets a specific type of shock, and Chef 3 gets a different one.
They divided the circuit into small "motifs" (like little sub-kitchens) based on who is measuring and who is waiting.

2. The Genetic Algorithm (The "Evolution" of Headphones)

They used a computer program called a Genetic Algorithm. Think of this as a breeding program for noise-canceling rhythms.

Generation 1: They tried 16 different random rhythms (like different songs) for the idle chefs.
Testing: They ran the experiment. Some rhythms worked okay; others made the cake worse.
Selection: They kept the "best" rhythms (the ones that kept the chefs steady) and "bred" them together to create new, slightly better rhythms.
Repetition: They did this over and over (9 generations). Just like evolution, the "fittest" rhythms survived and became the champions.

3. The Result: Custom-Tailored Headphones

The computer didn't just find a solution; it found the perfect, custom-tailored rhythm for every specific situation in the kitchen.

It learned that when Chef 1 tastes, Chef 2 needs a "Zap-Zap" rhythm.
It learned that when Chef 2 tastes, Chef 3 needs a "Beep-Beep-Beep" rhythm.
It even learned to pause the rhythm exactly when the "shouting" (feedforward) happens, so the chefs don't get confused.

The Proof: Baking the Perfect Cake

To prove this worked, they tried to bake a very difficult cake called the Quantum Fourier Transform (QFT). This is a fundamental recipe used in almost all quantum algorithms.

Without the new method: The cake was a disaster. For larger sizes (more than 14 ingredients), the result was just random noise. It was like trying to bake a cake in a hurricane.
With the new method: They successfully baked the cake with up to 20 ingredients. The result was clear, sharp, and accurate.
The GHZ State Test: They also tested it on a "super-entangled" state (a group of chefs holding hands so tightly that if one moves, they all move). This is the ultimate test of stability. With their learned rhythms, they could see the interference patterns clearly, proving the "team" stayed perfectly synchronized.

Why This Matters

This paper is a game-changer because it stops trying to guess how quantum computers break and starts learning how they break in real-time.

It's Scalable: You don't need a PhD in physics to design the fix; the computer learns it automatically.
It's Practical: It works on real, messy hardware (IBM's quantum computers), not just in theory.
It's Essential for the Future: As we move toward "Fault-Tolerant" quantum computers (computers that can fix their own mistakes), we need to be able to measure and correct errors while the computer is running. This "learning" method is the key to making that possible.

In short: They taught the quantum computer to listen to its own noise and invent its own custom noise-canceling headphones, allowing it to perform complex tasks that were previously impossible.

1. Problem Statement

Dynamic quantum circuits (DQCs) integrate unitary gates with mid-circuit measurements (MCMs) and classical feed-forward (FF) operations. While essential for fault-tolerant quantum computing (e.g., error correction) and efficient algorithms (e.g., Quantum Fourier Transform with Measurement, QFT+M), they introduce unique error sources not present in purely unitary circuits:

Measurement-Induced Errors: MCMs are significantly longer than unitary gates, causing idle qubits to decohere. Furthermore, the measurement process induces coherent errors (e.g., $Z$ -phase errors, crosstalk) on neighboring qubits due to resonator field population and AC Stark shifts.
Spatial and Temporal Heterogeneity: These errors are not uniform; they depend on the specific qubit being measured, the distance to idle qubits, and the specific timing of the feed-forward operations.
Limitations of Existing DD: Conventional Dynamical Decoupling (DD) sequences are often theoretically derived and fixed (e.g., $X\pi X\pi$ ). They fail to capture the complex, device-specific, and measurement-dependent noise landscape, leading to suboptimal error suppression in DQCs.

2. Methodology: Empirical Learning Framework (GADD)

The authors propose an empirical learning framework called GADD (Genetic Algorithm for Dynamical Decoupling) tailored for dynamic circuits. Instead of deriving sequences from first principles, they optimize them directly on hardware using a quantum-classical feedback loop.

