Q-gen: A Parameterized Quantum Circuit Generator

The quantum circuit computing model is analogous to classical digital computing, where computations are performed using a sequence of logic gates. In the quantum circuit model, quantum computations are carried out using quantum gates, which are a series of unitary matrix operators that introduce a state or phase transition on the qubit(s) it acts on. Qubits are the fundamental units of quantum information, capable of existing in a superposition of states and entangling with others. These properties enable quantum computers to solve certain problems exponentially faster than classical computers.

In a typical circuit model quantum computing workflow, the quantum algorithm is described using a high-level quantum circuit, and then circuit transpilation compiles the circuit into low-level hardware-specific instructions [7] for execution. In this process, the circuits are optimized to reduce extra SWAP gates [8, 9], and various policies [10, 11, 12] are applied to search for the best placement of the qubits on the quantum processor. Additionally, software error mitigation [13, 14] and hardware error correction [15, 16] techniques are developed to reduce the negative effect of decoherence and communication noise.

Several open-source frameworks exist to help automate or assist in creating quantum circuits. Qiskit [5], Cirq [17], and Braket [18] are three major quantum computing frameworks that provide pre-built components and templates for generating higher-level quantum algorithms. All the frameworks also support the OpenQASM [6] representation for lower-level representation of quantum circuits.

Currently, we believe that no projects are specifically targeted for high-level quantum circuit generation, but many works focused on quantum benchmarking usually include some low-level circuit dataset or a lightweight circuit generator with limited functionality. The SupermarQ [19] benchmark suite includes 8 application circuits and a circuit generator parameterized by the number of qubits. The QASMBench [20] benchmark suite provides more than 50 openQASM implementations of various scales of quantum circuits in different qubit settings. The Application-Oriented Benchmarks [21] offer 11 algorithms categorized into 3 groups with a specialized generator for creating random benchmark circuits.

Artificial Intelligence (AI) has experienced explosive development since the mid-2010s. In addition to the classical optimization methods mentioned in section II-A, there is a growing interest in using AI to help accelerate the development of quantum computing. This also gives a strong motivation for our work because machine learning models require a large number of quantum circuits with high variability for efficient training.

For circuit optimization, there are works on using Large Language Models (LLM) to design better quantum circuits [22] and using Transformers to simplify CNOT circuits [23]. Deep reinforcement learning techniques are also explored to minimize the number of T gates [24]. Many works proposed different machine learning models for predicting the output fidelity of quantum circuits [25, 26, 27, 28, 29], which can be used for finding better layout and improve CNOT routing. Works that have the potential for AI integration can also benefit from a large supply of quantum circuits. One example is circuit knitting, a technique of cutting the full circuits into smaller pieces for more efficient execution [30, 31, 32].

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95]images/alg_group_0.png Q-gen quantum circuit dataset statistics and example circuit of the Deutsch-Jozsa algorithm, the Bernstein-Vazirani algorithm, and Simon’s algorithm.

The Deutsch-Jozsa algorithm [33] determines whether a special boolean function $f$ is constant or balanced: a constant $f$ will always output all 0 or 1 regardless of the input, and a balanced $f$ will always output 0 for half of the input and 1 for the other half. Classically, at least two queries to $f$ are required to find the answer, but the quantum version of this algorithm only requires one query.

Q-gen can generate the oracle with both types of boolean functions, and the oracle width is decided by the problem size parameter. The constant oracle is a very simple oracle that only applies X gate to the output qubit, the balanced oracle adds more circuit complexity by creating entanglement with CNOT gates. Intuitively, as the problem size grows, the depth of the generated circuit remains constant for the constant oracle, and the number of single-qubit gates grows linearly, shown in figure II-C. If the oracle is balanced, the depth and number of CNOT gates both grow linearly.

This algorithm is useful for generating simple circuits (no CNOT gates) with constant depth but growing width. The only other algorithm in Q-gen that has a similar circuit complexity pattern is the quantum key distribution algorithm, but its number of single-qubit gates and measurements is doubled.

