What is an ansatz?
In variational quantum algorithms (VQAs), the ansatz corresponds to a parametrized quantum circuit, whose parameters are optimized to solve a task. More generally, an ansatz can represent a quantum state or a mathematical expression.
Usually, when designing an ansatz for a given algorithm, specific properties are considered such as the problem structure. In other words, we choose the form of the ansatz according to our best guess of where the solution lies in the search space. Ansatz choice impacts the trainability and expressivity of the resulting algorithm.
In QML specifically, the structure of the ansatz can be seen as the equivalent to the architecture of a neural network in classical ML.
The plural of ansatz is ansätze.
Frequently asked questions about ansätze
- What are common examples of ansätze?
The hardware-efficient ansatz, the unitary coupled cluster ansatz, or the QAOA ansatz.
- What would a photon-native ansatz be?
By definition, a photon-native ansatz uses gates that are native to photonics. In linear optics, a standard ansatz is a universal interferometer (e.g. rectangular) made of phase-shifters and beam-splitters.
- How do I choose an ansatz for my algorithm?
The choice depends on the problem, available hardware, and trade-offs between expressivity, circuit depth, and noise robustness. It may be inspired by physics, problem structure, or hardware constraints.
- What is the relationship between an ansatz and inductive bias?
In machine learning, inductive bias is the set of assumptions built into a model or algorithm that guide it toward certain solutions over others when learning from data. When designing a QML model, the choice of ansatz will impact the model’s inductive bias by restricting the set of quantum states or circuits explored during optimization.