Branch-Resolved Characterization of Feed-Forward Error in Dynamic Teleportation via Classical Choi Shadows

This paper presents a framework for characterizing feed-forward error in dynamic-circuit teleportation across individual measurement branches, demonstrating through experimental validation that the effectiveness of probabilistic readout error mitigation versus post-processing strategies depends critically on the specific measurement readout error of the hardware layout.

The Gist

Imagine you are trying to send a secret message to a friend using a magical, fragile box. This is what quantum teleportation is like: you take a piece of information, break it down, send the instructions, and your friend rebuilds it on their end.

In the world of quantum computers, this process often involves a "mid-circuit measurement." Think of this as opening a small window in the middle of the process to peek at the box. Based on what you see through the window (the measurement), you have to tell your friend exactly how to fix the box they are holding. This instruction is called "feed-forward."

The Problem: The Messy Window
The paper by Mason Edwards and Prabhat Mishra points out a big problem: looking through that window isn't perfect. Sometimes the window is dirty, or the light is bad, and you might misread what's inside. If you misread the signal, you tell your friend to fix the box in the wrong way.

Traditionally, scientists have looked at the average result of thousands of these attempts. They would say, "On average, the box was fixed 80% of the time." But this is like saying, "On average, the weather is nice," without realizing that it's actually pouring rain in one city and sunny in another. The paper argues that we need to look at each specific "branch" (each specific outcome of the measurement) individually to see where the errors are hiding.

The Experiment: Two Different Rooms
To test this, the researchers set up a "teleportation" game on a real quantum computer (IBM's "Fez" processor). They used two different physical setups (layouts) of the computer's chips:

1. The "Noisy Room" (Layout 1): In this setup, the "window" (the measurement tool) was very dirty. It made a lot of mistakes reading the signal.
2. The "Clean Room" (Layout 2): In this setup, the window was very clean and accurate.

They tried three different ways to fix the box after looking through the window:

• Method A (Physical Application): Immediately after looking, they physically turned a knob on the friend's box to fix it.
• Method B (Post-Processing): They didn't touch the box. Instead, they wrote down what the knob should have been, and later, when they analyzed the data, they mentally "re-labeled" the results as if the knob had been turned.
• Method C (PROM Mitigation): A fancy trick where they intentionally shook the window (added random noise) to make the errors more predictable, then used a mathematical "filter" to cancel out the noise and guess the true signal.

The Surprising Twist
The researchers expected that the "Clean Room" would always be better. But they found a surprising reversal:

• In the Noisy Room: The "Physical Application" (Method A) was actually the worst. The dirty window confused the physical knob, making the box worse. However, the fancy "PROM" trick (Method C) worked best. It was so good at cleaning up the messy signal that it produced the highest quality boxes.
• In the Clean Room: The "Physical Application" was still the worst, but this time, the "Post-Processing" (Method B) was the winner. Because the window was already so clean, the fancy PROM trick wasn't needed and actually added a little bit of unnecessary complexity. The simple mental re-labeling worked perfectly.

The "Branch-Resolved" Discovery
The most important takeaway is that if you had just looked at the average of all these results, you would have missed this story. You wouldn't have seen that the "best" method depends entirely on how dirty your measurement window is.

By looking at each specific outcome (each "branch") separately, they could see exactly how much error was introduced by the physical act of fixing the box versus just calculating it later. They found that in the noisy setup, the physical act of fixing the box added a small penalty (about 2-3% error), but in the clean setup, that penalty jumped significantly (about 7% error).

In Summary
This paper built a new "microscope" to look at quantum errors. Instead of just saying "the computer is 80% accurate," they showed that the computer behaves very differently depending on which specific path the data takes and how noisy the measurement tools are. They proved that sometimes, doing nothing physically and just fixing the math later is better, and sometimes, using a special noise-canceling trick is the only way to get a good result. It turns out there is no single "best" way to fix a quantum message; it depends entirely on the condition of the tools you are using.

Technical Summary

1. Problem Statement

Dynamic circuits, which utilize mid-circuit measurements and classical feed-forward, are essential for advanced quantum algorithms (e.g., error correction, teleportation, and adaptive circuits) on Noisy Intermediate-Scale Quantum (NISQ) devices. However, characterizing the errors associated with these operations remains challenging.

• The Gap: Existing characterization methods typically rely on outcome-averaged metrics (e.g., aggregate state fidelity). These methods obscure the specific error structures associated with individual measurement branches.
• The Challenge: Mid-circuit measurements branch the system's evolution into distinct, outcome-defined paths. Measurement noise (readout errors) can cause the system to take a "wrong" branch, leading to incorrect feed-forward corrections. Current methods struggle to isolate and quantify the error penalty of physical feed-forward operations versus post-processing adjustments on a per-branch basis.

2. Methodology

The authors propose a framework to characterize feed-forward error by reconstructing branch-resolved Choi operators using Classical Choi Shadows.