The Bernstein-Vazirani algorithm [34] solves problems similar to but harder than the Deutsch-Jozsa problem. The hidden function in this algorithm returns 0 or 1 based on the bitwise product of the input string with a hidden “secret” string $s$, the goal is to find $s$ with as few queries to the oracle as possible. The best classical solution must query every input bit of $f$. In contrast, the quantum version of this algorithm only requires one query.

Q-gen can generate the Bernstein-Vazirani oracle randomly or based on a binary string, with its width controlled by problem size. The oracle places the control of the CNOT gate on the input qubit if $s=1$ and leaves the qubit untouched if $s=0$, the target of all CNOT gates are placed on the additional phase-kickback qubit. The growth pattern of the Bernstein-Vazirani algorithm is similar to the Deutsch-Jozsa algorithm, but as problem size increases, the depth and number of CNOT gates all grow linearly at a higher rate.

This algorithm is good for generating elementary quantum circuits with some entanglement. Also, the output of the Bernstein-Vazirani circuit is easily verifiable without any post-processing, which makes it a good candidate for quantum benchmarking.

Simon’s algorithm [35] has proved that quantum computers can offer exponential speed-up over their classical counterparts. The oracle function $f$ in Simon’s problem can be one-to-one or two-to-one according to a secret string $s$. Similar to the Deutsch-Jozsa algorithm and the Bernstein-Vazirani algorithm, we want to determine the type of $f$ as fast as possible. While the best classical solution requires $O(\sqrt{2^{n}})$ queries, the quantum algorithm only requires $O(n)$ queries [36], achieving exponential speed-up.

The circuit structure of Simon’s algorithm is special in terms of circuit width. Although problem size is still defined as the oracle width, the generated circuit will have a qubit count that doubles the number of problem size. Upon measurements, the result will be multiple guesses related to the secret string $s$, and $s$ can be revealed with some moderate post-processing. Simon’s algorithm requires more CNOT gates than the other two query algorithms in Q-gen, and the growth rate for the number of CNOT gates is approximately 2$\times$ of circuit depth, as shown in figure II-C.

Simon’s algorithm has the highest circuit complexity compared with the Deutsch-Jozsa algorithm and the Bernstrin-Vazirani algorithm, considering their simulation time and the number of CNOT gates at the same problem size. Since it requires multiple shots and post-processing to find the correct answer, Simon’s algorithm is good for testing the full quantum$+$classical hybrid computing workflow.

Quantum Fourier transform (QFT) [37] is the quantum implementation of the classical Discrete Fourier Transform (DFT). To perform DFT on $2^{n}$ elements, classical algorithms like the Fast Fourier Transform (FFT) require $O(n2^{n})$ operations. QFT only needs $O(n^{2})$ operations, which is exponentially fewer than the classical algorithms [3]. Note that QFT does not directly solve any specific problems, if the qubits are measured in the computational basis, the result will appear random as they have been transferred to the Fourier basis.

Q-gen takes problem size as the circuit width and generates the barebone QFT subroutine. Although this circuit can be directly integrated into other higher-level algorithms, it is not directly verifiable. Q-gen provides three generation parameters that make verification possible: the qubits can be initialized in the Fourier basis using an integer input, and the QFT subroutine can be inverted to transform the qubits into the computational basis. By measuring the qubits, the result will be the integer number used for initialization. While QFT’s circuit complexity growth pattern is exponential, as shown in figure III-D, it is still a very efficient implementation since the depth only reaches around 4,000 when problem size is 47 qubits.

QFT is the building block for many practical quantum algorithms, so reducing its circuit complexity and at the same time increasing its accuracy is an important research topic. For instance, the approximate QFT algorithm can ignore phase rotations below a certain threshold and still yield the correct result with acceptable error. Q-gen’s parameterized generation offers a quantitative approach to quickly analyze the behavior of various QFT circuits.

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95]images/alg_group_1.png Q-gen quantum circuit dataset statistics and example circuit of quantum Fourier transform (QFT), quantum phase estimation (QPE), and Shor’s algorithm.

Quantum phase estimation (QPE) [38] is one of the key applications of QFT. This algorithm uses phase-kickback and inverse-QFT to estimate the phase/eigenvalue of any unitary operator $U=e^{2\pi i\theta}$. The phase angle $\theta$ is recorded in the counting qubits by the iterative controlled-U gates applied on the eigenstate qubit. With $n$ counting qubits and the measurement result equal to $x$, $\theta$ can be estimated as $\theta\approx x/2^{n}$.

