Quantifying the Hadamard Resilience Law: Discovery of the Coherence Gap in NISQ-Era Classifiers

This paper reports that while the Hadamard Test Perceptron maintains high accuracy on the IBM Kingston processor despite significant signal collapse, a critical "Coherence Gap" emerges at high feature depths due to coherent phase errors exceeding hardware limits, thereby identifying these errors rather than depolarizing noise as the primary barrier to scaling quantum linear layers.

The Gist

The Big Picture: A Noisy Quantum Classroom

Imagine you are trying to teach a student (the quantum computer) to recognize handwritten numbers (like the digits 0–9 in the MNIST dataset). In a perfect world, the student has a crystal-clear view of the numbers. But in the real world, the "classroom" is incredibly noisy. The lights are flickering, people are shouting, and the student's eyes are blurry.

This paper investigates a specific question: Can this noisy student still get the right answer, even if they can't see the details clearly?

The researchers tested this on a real quantum computer (the "ibm kingston" processor) and discovered two major things: a "Superpower" the computer has, and a "Wall" that stops it from working on big problems.

---

1. The "Kingston Constant": The Signal Shrinkage

First, the researchers looked at how much the noise messed up the data.

• The Analogy: Imagine you are trying to hear a friend whisper a secret across a crowded, loud stadium. The volume of their voice (the signal) gets crushed by the noise.
• The Finding: On the IBM Kingston processor, the "whisper" was crushed by 93%. The signal shrank so much that it looked almost like static. The researchers call this massive shrinkage the "Kingston Constant."
• The Result: Even though the signal was 93% smaller, the computer could still tell the difference between a "1" and a "2." It was like hearing a whisper so faint you couldn't make out the words, but you could still tell who was speaking.

2. The "Hadamard Resilience Law": The Rank-Order Superpower

This is the paper's main discovery. Usually, we think if the signal gets too weak, the computer fails. But this paper found a "law" that says otherwise.

• The Analogy: Imagine a race where the runners are covered in thick fog. You can't see their faces or their exact speed. However, you can still see that Runner A is ahead of Runner B, and Runner B is ahead of Runner C.
• The Finding: The quantum computer uses a trick called the "Hadamard Test." Even though the noise shrinks the numbers (the speed of the runners), it doesn't scramble the order (who is winning).
• The Law: As long as the computer can figure out which number is "winning" (the highest rank), it doesn't matter if the numbers are tiny or huge. This is why the computer still got 93.9% accuracy on the test, even with that 93% signal loss. The computer is "resilient" because it only needs to know the order, not the exact value.

3. The "Coherence Gap": The Invisible Wall

However, the superpower has a limit. The researchers tried to make the problem harder by using more features (making the "fog" thicker and the race longer).

• The Analogy: Imagine the race track gets so long that the runners have to run for hours. Eventually, the fog gets so thick that the runners start tripping over each other, or they get confused about which lane they are in. The order gets scrambled.
• The Finding: When the researchers increased the complexity to 256 features (a deep circuit), the computer suddenly failed.
  • The Simulation: A computer simulation (a "Digital Twin") that only accounted for random noise still worked perfectly.
  • The Real Hardware: The real quantum computer crashed. The accuracy dropped to about 53% (basically guessing like a coin flip).
• The "Coherence Gap": This huge difference between the simulation and the real machine is called the Coherence Gap. It proves that the problem isn't just "random noise" (like static); it's a specific type of "systematic error" (like a broken compass). The quantum bits (qubits) are getting confused about their timing and phase, causing the "order" of the runners to get scrambled.

4. The "Coherence Wall"

The paper identifies a specific point where the computer hits a wall.

• The Analogy: Think of a battery. If you run a small circuit, the battery lasts. If you try to run a massive circuit (like the 256-feature one), the battery dies before the task is finished.
• The Finding: The circuit for the big problem was about 10,000 steps deep, but the IBM Kingston processor can only handle about 3,500 steps before the signal dies out completely.
• The Conclusion: The "Hadamard Resilience Law" works great for small problems, but it hits a "Coherence Wall" when the problem gets too big for the current hardware.

Summary of the "Golden Path"

The researchers found a clever way to prove their theory without running millions of slow tests:

1. They ran a few quick tests to measure exactly how much the "Kingston Constant" shrinks the signal.
2. They used that data to build a "Digital Twin" (a perfect simulation of the noisy machine).
3. They proved that if the only problem was random noise, the computer would work perfectly.
4. Since the real computer failed at the big size, they proved that the real culprit is not random noise, but coherent errors (timing/phase mistakes) that current simulators don't catch.

The Bottom Line

• Good News: Quantum computers are surprisingly tough. They can still classify numbers correctly even when the signal is 93% weaker than it should be, as long as the "order" of the answers stays the same.
• Bad News: They hit a hard wall when the problems get too big (256 features). The hardware isn't stable enough to keep the "order" straight for deep, complex circuits.
• The Fix: To go bigger, we can't just add more noise; we need to fix the "timing" errors (coherence) or split the big problem into smaller pieces that fit on the current hardware.

