Benchmarking Quantum Operations: RB vs. QPT

Hello. In our last lesson, we took a hands-on approach to randomized benchmarking (RB), implementing it on a noisy simulator. We saw how RB uses sequences of random Clifford gates to extract a single, powerful metric: the average gate fidelity. This metric, encapsulated in the Error Per Clifford (EPC), tells us how well our gates perform on average and has the crucial property of being robust to state preparation and measurement (SPAM) errors.

However, an average can sometimes conceal more than it reveals. If a gate is failing, is it due to a systematic over-rotation (a coherent error), or is it decohering (an incoherent error)? RB, by design, cannot distinguish between these cases. This lesson addresses that gap. Our goal is to compare and contrast randomized benchmarking with quantum process tomography in terms of the information they provide and their experimental scalability. We will treat these as two complementary tools in the quantum characterization toolkit, each with its own distinct purpose, strengths, and weaknesses.

What is Quantum Process Tomography?

While randomized benchmarking provides a single figure of merit, Quantum Process Tomography (QPT) aims for a much more ambitious goal: to obtain a complete mathematical description of a quantum process, such as a gate or a noise channel.

If RB gives you the average grade of a student across all subjects, QPT provides the full, detailed report card, including the teacher's comments on every subject, highlighting specific strengths and weaknesses. The output of QPT is a mathematical object—typically a process matrix $\chi$ or a set of Kraus operators $\{E_k\}$—that fully defines the quantum map $\mathcal{E}$ describing the noisy gate.

The general procedure for QPT is as follows:

1. Prepare a basis of input states: For a single qubit, this means preparing states like $|0⟩,|1⟩,|+⟩,|i⟩$.
2. Apply the process: The gate you want to characterize is applied to each of these input states.
3. Characterize the output states: For each output, you must perform full quantum state tomography. This involves repeatedly preparing the state and measuring it in a complete set of bases (e.g., X, Y, and Z bases) to reconstruct its density matrix.
4. Reconstruct the process: From the set of input-output density matrix pairs, you can mathematically solve for the process matrix $\chi$ that describes the transformation.

With this complete description, we can now dig into the core comparison with RB.

Axis 1: Information Provided

This is where the difference between the two techniques is most stark.

Quantum Process Tomography (QPT)

As the "tomography" in its name implies, QPT provides a complete, high-resolution picture of the quantum process. From the reconstructed process matrix $\chi$, you can calculate anything you want to know about the gate's performance:

• Average Gate Fidelity: This can be directly computed from $\chi$.
• Coherent Errors: You can identify systematic errors, such as whether a gate is consistently over- or under-rotating the qubit state.
• Incoherent Errors: You can characterize the decoherence properties, such as the rates of dephasing ($T_2$) or amplitude damping ($T_1$).
• Error Correlations: It can reveal if certain types of errors are correlated.
• Worst-Case Performance: You can compute metrics like the diamond norm distance, which quantifies the worst-case error over all possible input states.

This detailed breakdown provides what is often called "actionable advice." An experimentalist can use the QPT results to diagnose the physical cause of an error and attempt to recalibrate their control pulses to fix it.

Randomized Benchmarking (RB)

In contrast, RB provides a single piece of information: the average gate fidelity. It is a benchmark that tells you how well your gates perform on average, but it offers no insight into why they might be failing. The averaging process (twirling) that makes RB so elegant intentionally washes out the details of the noise, converting it all into an effective depolarizing error rate. A gate with a coherent rotation error and a gate with pure dephasing could, in principle, produce the same RB decay curve.

Axis 2: Scalability and Experimental Cost

If QPT provides so much more information, why isn't it used all the time? The answer lies in its staggering resource cost.

Quantum Process Tomography (QPT)

QPT scales disastrously with the number of qubits, $n$. To characterize a process on $n$ qubits (a Hilbert space of dimension $d=2^n$), you need to determine the $\chi$ matrix, which has $d^4-d^2=(2^n{)}^4-(2^n{)}^2$ independent real parameters.

The number of experiments required scales exponentially. A full QPT on $n$ qubits requires:

• Preparing $4^n$ different basis input states.
• For each output, performing full state tomography, which requires $3^n$ different measurement settings.
• The total number of unique experimental settings scales as $O(12^n)$, and the classical post-processing to reconstruct the matrix is also computationally intensive.

This exponential scaling makes full QPT infeasible for anything beyond two or three qubits.

Randomized Benchmarking (RB)

This is RB's greatest strength. The number of required experimental settings is independent of the number of qubits. For an $n$-qubit RB experiment, you still prepare a single initial state (e.g., $|00...0⟩$) and measure in a single basis (the computational basis). The complexity is determined by:

1. The maximum sequence length, $m$.
2. The number of random sequences you average over, $K$.

While generating the classical description of an $n$-qubit Clifford gate and its inverse becomes hard for large $n$, the experimental cost of RB does not have the prohibitive exponential dependence on $n$ that plagues QPT. This is why RB, and not QPT, is the industry standard for assessing the performance of multi-qubit processors.

Axis 3: Robustness to SPAM Errors

There's one more critical difference that often makes RB the more practical choice even for a single qubit.

Quantum Process Tomography (QPT)

Standard QPT is extremely sensitive to State Preparation and Measurement (SPAM) errors. If your initial state isn't prepared perfectly, or your measurement apparatus is miscalibrated, QPT has no way of distinguishing these errors from errors in the gate itself. It will faithfully report the combined error of (preparation + gate + measurement), attributing it all to the gate.

Randomized Benchmarking (RB)

As we saw in the previous lessons, RB is inherently robust to SPAM errors. In the decay model,
$F_{\text{seq}}(m)=A_0{p}^m+B_0$
the SPAM errors are absorbed into the constant offsets $A_0$ and $B_0$. The decay parameter $p$, from which the gate fidelity is calculated, is independent of them. This is a massive practical advantage, as it decouples the gate characterization from the (also imperfect) preparation and measurement steps.

It's worth noting that a more advanced and even more costly technique called Gate Set Tomography (GST) was developed to solve this problem for QPT. GST characterizes the gates, the initial state, and the measurement operators all simultaneously in a self-consistent way, but its resource cost scales even more unfavorably than standard QPT.

Summary: A Tale of Two Tools

We can summarize the comparison in a table:

These techniques are not competitors but rather complementary parts of a complete characterization suite. You might use RB to quickly survey the performance across a 50-qubit chip. If a particular qubit's gates are performing poorly, you might then zoom in on that specific qubit with QPT to diagnose the physical root of the problem.

Conclusion

Today we've drawn a clear distinction between two fundamental methods for quantum device characterization.

Key Takeaways:

• QPT provides a complete picture: It reconstructs the full process matrix, offering detailed insights into all error mechanisms (coherent, incoherent), but at a high cost.
• RB provides a scalable average: It yields a single, robust metric (average gate fidelity) that scales to many qubits and is insensitive to SPAM errors, but it hides the details of the noise.
• The fundamental trade-off: The choice between QPT and RB is a classic trade-off between information and scalability. QPT is a deep, narrow diagnostic tool, while RB is a broad, shallow benchmarking tool.
• Complementary roles: In practice, both are essential. RB identifies problem areas at scale, and QPT can be used to perform a deep dive on those specific problems.

Preview of the next lesson:
We have discussed what QPT aims to achieve and how it compares to RB. In the next lesson, we will delve into the mechanics of how it works. We will derive the measurement scheme for quantum process tomography and the linear inversion procedure for reconstructing the process matrix ($\chi$-matrix), putting the mathematical machinery behind today's high-level overview.