The QPE circuits from Q-gen have one eigenstate qubit for $U$ to act on, and $n$ counting qubits, controlled by problem size. The number of counting qubits directly affects the estimation accuracy of the circuit, when $\theta$ has a denominator equal to $2^{n}$, it is possible to retrieve the exact angle of $\theta$. Due to the high implementation cost of the iterative controlled-U gates, the complexity of QPE circuits grows exponentially, reaching a depth more than $5.0\times 10^{5}$ at problem size = 17. This causes other QPE-based algorithms to have an even higher gate count, like the quantum walk search algorithm.

QPE has many practical applications because some classically hard problems can be condensed to phase estimation, the most famous examples include period-finding and prime factorization. However, it is still hard to implement QPE on current NISQ hardware due to its high circuit complexity. Although we can reduce the number of counting qubits, it also causes the estimation accuracy to drop. Q-gen can generate a large sample of QFT circuits to form a baseline of estimation accuracy, which can then be used to study the trade-off between reduced gate count and loss of accuracy.

Shor’s algorithm [39] can factor any large number $N$ in polynomial time, better than every known classical algorithm. Under the hood, the core problem that contributes to this quantum speedup is the period-finding problem. If the period $r$ of the modular exponential function $f(x)=a^{x}modN$ can be found in polynomial time, the factor(s) of $N$ can also be found in polynomial time. Shor’s algorithm uses QPE to accelerate the period-finding process, the subsequent search for factor(s) can be done efficiently using classical algorithms.

Shor’s algorithm is a quantum-classical hybrid algorithm that utilizes classical pre-processing to return a result before the QPE subroutine if $N$ is simple. To ensure the algorithm reaches the quantum part, $N$ has to be reasonably hard to factor. Specifically for Shor’s algorithm, $N$ has to be odd and not formed by $m^{n}$ for $m\geq 1$ and $n\geq 2$. Additionally, a random guess of $a$ is required to kick-start the modular exponential function, and $a$ must be a coprime of $N$ [3]. Q-gen automatically generates the suitable $a$ and provides a list of $N$ (as problem size, up to 123) ready to use for the quantum circuit. The main difficulty in Shor’s algorithm is the quantum implementation of the modular exponential function, many customized implementations exist focusing on minimizing the gate count or using a specialized basis gate set [40, 41]. Q-gen implements a relatively efficient modular exponential circuit using $2n+3$ qubits with $n$ equal to the bit-length of $N$ [42, 43]. Most importantly, this implementation is general regardless of $N$, ideal for Q-gen’s parameterized generation process.

Like most quantum algorithms, Shor’s algorithm only displays quantum advantage over classical algorithms when the problem size is sufficiently large [44]. Although the ability to factor large prime numbers in polynomial time potentially breaks the RSA cryptosystem, it is still unclear whether this can be realized on a practical quantum computer in the foreseeable future. As of 2023, the largest factored RSA number has 829 bits (RSA-250) [45]. A quick analysis using Q-gen shows that factoring the same number using a quantum computer requires at least 1,661 qubits, excluding the auxiliary qubits.

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95]images/alg_group_2.png Q-gen quantum circuit dataset statistics and example circuit of Grover’s algorithm, quantum counting, and quantum walk algorithm.

Grover’s algorithm [46] uses the Grover oracle $G$ and Diffusion operator $D$ to accelerate search problems. $G$ contains logical computations that mark the target states with a negative phase, and $D$ then amplifies the measurement probability of the marked states while simultaneously reducing the probability of other states. For an unstructured database with $N$ items, Grover’s algorithm can find all $M$ targets after approximately $\sqrt{N/M}$ Grover iterations ($G+D$), providing a quadratic speedup over the classical algorithms.

Q-gen can automatically generate $G$ based on problem size with the optimal $M$ that only requires 1 Grover iteration. The target state in $G$ can also be specified as an input string. Q-gen will append the appropriate number of Grover iterations to the circuit, or take this number as another input. The width of Grover’s circuit always equals the number of measurement gates and they both grow linearly, but its circuit complexity grows exponentially as the search space increases. Noticeably, compared with the quantum counting algorithm, the number of CNOT gates grows faster than the number of single-qubit gates, as shown in figure III-F.