Spatiotemporal Partitioning (Motifs):
- The circuit is partitioned into temporal intervals ( $T_i$ ) and qubit subregisters ( $R_j$ ).
- This creates "motifs" ( $M_{ij}$ ), which are subcircuits representing specific measurement operations and their associated idle periods.
- This partitioning acknowledges that noise is localized; errors induced by measuring qubit $A$ on qubit $B$ differ from those on qubit $C$ .
Genetic Algorithm Optimization:
- Utility Function: A utility function is defined based on the performance of the motif (e.g., similarity to the ideal output distribution or survival probability in randomized benchmarking).
- Evolution: A population of DD sequences (composed of pulses from a decoupling group $G$ ) is evolved. High-performing sequences are selected, and new sequences are generated via crossover and mutation.
- Hardware-in-the-Loop: The algorithm executes training circuits on the quantum device (IBM Eagle ibm_kyiv), measures outcomes, calculates utility, and updates the population classically.
Handling Constraints:
- Frequency Collisions: The framework identifies qubit pairs prone to measurement-induced frequency collisions and applies "constrained" sequences (mostly Identity) to avoid skewing the learning process.
- Feed-Forward Timing: The DD sequences are specifically timed to suppress errors during the measurement pulse and the subsequent feed-forward delay ( $\tau_{FF}$ ), ensuring pulses occur immediately before and after the classical control logic.

3. Key Contributions

Motif-Based Learning: Introduction of a spatiotemporal motif structure that allows parallel optimization of DD sequences for specific local error environments, capturing qubit-dependent and layer-dependent noise.
Empirical vs. Theoretical: Demonstration that empirically learned sequences systematically outperform theoretically derived sequences (like MDD and FFDD) which rely on idealized noise models.
Scalability: The method scales to large circuits (up to 20 qubits in the study) without requiring full process tomography or exponential classical resources, as it relies on local motif optimization.
Application to QFT+M: Successful application of learned strategies to the Quantum Fourier Transform with Measurement on connected qubit chains, a regime previously difficult to simulate due to crosstalk.

4. Key Results

The experiments were conducted on IBM's 127-qubit ibm_kyiv superconducting processor.

Benchmarking (MCM-RB & DC-RB):
- Error Reduction: Empirically learned DD sequences achieved a 3-fold reduction in average dynamic circuit error rates compared to no DD.
- Comparison: In Dynamic Circuit Randomized Benchmarking (DC-RB) on 20 qubits, GADD significantly outperformed both "MDD" (standard DD) and "FFDD" (feed-forward aware DD).
  - Z_c1 Block: Mean Error Per Layer (EPL) reduced from 0.12 (No DD) / 0.11 (MDD) / 0.091 (FFDD) to 0.068 (GADD).
  - I_c1 Block: Mean EPL reduced from 0.11 (No DD) / 0.099 (MDD) / 0.074 (FFDD) to 0.059 (GADD).
- GADD successfully suppressed both "always-on" crosstalk (e.g., $ZZ$) and measurement-induced crosstalk, which theoretical sequences failed to address simultaneously.
QFT+M Performance:
- Process Fidelity ( $F_{proc}$ ): On connected 1D chains, GADD enabled non-trivial QFT process fidelity ( $F_{proc} \geq 1\%$ ) up to $N=16$ qubits.
- Scaling: In contrast, standard "staggered" DD ( $X_p X_m$ ) failed to maintain signal above the random guessing threshold ( $2^{-N}$ ) for $N > 14$ . GADD extended the useful range to $N=20$ .
- Hilbert Space: This represents a >64-fold increase in the dimensionality of the Hilbert space where reliable QFT can be executed.
GHZ State Characterization:
- Applied to a 10-qubit entangled GHZ state with a local phase flip.
- Signal-to-Noise Ratio (SNR): GADD achieved an SNR of 15–20, approaching the theoretical bound, compared to significantly noisier results with standard DD.
- This demonstrated the preservation of many-body interference patterns, proving the suppression of spatially correlated measurement-induced errors.

5. Significance

Practical Error Suppression: The work establishes empirical learning as a viable, scalable strategy for error suppression in dynamic circuits, bypassing the need for precise, often unavailable, noise models.
Enabling Fault Tolerance: By effectively suppressing measurement-induced errors, this approach is directly relevant to quantum error correction (QEC) cycles, where repeated syndrome extraction and feed-forward are critical.
Platform Agnostic: While demonstrated on superconducting qubits, the framework relies only on the ability to execute parameterized circuits and collect statistics, making it applicable to trapped-ion and neutral-atom platforms.
Algorithmic Advancement: It unlocks the potential of dynamic algorithms (like QFT+M) on densely connected hardware, moving beyond disjoint qubit selections used in previous demonstrations.

In conclusion, the paper demonstrates that learning error suppression strategies directly from hardware data is superior to theoretical derivation for dynamic circuits, providing a robust pathway toward high-fidelity, large-scale quantum computation involving measurement and feedback.