⚛️ quantum physics

When does numerical pulse optimization actually help? Error budgets,robustness tradeoffs, and calibration guidance for transmon single-qubit gates

This paper demonstrates that while numerical GRAPE optimization guarantees zero coherent error and superior amplitude robustness, properly calibrated analytical DRAG pulses are often sufficient for practical transmon gates at standard durations, offering comparable performance with greater robustness to frequency detuning, thereby providing a clear framework for selecting pulse strategies based on specific hardware error budgets and gate time constraints.

Original authors: Rylan Malarchick

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rylan Malarchick

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

Imagine you are trying to teach a very sensitive, wobbly robot arm to pick up a delicate egg and move it to a plate without dropping it or cracking the shell. In the world of quantum computing, that "robot arm" is a qubit (a quantum bit), the "egg" is a quantum state, and the "move" is a logic gate (a calculation step).

The problem is that the robot arm is attached to a spring (the anharmonicity) that makes it want to wobble into a third, unwanted position (leakage) if you move it too fast. Also, the room is drafty (decoherence), which can knock the egg over no matter how steady your hand is.

This paper asks a simple but crucial question: Do we need a super-computer to figure out the perfect way to move the robot arm, or is a simple, well-practiced human technique good enough?

Here is the breakdown of the three "techniques" the authors tested:

1. The Three Techniques

The Gaussian Pulse (The "Brute Force" Approach): This is like telling the robot arm, "Move fast and straight!" It's simple, but if you move too fast, the spring makes the arm wobble into the wrong spot, and you drop the egg.
The DRAG Pulse (The "Smart Human" Approach): This is like a skilled human who knows the physics of the spring. They add a tiny, calculated "counter-movement" (a derivative) to cancel out the wobble. It's an analytical formula—simple math that works very well.
The GRAPE Pulse (The "Super-Computer" Approach): This uses complex numerical optimization (a super-computer algorithm) to find a weird, jagged, perfectly unique movement pattern that theoretically eliminates all wobble, no matter how fast you go.

2. The Big Surprise

The authors ran simulations using real-world hardware numbers (from a machine called IQM Garnet). They expected the Super-Computer (GRAPE) to win every time. It didn't.

Here is what they found, using everyday analogies:

The "Decoherence Floor" (The Drafty Room)

Imagine the room has a draft (noise) that will knock the egg over if you take longer than a certain time. No matter how perfect your movement is, you can't beat the draft. This is the decoherence floor.

The Finding: The "Smart Human" (DRAG) is already so good that the only reason they aren't perfect is because of the draft in the room. The Super-Computer (GRAPE) can only improve the score by about 20% over the Smart Human.
The Lesson: If the room is drafty, spending weeks programming a super-computer to move the arm 20% better is a waste of time. You should just fix the draft (improve the hardware) instead.

The "Speed Limit" (When to use the Super-Computer)

The Finding: If you try to move the arm very fast (under 15 nanoseconds), the "Smart Human" math breaks down. The wobble becomes too complex for simple formulas.
The Lesson: If you need to move at lightning speed, you do need the Super-Computer (GRAPE). But for normal speeds (20 nanoseconds and up), the Smart Human is plenty fast and accurate.

The "Tuning Fork" Problem (Robustness)

This was the biggest surprise.

The Finding: The Super-Computer (GRAPE) found a movement pattern that was incredibly precise only if the robot arm was tuned to the exact right frequency. If the arm drifted even slightly (like a guitar string going out of tune), the Super-Computer's complex pattern failed miserably.
The Lesson: The "Smart Human" (DRAG) used a smooth, simple movement that was much more forgiving. If the robot arm drifted, the Smart Human still did a great job. In the real world, where things drift and change, the simple, robust method is often better than the fragile, perfect one.

3. The Final Verdict: What Should Engineers Do?

The paper gives three clear rules for building quantum computers:

Don't over-engineer for now: If you are building a standard quantum gate today (taking 20ns or more), just use the DRAG method. It's simple, it's robust, and it's already as good as the hardware allows. Using the complex Super-Computer method adds a lot of headache for very little gain.
Fix the hardware first: Since the biggest error source is the "draft" (decoherence/T2 time), engineers should focus on making the hardware more stable. No amount of pulse optimization can fix a broken machine.
Save the Super-Computer for the future: If you need to go super fast (under 15ns) or if you build a machine so perfect that the "draft" disappears, then you should use the complex numerical optimization (GRAPE).

Summary Analogy

Think of it like driving a car:

Gaussian is driving with your eyes closed (fast, but you crash).
DRAG is driving with a GPS and a steady hand. You get to the destination safely and quickly.
GRAPE is a self-driving AI that calculates the perfect trajectory down to the millimeter.

The paper says: "Why use the AI to drive on a bumpy, rainy road when a steady human with a GPS is already getting you there safely? Only use the AI if you need to drive at 200 mph or if the road becomes perfectly smooth."

For today's quantum computers, the road is still a bit bumpy, so the steady human (DRAG) is the best choice.

1. Problem Statement

While Numerical Optimal Control (NOC), specifically the GRAPE algorithm, is theoretically capable of discovering pulse shapes that suppress coherent gate errors to machine precision, its practical necessity in current experimental settings remains unclear.

The Gap: Many existing comparisons between NOC and analytical methods (like DRAG) are flawed because they use uncalibrated baselines, ignore the multi-level nature of transmons (leakage), or omit decoherence.
The Core Question: Does the marginal fidelity gain from complex numerical optimization justify the implementation cost when the system is already limited by decoherence (T1/T2) and calibration uncertainties (frequency drift)?
Context: The study focuses on single-qubit gates on superconducting transmon qubits, where the trade-off between gate speed and leakage to non-computational states ( $|2\rangle$ ) is critical.