Q-gen offers a quick way to statistically and empirically analyze the performance of Grover’s algorithm, especially when the search space is large. The parameterized circuit generation also enables an in-depth investigation of the algorithm. For example, testing how different numbers of Grover iterations affect result accuracy.

The quantum counting algorithm [47] can estimate the number of solutions $M$ inside $N$ items, it is a combination of Grover’s algorithm and quantum Fourier transformation (QFT). This algorithm can be seen as a prerequisite for Grover’s algorithm since Grover’s algorithm requires $M$ to calculate the correct number of iterations. It can also determine whether a solution even exists inside $N$ items, which can accelerate certain NP-complete problems like the Hamiltonian Cycle problem [3].

There are two groups of qubits inside the quantum counting algorithm: the searching qubits iterate over the search space $N$, and they use controlled Grover’s operator to mark $M$ on the counting qubits. The default setting of Q-gen takes problem size as the number of counting qubits and assigns an equal number of searching qubits to ensure optimal search results. The number of solutions $M$ can be randomly generated or specified. Under the default setting, as problem size increases, the width of the generated circuits grows faster than the number of measurement gates. Compared with Grover’s algorithm, The number of single-qubit gates grows faster than the number of CNOT gates.

Q-gen’s generation settings can be easily changed to investigate the efficiency and accuracy of quantum counting under different problem spaces. More importantly, this algorithm is another great candidate for benchmarking the quantum$+$classical hybrid computing workflow, since it contains two famous quantum subroutines: Grover’s operator and inverse-QFT, plus moderate classical post-processing.

Quantum walk [48] is the quantum version of the random walk search algorithm. On each walking step, the “walker” enters superposition to search all the nodes on a graph simultaneously. For a graph with $N$ nodes and has $M$ targets, approximately $1/\sqrt{|M|/N}$ iterations are required to find all targets [49]. This is another quantum search algorithm that provides a quadratic speedup.

Q-gen implements the coined quantum walk, which uses a coin to direct how the “walker” moves. Specifically, Q-gen uses the Grover coin, equivalent to the Diffusion operator $D$ from Grover’s algorithm. Each iteration contains a phase oracle that marks the target states, followed by quantum phase estimation using the coined walk steps. Q-gen takes problem size as the width of the coin, then automatically determines the width of the node qubits, and the number of counting qubits for phase estimation. Due to the extremely high gate cost of the coined walk steps, under the default setting, a 3-qubit coin will generate a circuit with more than $1.0\times 10^{6}$ CNOT gates and more than $1.5\times 10^{6}$ single-qubit gates.

Although this is a resource-intensive algorithm, many parameters are still available in Q-gen to simplify its circuit structure. The number of counting qubits for phase estimation can be reduced, or the width of the node qubits can be set smaller. The number of iterations can also be specified, potentially lowering the search accuracy.

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95]images/alg_group_3.png Q-gen quantum circuit dataset statistics and example circuit of quantum key distribution, superdense coding, and quantum teleportation.

Quantum key distribution [50] is a secure quantum communication protocol based on one primary quantum principle: measurement collapses superposition. Suppose a sender prepares a qubit in a specific basis and sends it through a quantum communication channel. In that case, the receiver can retrieve the same information by measuring the qubit in the same basis. However, if the qubit is measured in a different basis before reaching the receiver, the qubit collapses prematurely and the receiver will get a random result.

Q-gen generates the full quantum circuit simulating the key distribution process. The problem size controls the qubit width of the circuit, equivalent to the bit width of the key. The interception parameter controls whether an “attacker” will be inserted before the final measurement. This parameter causes the circuit to have mid-circuit measurements, which is the only occurrence among Q-gen’s algorithms. Q-gen can also simulate the circuit and output the information measured at each stage, including if the “attacker” has been detected. The generated circuits only contain single-qubit gates with a linear growth pattern. The depth has two fixed variations depending on whether interception is true.