A. Experimental Protocol: Dynamic Teleportation

• Setup: A 6-qubit system is used: a reference qubit ($R$), an input qubit ($A$), three Alice resource qubits ($1, 2, 3$), and a Bob output qubit ($B$).
• Entanglement Resources: Two types of 4-qubit $W_4$ states are used as the shared entanglement resource between Alice and Bob:
  1. Symmetric $W_4$: Maximally entangled across the partition but results in a dephasing channel (less than 1 ebit).
  2. Perfect $W_4$: Constructed such that Bob's reduced state is maximally mixed, enabling unit-fidelity teleportation (logically equivalent to a Bell state).
• Process: Alice performs a Bell measurement on qubits $A$ and $3$. The outcome ($m$) determines a Pauli correction ($U_m$) to be applied to Bob's qubit.

B. Correction Strategies

The study compares three distinct approaches to applying the necessary corrections:

1. Physical Application: The Pauli gate is physically applied to qubit $B$ immediately after measurement.
2. Post-Processing: No physical gate is applied. Instead, the measurement outcome is tracked, and the correction is applied classically to the final measurement results (relabeling).
3. PROM Mitigation (Probabilistic Readout Error Mitigation): Uses Bit-Flip Averaging (BFA) to symmetrize the readout channel. By applying random bit-flips before measurement and using quasi-probability re-weighting (PROM), the method attempts to target the ideal branch despite readout noise.

C. Characterization Framework

• Classical Choi Shadows: Instead of full quantum process tomography (which scales exponentially), the authors use randomized Pauli measurements to construct "snapshot" operators. These are aggregated to estimate the Branch Choi Operators ($\hat{C}_\ell$).
• Branch Labeling:
  • For unmitigated strategies, the label $\ell$ is the reported measurement outcome.
  • For PROM, the label is a corrected effective label ($c = r \oplus u \oplus f$) derived from the calibration and sampling distribution.
• Metrics:
  • Branch Quality ($\hat{q}_\ell$): The overlap between the reconstructed branch state and the ideal state.
  • Feed-Forward Penalty ($\Delta_{FF}$): The difference in branch quality between physical application and post-processing ($\hat{q}_{post} - \hat{q}_{phys}$).
  • PROM Gain ($G_{q,\ell}$): The improvement in branch quality provided by PROM over physical application.

3. Key Contributions

1. Branch-Resolved Framework: Introduced a method to characterize feed-forward errors specifically for individual measurement branches, avoiding the information loss inherent in outcome-averaged analyses.
2. Controlled Resource Comparison: Developed a pair of 4-qubit $W_4$ entanglement resources (Symmetric and Perfect) that compile to identical circuit costs but induce different ideal teleportation channels, allowing for a controlled study of entanglement structure effects.
3. Validation of Shadow Estimators: Experimentally validated the accuracy of the non-mitigated Choi shadow estimators against full branch Choi state tomography on a 2-qubit subsystem.
4. Layout-Dependent Reversal Discovery: Demonstrated that the relative effectiveness of PROM versus post-processing is not absolute but depends heavily on the readout error characteristics of the specific physical qubit layout.

4. Experimental Results

Experiments were conducted on the IBM Fez (156-qubit Heron) processor using two distinct physical qubit layouts with significantly different readout assignment errors (RAE):

• Layout 1 (High RAE): Qubits used for mid-circuit measurement had higher readout error (~5.1%).
• Layout 2 (Low RAE): Qubits used had lower readout error (~0.6%).

Key Findings:

• Validation: The Choi shadow estimators closely matched full tomography results (RMSE < 0.015), confirming the reliability of the shadow approach for larger systems.
• Layout 1 (Noisy Readout):
  • PROM was superior: PROM mitigation produced the highest branch qualities for all branches.
  • Low Penalty: The operational feed-forward penalty (difference between physical and post-processing) was modest (~0.02–0.03).
  • Reasoning: In high-noise environments, the physical application of gates combined with readout errors degrades performance significantly. PROM effectively mitigates the readout noise, outperforming both physical and pure post-processing.
• Layout 2 (Clean Readout):
  • Post-Processing was superior: Post-processing achieved the highest branch qualities, exceeding PROM.
  • High Penalty: The feed-forward penalty increased significantly (~0.07).
  • Reasoning: When readout error is low, the overhead and noise introduced by physically applying correction gates (and the associated BFA/PROM sampling noise) outweigh the benefits. Post-processing avoids these physical overheads entirely.
• Reversal Phenomenon: The study observed a reversal in the relative order of branch qualities between PROM and post-processing when switching from the noisy layout to the clean layout. This highlights that the "best" strategy is context-dependent.

5. Significance

• Granular Error Insight: The framework reveals error structures (branch-specific penalties) that are invisible to standard, outcome-averaged benchmarking.
• Hardware-Aware Optimization: The results demonstrate that mitigation strategies like PROM are not universally superior. Their efficacy is inversely correlated with the quality of the hardware's readout channel.
  • Noisy Hardware: Requires active mitigation (PROM) to compensate for readout errors.
  • Clean Hardware: Passive strategies (Post-processing) are more efficient as they avoid the noise introduced by physical gate operations and probabilistic sampling.
• Scalability: By validating the shadow estimator against tomography, the paper provides a scalable pathway for characterizing complex dynamic circuits on future, larger quantum processors where full tomography is infeasible.

In conclusion, this work establishes that the choice of feed-forward correction strategy must be dynamically adapted to the specific noise profile of the quantum hardware, and that branch-resolved analysis is critical for making these optimizations.