2. Methodology

The author conducted a systematic comparison of three pulse strategies under identical conditions using a three-level transmon model parameterized by IQM Garnet hardware specifications ( $T_1 = 37\,\mu\text{s}$ , $T_2 = 9.6\,\mu\text{s}$ , anharmonicity $\alpha/2\pi = -200\,\text{MHz}$ ).

Pulse Strategies Compared:
1. Gaussian: Standard analytical pulse (I-channel only).
2. DRAG (Derivative Removal by Adiabatic Gate): Analytical correction adding a derivative component to the Q-channel to suppress leakage. The parameter $\beta = -1/(2\alpha)$ was used, ensuring amplitude and gate-time independence.
3. GRAPE (Gradient Ascent Pulse Engineering): Numerical optimization of piecewise-constant amplitudes over 50 time slices to maximize subspace fidelity in the full 3-level Hilbert space.
Simulation Framework:
- Dynamics: Open-system evolution modeled via the Lindblad master equation (including amplitude damping $T_1$ and pure dephasing $T_2$ ).
- Metrics: Process fidelity ( $F$ ), leakage probability ( $P_2$ ), and error budget decomposition (coherent vs. decoherent errors).
- Robustness Analysis: Sweeps of qubit frequency detuning ( $\pm 5\,\text{MHz}$ ) and control amplitude errors ( $\pm 5\%$ ) to simulate calibration uncertainties.
Tools: The QubitPulseOpt open-source framework (QuTiP-based) was used for all simulations.

3. Key Contributions

Systematic Error Budget Decomposition: The paper breaks down total infidelity into coherent errors (leakage/control), $T_1$ , $T_2$ , and control noise, establishing a clear hierarchy of error sources.
Robustness Trade-off Discovery: It identifies a counter-intuitive trade-off where the numerically optimized GRAPE pulses are less robust to frequency detuning than the analytical DRAG pulses, despite having higher nominal fidelity.
Calibration Guidance: It provides concrete, regime-specific recommendations for when to use DRAG versus GRAPE, moving beyond "GRAPE is always better" to "GRAPE is better only if..."
Correction of DRAG Scaling: The author clarifies that the DRAG parameter $\beta$ depends solely on anharmonicity, not gate time or Rabi frequency, correcting misconceptions in some literature.

4. Key Results

A. Performance vs. Gate Time

Long Gate Times ( $\ge 20\,\text{ns}$ ): Properly calibrated DRAG operates extremely close to the decoherence floor.
- At $T=20\,\text{ns}$ , DRAG achieves an infidelity of $8.4 \times 10^{-4}$ (including decoherence), which is only 1.2 $\times$ worse than the theoretical limit set by GRAPE ( $7.2 \times 10^{-4}$ ).
- GRAPE eliminates coherent error to machine precision ( $<10^{-15}$ ), but this gain is negligible compared to the dominant $T_2$ dephasing error.
Short Gate Times ( $\lesssim 15\,\text{ns}$ ): DRAG's first-order perturbative correction fails.
- Below $15\,\text{ns}$ , DRAG infidelity rises sharply (dominated by higher-order leakage), while GRAPE maintains near-perfect fidelity.
- Crossover Point: $\approx 15\,\text{ns}$ . Below this, GRAPE is necessary.

B. Robustness to Calibration Errors

Frequency Detuning (Dominant Uncertainty):
- DRAG: Superior robustness. Minimum fidelity over $\pm 5\,\text{MHz}$ detuning is 0.990.
- GRAPE: Poor robustness. Minimum fidelity drops to 0.931.
- Reason: GRAPE pulses have richer spectral content that couples strongly to off-resonant transitions when the qubit frequency shifts. DRAG's smooth, analytical shape is naturally resilient.
Amplitude Errors:
- GRAPE: Superior robustness (Min fidelity 0.994).
- DRAG: Slightly lower (Min fidelity 0.990).

C. Error Budget Hierarchy (at $T=20\,\text{ns}$ )

For both DRAG and GRAPE (once coherent errors are suppressed):

$T_2$ Dephasing: The dominant error source ( $\sim 5.8 \times 10^{-4}$ ).
$T_1$ Relaxation: Secondary source ( $\sim 2.1 \times 10^{-4}$ ).
Control Noise: At 1% noise, comparable to $T_1$ ; at 5%, it dominates the error budget.

Conclusion: Improving hardware $T_2$ yields higher fidelity gains than further pulse optimization for gate times $\ge 20\,\text{ns}$ .

5. Significance and Recommendations

The paper shifts the paradigm from "maximizing theoretical fidelity" to "optimizing practical utility."

Recommendation 1 (Standard Operation): For gate times $\ge 20\,\text{ns}$ on current hardware, DRAG is sufficient. The marginal gain from GRAPE (1.2 $\times$ ) is likely offset by GRAPE's inferior robustness to frequency drift and increased calibration complexity.
Recommendation 2 (High-Speed/Future Hardware): GRAPE is necessary for gate times $\le 15\,\text{ns}$ or for applications targeting error rates significantly below the current decoherence floor (e.g., future fault-tolerant thresholds).
Recommendation 3 (Noisy Environments): In charge-noise-limited systems where frequency drift is common, DRAG is preferable to unconstrained GRAPE. If robustness is required, "robust optimal control" (incorporating detuning uncertainty into the cost function) is the suggested path forward.
Hardware Focus: The highest-leverage upgrade for current transmon devices is improving $T_2$ coherence, not pulse complexity.

Final Takeaway: Numerical optimization is a powerful tool, but for near-term transmon single-qubit gates, a properly calibrated analytical DRAG pulse often provides the best balance of fidelity, robustness, and implementation simplicity. The "decoherence floor" is the true bottleneck, not the pulse shape, for standard gate durations.