This key distribution protocol is not designed to be risk-free. For example, if the “attacker” inadvertently picked the same measurement basis as the sender, the interception will go undetected. However, adding more qubits to the key can significantly lower this risk. Q-gen’s generation parameters provide a quick method to analyze the quantum circuit’s complexity under different security conditions.

Superdense coding [51] is a quantum communication protocol that encodes 2 bits of classical information using 1 qubit. This protocol requires the sender and receiver to share a pair of entangled qubits before the transmission starts, usually prepared by a third party. The sender can pick from 4 quantum gates to apply on the qubit, corresponding to 4 possible binary numbers using 2 classical bits. After the sender receives this qubit, the message can be decoded by disentangling the qubit pair. Superdense coding is also secure, as the entanglement will be destroyed if any qubit is prematurely measured before reaching the receiver.

Q-gen generates the three steps of superdense coding: entangling, encoding, and disentangling. The problem size controls the width of the circuit, equal to the number of qubits used for transmission. The number of CNOT gates grows linearly alongside the depth of the circuit. The number of qubits to encode, and the single-qubit gates used for each encoding, can all be specified or randomly generated.

The superdense coding circuits are highly structured and easy to scale up. This makes it a great candidate for benchmarking the entanglement ability, and the measurement fidelity of the quantum hardware.

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95]images/alg_group_4.png Q-gen quantum circuit dataset statistics and example circuit of the quantum approximate optimization algorithm (QAOA), the variational quantum eigensolver (VQE), and the variational quantum classifier (VQC).

Quantum teleportation [52] recreates the sender’s qubit on the receiver’s side using a pair of entangled qubits and classical bits. The sender’s qubit is processed with one of the qubits from the entangled pair, and then measured to obtain a 2-bit classical data. This data is transferred to the receiver through a classical channel so the receiver can reconstruct the sender’s qubit on top of the other qubit from the entangled pair. This protocol can be seen as the opposite of superdense coding.

The minimal teleportation circuit in Q-gen involves 3 qubits. The state to be teleported is initialized on $q_{0}$, the entangled pair is created on $q_{1}$ and $q_{2}$. After the deferred measurement on $q_{0}$ and $q_{1}$, the original state is recreated on $q_{2}$. An inverse-initializer is then placed on $q_{2}$, if the state is successfully teleported, it should inverse the qubit back to $\ket{0}$. The problem size defines how many sets of teleportation circuits to create, thus the generated circuit is analogous to a classical parallel communication bus. The number of CNOT and single-qubit gates follow a linear growth pattern, and the circuit’s depth is always fixed, as shown in figure III-I

The quantum teleportation algorithm in Q-gen is easily verifiable because the ideal measurement should always be all $\ket{0}$. Additionally, because every set of teleportation circuits is physically isolated, this algorithm can potentially detect crosstalk errors between different regions on a quantum processor.

The quantum approximate optimization algorithm (QAOA) [53] can approximate the solution of the combinatorial optimization problem. For complex optimization problems like MaxCut and Max-kXOR, QAOA gives higher quality solutions and takes less time compared to the classical algorithms [54]. The cost function from the combinatorial optimization problem is encoded into a quantum circuit called variational form, then the expectation value can be measured repeatedly to optimize the parameters of the cost function, ultimately converging to the solution.

The QAOA circuit in Q-gen is based on the popular MaxCut problem, which tries to partition a graph so that the number of edges between the two sets of nodes is maximum. Q-gen generates the variational form based on the input graph and randomly initializes all the parameters. The number of nodes in the graph is defined by problem size, and each edge in the graph is translated to an Rzz gate which decomposes into two CNOT gates and one Rz Gate. The depth and gate counts of the QAOA circuit grow linearly with problem size. Even if the input graph is fixed and the repetitions of the variational form increase, the circuit’s depth still grows linearly, making QAOA a very efficient quantum algorithm [53].

QAOA is a heuristic algorithm that gives better approximations after every measurement and optimization cycle. Although it has the potential to outperform certain classical algorithms, the implementation of QAOA still heavily depends on the topology of the input graph. Q-gen can help researchers quickly explore the circuit structure of QAOA under different input graphs.

