Quantum Program Synthesis Through Operator Learning and Selection

Abstract

Programming for quantum computers is complicated and time consuming, because quantum operations are counterintuitive and their combined effects are difficult to understand. Existing tools allow automatic synthesis of quantum programs, which releases the burden of handwriting. However, many existing systems arrange predetermined operators in successive manner to gradually reduce the gap with requirements; these methods are quick but often produce lengthy programs, and they are difficult to adopt for new operators. Other systems depend on stochastic or heuristic search; they identify near-optimal programs for certain cases, but it is not easy to tune the algorithms for a wide range of cases. We propose a system that produces compact programs for most cases and easily evolves with new operators. The system automatically learns the roles of available operators by composing various possible programs. Based on the knowledge, it selects a subset of operators most appropriate for requirements and uses them to compose a program. The learning is geared toward concise programs; thus, the system tends to produce programs with the fewest operators possible. We implemented the system and evaluated it by synthesizing over 400 programs. In comparison with a state-ofthe-art system, the proposed system produced programs with approximately 40-times fewer operators at the cost of increased synthesis time from seconds to minutes. We also observed that the system successfully adopted new operators by learning their differences from existing operators and utilizing them in right places. We believe that the system provides a basis of utilizing machine learning for quantum program synthesis.

Implementation of Grover’s Iterator for Quantum Searching With an Arbitrary Number of Qubits

“…Search-based techniques [7,[18][19][20][21][22] explore a more diverse space of potential implementations to find a compact one, but the exploration is time consuming.…”

“…There are also existing studies on the implementation of step 2 (the oracle) for various applications, such as cryptanalysis [26]. Previous literature offers implementations for step 3 that are tailored to specific numbers of qubits [5,18,27]. For instance, Fig.…”

“…To minimize operator counts in unitary-matrix synthesis, certain strategies scout through a multitude of potential implementations, favoring those that employ the fewest operators [18,7,19,20]. Given the vastness of the search area, these methods prioritize efficient ways to narrow the possible choices.…”

“…Given the vastness of the search area, these methods prioritize efficient ways to narrow the possible choices. For instance, some methods learn various potential implementations and leverage this knowledge to isolate a subset of operators that appears most efficient for constructing the desired matrix [18]. Alternatively, other methods adopt stochastic search techniques, such as ant colony optimization [7] and genetic algorithms [19,20], which refine an initial program by randomly applying adjustments-adding, replacing, or reordering operators.…”

“…These methods can quickly produce implementations, but they often lead to an exponential increase in the number of operators as the qubit count grows [17]. Other methods pursue a broader exploration of potential implementations, which can result in configurations with fewer operators [18,7,19,20,21,22]. However, this exploration is timeconsuming, making it feasible only for systems with a relatively small number of qubits in a practical timeframe.…”