Speeding Up Quantum Hardware Calibration with ‘Restless Measurements’ Techniques
By Robert Davis and Daniel J. Egger
Quantum logic gates are one of the fundamental building blocks of quantum circuits; they give us the ability to manipulate qubits and execute quantum computations. In superconducting quantum computing, we implement quantum gates by zapping our qubits with carefully designed pulses of microwave energy. But if we want accurate, high-quality gates, we need to carefully tailor and calibrate these pulses to match individual quantum systems.
In a new paper published in Physical Review Applied, researchers detail a set of novel techniques for improving the calibration and characterization of quantum gates. These two processes play an important role in maintaining the accuracy and fidelity of quantum computations. However, quantum computers must perform these gate maintenance tasks frequently, usually in between normal quantum jobs. This eats up valuable processor time that could instead go to real quantum applications. The IBM Quantum researchers suggest their new approach could reduce the time quantum computers must spend on quantum gate calibration and characterization tasks.
Calibration and characterization for high-quality quantum gates
Quantum gate calibration and characterization tasks are very important in quantum computing because quantum hardware “drifts” over time, undergoing small physical changes that can throw off the accuracy of quantum gates. To compensate, we run calibration and characterization experiments that tell us how we need to adjust our gate pulse parameters (calibration tasks), and that allow us to measure the effectiveness of those calibrations (characterization tasks).
We determine the specific calibrations we need to make by applying pulse sequences that are designed to amplify various gate errors, making it easier to see where problems occur. Quantum hardware engineers use the outcomes of these experiments to determine which pulse parameters are in need of calibration, and by how much.
Once that’s complete, we can determine how well we did by measuring the quality and fidelity of the newly calibrated quantum gate. To perform those measurements, we use quantum gate characterization techniques like randomized benchmarking and quantum process tomography. These characterization techniques are similar to gate calibration methods in that they involve running time-consuming experiments on quantum hardware.
Bottlenecks in characterization and calibration algorithms
Traditional characterization and calibration algorithms are just like most other quantum algorithms. They start by initializing some number of qubits, then they run a series of gates on those qubits in a quantum circuit, measure the results, initialize the qubits again, and continue iteratively from there.
However, as the team reports, the “initialization” part of that procedure may not always be necessary.
There are many ways to initialize a qubit. If you’ve ever spent hours standing in line for a popular theme park ride or during a busy day at your local supermarket, then you’re already familiar with one method, a sophisticated technique known as “waiting.” Qubits can only maintain quantum states for brief periods of time — just fractions of a second — after which they decay to their ground (or “0”) state. This is the qubit’s “T1” time, or the “lifetime of the qubit.” To reinitialize qubits after running a circuit, one can simply wait somewhere between 5x and 10x the T1 time. Once that time elapses, the qubits will have decayed to their ground state, and are ready to execute the next circuit.
This “passive” reset works well for quantum systems that have short T1 times. However, today’s quantum computers have significantly longer T1 times than their predecessors, making passive reset highly inefficient.
Researchers eventually developed an alternative technique called “active” reset, where rather than waiting for qubits to return to their ground state, we instead perform some active operation to force the qubit back into it. There are a few different ways to perform that reset operation, but all require a non-negligible amount of time to implement. Active reset techniques aren’t always perfect, either. In many cases, engineers have to build in an additional delay before running the next circuit to ensure the system has really reached the ground state.
That delay is shorter than the waiting period in passive reset — on IBM Quantum systems, an active reset might require just one T1’s worth of time compared to 5–10x for passive reset. The research team was able to speed things up even further.
No rest for the qubit
In quantum physics, we often say that measuring a quantum state destroys its information, but that’s a simplification. When we measure a superconducting qubit with a well-tuned measurement pulse, for example, we aren’t destroying all its information. Instead, the measurement “projects” the qubit into either a 0 or 1 state. If we run a quantum circuit, measure the results, and find that all our qubits are in the 0 state, we could theoretically skip the reset process and run our next quantum circuit immediately, since all the qubits would already be in the ground state.