[t](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.9]images/alg_system.png The complete Q-gen algorithm system illustration. The algorithm’s complexity rating is designed to be compared vertically within its category instead of horizontally across different algorithm categories. To help form the connection between different algorithms, we include some generalized quantum computing problems, indicated by the dashed circles. The connections are gathered from various textbooks, lecture notes, and scientific papers. [3] [55] [56] [57] [58]

The variational quantum eigensolver (VQE) [59] can estimate the minimum eigenvalue of a physical system represented by a matrix. Since the current quantum hardware is not capable of running deep QPE algorithms, VQE is designed as an alternative to efficiently approximate the solution with the help of a classical optimizer [59]. VQE has many practical applications in the chemistry and physics fields. For example, finding the ground state energy of molecules [60].

The implementation of the variational forms for VQE depends on the specific system being simulated. Q-gen fully utilizes Qiskit’s circuit library to give a wide variety of generation parameters. The choices of the single-qubits gates for the initial mixing unitary are generated according to the gates parameter, the entanglement pattern (full, linear, etc.) is generated based on the entanglement parameter, and the repeat parameter controls the number of repetitions of the full variational form. Like the other variational algorithms in Q-gen, the rotation angle of the gates can all be specified or randomly generated.

Because the VQE algorithm in Q-gen is highly customizable, it enables explicit control of the percentage of single-qubit and two-qubit gates, which produces circuits with vastly different structures.

The variational quantum classifier (VQC) is the quantum version of the traditional neural network classifier [61]. By embedding classical data into a quantum feature map circuit, a variational form can be trained on this feature map just like training the classical neural networks. Machine learning is another promising application of variational quantum algorithms. As the datasets and models get larger, quantum computing can potentially make training more efficient by utilizing the exponentially growing parameter space represented by more entangled qubits.

Q-gen implements a common VQC circuit using the ZZFeatureMap and the RealAmplitudes variational form [62]. The repetition number of the feature map and the variational form can be separately controlled, depending on the dataset embedding design and the number of trainable parameters. Additionally, different feature maps provided by the Qiskit library can be easily swapped in, like the PauliFeatureMap and ZFeatureMap [63]. Compared with QAOA and VQE circuits in Q-gen, VQC circuits contain significantly more CNOT gates due to the entanglement introduced by the feature map circuit, as shown in figure III-K.

By taking advantage of the striking similarities between the connected neurons and the entangled qubits, VQC demonstrates how quantum computing can augment classical computing tasks. The circuit generation parameters in Q-gen can be tuned similarly to the traditional configuration variables like learning rate and batch size.

The most intuitive way to demonstrate Q-gen’s capability as a circuit generator is to show that it can easily generate a large variety of quantum circuits. We realized that organizing the generated circuits into a well-structured dataset and providing open-source access can greatly assist studies that require large-scale quantum circuit testing and benchmarking, or even push new research directions on AI$+$Quantum.

However, the quantum algorithms provided by Q-gen have different structures and targeted applications. A good implementation of Grover’s algorithm should find the marked state in fewer iterations. In contrast, a good quantum key distribution protocol should protect the message from the attacker with minimum encryption overhead. The complex relationships between different quantum algorithms affect how researchers evaluate and optimize the performance of their novel projects. For example, a circuit optimization tool aimed at reducing the number of CNOT gates should not pick the Deutsch-Jozsa algorithm or the quantum key distribution algorithm for testing, due to their low utilization rate of CNOT gates. For studies aimed at improving the QFT algorithm, it is also valuable to test its performance on the QPE algorithm, because it is a direct extension of the QFT algorithm.

Therefore, to facilitate understanding and effective use of all the available algorithms in Q-gen, we summarize them into an organized system with 5 algorithm categories based on their theoretical origin, circuit structure, and targeted application. The complete Q-gen algorithm system is visualized in figure III-M, we also hope this figure can offer a clear and direct introduction for new researchers to dive into quantum computing.

The algorithms under one Q-gen category share similar characteristics, like circuit structure and measurement output pattern. Given the high variability of the available generation parameters, it is possible to generate some outlier circuits that have unrealistic gate arrangements. Still, when creating the Q-gen dataset, we try to keep the parameters reasonable so the generated circuits are close to practical application circuits. The generation details for each algorithm category are explained below:

1. 1.