Technical Summary

Technical Summary: Quantifying the Hadamard Resilience Law and the Discovery of the Coherence Gap

Problem Statement
The transition of quantum machine learning (QML) algorithms from theory to Noisy Intermediate-Scale Quantum (NISQ) hardware is frequently hindered by systemic decoherence. While theoretical models often assume high gate fidelity, physical superconducting processors suffer from signal decay that can render deep quantum circuits ineffective. Specifically, there is a lack of understanding regarding whether the decision boundaries of quantum classifiers can survive the catastrophic signal magnitude collapse observed on current hardware. This paper investigates the robustness of the Hadamard Test—a fundamental primitive for quantum linear layers—against the specific noise profile of the ibm_kingston processor.

Methodology
The authors employed a hybrid approach combining physical hardware execution with a calibrated "Digital Twin" simulation:

1. Hardware Calibration: Using 100 persistent probes derived from trained MNIST weight matrices, the team ran Hadamard tests on the ibm_kingston backend. These results were compared against an ideal simulator to calibrate a noise model, identifying specific gate noise ($\lambda$) and readout error ($\epsilon$) parameters.
2. Stochastic vs. Physical Comparison: The study contrasted the performance of a stochastic noise model (Digital Twin) against physical hardware execution. This involved analyzing signal compression, rank preservation, and classification accuracy across varying feature depths ($N$), specifically testing $N=16$ and $N=256$.
3. Parametric Analysis: The authors performed a parametric sweep to identify critical thresholds where the "Argmax" operator (used for classification) fails to recover the signal, and conducted a depth ablation study to isolate the impact of circuit depth on coherence.

Key Contributions
The paper introduces four primary concepts and findings:

• The Kingston Constant ($\kappa \approx 0.07$): A characterization of the ibm_kingston processor's systemic signal compression. The hardware exhibits a 93% collapse in signal magnitude, mapping theoretical expectation values (range $[-15, 15]$) into a narrow hardware band (approx. $[-1, 1]$).
• The Hadamard Resilience Law: A proposed law stating that quantum classifiers can maintain high inference accuracy (>93%) even when numerical fidelity collapses due to depolarizing noise. This resilience is attributed to the preservation of the rank order of class probabilities rather than their absolute magnitudes. The Argmax operator acts as a non-linear filter, remaining functional as long as the hardware-compressed margin exceeds the statistical shot-noise floor.
• The Resilience Threshold ($\epsilon_{crit} \approx 0.65$): A parametric limit identified where the Argmax operator fails to recover the signal. Below this gate noise threshold, the system maintains high rank correlation; above it, the signal is lost to statistical variance.
• The Coherence Gap ($\Delta\rho \approx 0.91$): A critical divergence discovered at high feature depths ($N=256$). While stochastic simulations predict high resilience, physical hardware performance collapses. This gap isolates coherent phase errors and crosstalk (rather than depolarizing noise) as the primary barrier to scaling quantum linear layers.

Experimental Results

• Signal Decay and Rank Preservation: Despite the 93% signal magnitude collapse, the monotonic trend of the signal remained intact for lower depths, allowing for correct classification. The Digital Twin (stochastic model) achieved 94.2% accuracy with a Spearman $\rho$ of 0.9856, confirming the theoretical resilience.
• The Coherence Wall at $N=256$: On the physical ibm_kingston processor, the circuit depth for 256 features ($D \approx 10,392$) exceeded the hardware's characteristic coherence limit ($D_{max} \approx 3,465$). This resulted in a "Coherence Wall" where the Spearman $\rho$ dropped to 0.0763 and sign fidelity fell to 53.0% (effectively random guessing), despite the Digital Twin predicting robust performance.
• Shot Noise Efficiency: The study found that rank correlation stabilizes rapidly with shot count. At 512 shots, the system achieved a Spearman $\rho$ of 0.9096, indicating that statistical shot noise is not the primary bottleneck for current NISQ implementations.
• Resource Analysis: The analysis confirmed that the "Coherence Gap" is caused by the accumulation of coherent phase errors in deep CNOT-dense circuits, which scramble the rank ordering required for the Hadamard Resilience Law to hold.

Significance and Claims
The paper claims to establish two critical pillars for NISQ-era quantum machine learning:

1. Validity of the Hadamard Resilience Law: It demonstrates that the Hadamard Test, when coupled with an Argmax operator, is mathematically robust against depolarizing noise. High-performance classification is achievable even when signal magnitudes collapse by over 90%, provided the noise is stochastic and uniform.
2. Identification of the Coherence Gap: The work redefines the scaling limitations of quantum classifiers. It argues that the barrier to scaling is not the stochastic noise floor (which the Resilience Law overcomes) but the accumulation of coherent phase errors in deep circuits. This "Coherence Gap" serves as a new benchmark for NISQ backends, suggesting that future hardware evolution must prioritize the suppression of coherent phase drift and crosstalk over mere gate fidelity improvements.

The authors conclude that while the Hadamard Resilience Law provides a theoretical pathway for functional utility in classification tasks decades before full error correction, current hardware is limited by a "Coherence Wall" at $N=256$. They propose that overcoming this requires architectural shifts, such as distributed linear transformations (Quantum MapReduce), to reduce circuit depth below the coherence limit.