03 / PHYSICAL CONSTRAINTS
What must quantum hardware deliver for AI?
Quantum hardware must execute a useful workload accurately enough and cheaply enough to justify its place in the full AI system. Qubit count alone cannot answer that question.
Why is a physical qubit not the same as a useful logical qubit?
A physical qubit is a component of the device, while a logical qubit is information encoded across physical resources to protect it against errors. The distinction separates raw hardware scale from the reliability available to an algorithm.
Real devices suffer imperfect gates, measurement errors and unwanted interactions with their environment. Quantum error correction gathers information about errors without directly reading out the protected quantum state. Doing that repeatedly requires additional qubits, control operations and classical computation. Consequently, an algorithm described using logical qubits needs a separate physical resource estimate. There is no universal conversion ratio that remains valid across codes, devices, target accuracies and workloads.
What does the noisy intermediate-scale regime imply?
The noisy intermediate-scale regime describes quantum processors whose size makes them interesting to study but whose noise constrains the computations they can sustain. Preskill’s Quantum Computing in the NISQ era and beyond sets out the opportunities and limitations behind that description.
For quantum machine learning, shallow circuits may reduce exposure to noise, but shallowness can also restrict what the model represents. Repeating a noisy experiment supplies more samples; it does not automatically remove systematic bias. A hybrid optimiser may compensate for some device effects, yet a model trained on one calibration can behave differently after the hardware drifts. Those issues belong in evaluation, not only in a hardware appendix.
How does a surface code protect information?
A surface code protects logical information through local checks that reveal patterns associated with errors. A classical decoder interprets the resulting syndrome measurements and infers how to track or correct the damage.
The review Surface codes: Towards practical large-scale quantum computation explains the architecture and its resource demands. Suppression of logical errors depends on operating in an appropriate regime, increasing protection and controlling relevant error mechanisms. Correlated noise and leakage can complicate simplified models. Reliable storage is also not the whole task: a useful computation needs logical operations, state preparation and readout. Some required operations can demand substantial supporting machinery beyond the protected data qubits.
Is error mitigation a substitute for error correction?
Error mitigation is not a general substitute for error correction because it usually estimates less-biased results from noisy experiments rather than protecting an arbitrarily long computation. Its usefulness depends on the observable, noise assumptions and sampling overhead.
Zero-noise extrapolation, for example, estimates behaviour towards a lower-noise limit using experiments at different effective noise levels. That procedure requires modelling choices and can magnify uncertainty. Error correction instead aims to suppress logical errors through encoding and repeated checks. For a learning experiment, either approach should be assessed against the quality of the final prediction and its total cost. Reporting a cleaner expectation value without the extra circuit executions gives an incomplete comparison.
How do hardware platforms change the trade-offs?
Hardware platforms change gate implementation, connectivity, control and error behaviour, so the same abstract circuit can have different physical costs. Superconducting circuits, trapped ions, neutral atoms and photonic approaches should not be compared through one headline specification.
Superconducting designs use engineered electrical circuits; trapped-ion systems manipulate charged atoms; neutral-atom systems organise uncharged atoms, often with optical control; photonic approaches encode and process information using light. Within each family, architectures vary. A circuit with interactions between distant logical partners may incur routing costs on one device and different control costs on another. Appropriate comparison therefore starts with the intended workload and a compilation strategy, not a presumed universal winner.
What belongs in a hardware resource estimate?
A hardware resource estimate should include the target output accuracy, logical operations, physical protection, runtime and classical support needed to finish the task. It should also make the noise model and compilation assumptions visible.
For AI, repeated training or inference calls can multiply costs that look modest for one circuit. Queueing, reset, measurement and data movement can matter alongside gate execution. A proposal for fault-tolerant quantum simulation should be judged differently from a small variational demonstration. Neither should be turned into an AGI delivery schedule. Hardware capability is one necessary input to a proposed quantum-assisted system; the quality of its learning objective and the usefulness of its outputs remain separate questions.
Where can classical AI help with these constraints?
Classical AI can help interpret syndrome data, tune controls and prioritise experiments. The AI for quantum research guide examines those roles and the evidence needed to trust them.