We only perform resets because most quantum circuits output at least some qubits that are in the 1 state, and we generally want to flip those to the 0 state before we start our next circuit. However, it turns out that there is a large class of circuits for which that kind of reset isn’t necessary.
The IBM Quantum team found that if they skipped the qubit reset process and begin the next circuit with some qubits in the 0 state and some in the 1 state, they could use classical post-processing to relabel the 1-state qubits and easily get the results those qubits would have outputted if they had started in the ground state. The researchers call this “restless measurement” because the qubits can execute circuit after circuit with no resets between them.
Let’s take a closer look at how this works:
Segment “a” of the image above depicts a single-qubit job with two circuits that repeats four times. The first circuit consists of only an X gate, while the second circuit consists only of a Hadamard gate. The gray boxes that appear to the right of each gate represent the measurement we take after the circuit is complete.
The qubit starts off in the ground 0 state for our very first “shot,” or iteration, of the circuits. We can then see in the system memory that the first circuit outputs a 1, and the second outputs a 0. After that, the qubit is left in the ground state and we run the circuit again without resetting the qubit.
On this next run, we can see that the second circuit outputs a 1 state. Once again, because we are using the restless measurement scheme, we simply begin our third shot of the circuit without resetting the qubit, and record the results. Once all four shots are complete, we begin our classical post-processing, which we can see detailed in segment “b.”
Our initial output is a chronological list of the outputs our circuits produced, which isn’t very useful since some of our circuit executions didn’t start off in the ground state. To figure out what the outputs would look like if each shot had started off in the ground state, we review each item in the time-ordered list, compare each shot with the previous value, and perform an “exclusive OR” operation on them.
This is just a concise way of saying: “If the state of the qubit did not change, count it as a 0. If the state of the qubit changed, count it as a 1.” Classical computers can perform this kind of logical operation extremely quickly, and of course, processing time on a classical CPU is much less computationally expensive than time on a quantum processor.
Quantifying the ‘restless’ speedup
Data acquisition using restless measurements may be dozens of times faster than measurements taken with active or passive reset (including any associated delays), but the actual speed-up they deliver to common calibration and characterization tasks is more modest.
That’s because circuit executions are only a small part of the jobs we run for quantum gate calibration and characterization. We also spend considerably more time using classical computers to prepare the circuits and transfer their data to the control electronics that will generate the pulses for the quantum processor.
The team notes that future improvements in circuit preparation and data transfer may yield additional speedups. This also ties into the goal of increasing the number of quantum circuits a quantum processing unit can execute in a given unit of time, a metric that’s referred to as Circuit Layer Operations per Second (CLOPS).
Still, the actual speed-ups that the researchers found on the quantum processing unit were significant. The team found restless measurements brought a 5.3x speed-up to the QPU time of randomized benchmarking quantum gate characterization on two qubits, and a much larger 38.7x speed-up for restless measurements in the QPU time for single-qubit quantum process tomography characterization. They also found a speed-up of up to 38.3x for calibration tasks. Speed-ups varied considerably depending on the specifics of the experiments the researchers were running, and the active reset schemes to which those experiments were compared.
The team says that quantum hardware engineers can begin using at least some of the techniques demonstrated in their paper today. Engineers can easily start running randomized benchmarking and single-qubit calibrations, for example, using the approach outlined in the new paper. At the same time, they also note that techniques like restless measurements will never be generically useful for a large quantum circuit, though they can be useful in specific cases.
Over the long-term, the IBM Quantum team hopes that continued research in this area will yield a new approach to quantum gate calibration and characterization that significantly increases the availability of quantum hardware for real world quantum applications. For more detail on their findings and proposed next steps, be sure to check out the full paper, here. To experiment with restless measurements for yourself, take a look at the restless measurements experiment tutorials on Qiskit.org.