   The quantum query circuits are relatively easy to generate in terms of average generation time and circuit size. All the oracles in the three algorithms are set to contain random output states, and the oracle of the Deutsch-Jozsa algorithm is always set to be a constant oracle.

2. 2.

   The quantum Fourier transform circuits are significantly harder to generate. The QFT circuits are generated as inverse-QFT and measurement gates to produce verifiable histograms. The QPE circuits have fixed input initialization of $\theta=\frac{1}{8}$, the same phase angle as the T gate. The input number of QFT circuits and the initial guess ($a$) of Shor’s circuits are both randomly generated.

3. 3.

   The quantum search circuits take the longest average time to generate in Q-gen if normalized to the same circuit width. Due to the exponentially growing gate count, it also produces the deepest circuits compared to other algorithm categories. For Grover’s algorithm and the quantum walk algorithm, we specify the number of solutions as $2^{(\text{{problem size}}-2)}$ so that the number of iterations will always be 1. Although the number of solutions is fixed, the individual solution state inside the oracle is still randomly picked. The quantum counting circuits always generate the same number of counting qubits and searching qubits according to problem size.

4. 4.

   The quantum communication circuits have the simplest generation parameters in Q-gen. The quantum key distribution circuits contain interceptions from the attacker, the superdense coding circuits encode half of the qubits, and the quantum teleportation circuits randomly initialize the states to teleport.

5. 5.

   The variational quantum algorithms can produce circuits with vastly different structures depending on the generation parameters. The QAOA circuits are generated with cyclic input graph, and the variational form of the VQE circuits is generated with RY/RZ gates and full entanglement. The repetition number of the variational form for the VQE and VQC circuits is fixed to 1. For all the circuits in this category, the rotation angles of the single-qubit gates are randomly generated.

The minimum problem size is set to 2 for every algorithm (except Shor’s algorithm) because most of them require at least 2 qubits, it is then incremented by 1 for each generation step until the circuit gets too large for generation or simulation. For Shor’s algorithm, its circuit structure is decided by the input number to be factored. We generated Shor’s circuit with a list of problem size from a subset of odd composite numbers, up to 123. Table I summarizes the full generation details and the statistics of the Q-gen quantum circuit dataset.

We want to provide each circuit in the Q-gen dataset with its corresponding noise-free output to form a straightforward input-output pair, either in measurement count or state vector format. This enables researchers to utilize the Q-gen dataset like any machine learning dataset without worrying about the simulation overhead of large circuits, which can easily take hours to days.

For circuit simulation, we obtain the measurement count using the Qiskit Aer simulator running on a rack server with two Xeon Gold 6354 processors and the Nvidia A100 GPU. The simulation times for all the Q-gen dataset circuits are plotted in figure IV-A5. For most algorithms, the simulation time difference between GPU and CPU is negligible. However, the simulation of Shor’s algorithm and quantum teleportation quickly ran out of memory on the CPU. The GPU can significantly outperform the CPU on these two algorithms thanks to its 80GB graphics memory. The state vector simulations are performed on the Qulacs simulator due to its efficiency when running large quantum circuits. The circuits are converted from Qiskit to Qulacs format with the Naniwa converter, and the simulation results can be cross-verified with the state vector output from the Aer simulator.

Another addition to complement the usability of the Q-gen quantum circuit dataset is the algorithm complexity rating, denoted by the star ($\medwhitestar$) symbols in figure III-M. We define an algorithm’s complexity based on its simulation time, generation time, and gate statistics. Note that this rating should not be confused with the circuit complexity mentioned in section III. A quantum circuit’s complexity can be quantified regardless of its application, most commonly using the circuit depth. However, it is hard to come up with a comprehensive and uniform metric to define the complexity of a quantum algorithm across all the fields of quantum computing. Under the Q-gen algorithm category system, the individual algorithms within the same category have the same application and similar circuit structure, so we assign the ratings independently for each algorithm category. This gives researchers a clear hierarchical view of the complexity of the quantum algorithms, and helps software developers quickly grasp the coding difficulty if they want to commit new algorithms to the Q-gen algorithm system.