A sub-theory is a subset of operations, closed under joint measurements.

Main idea: If a sub-theory is contextual, it needs \(O(n^2)\) space to be classically simulated (vs \(O(n)\) for some noncontextual sub-theories, such as qubit stabilizer codes.) This quadratic scaling matches the Gottesman-Knill algorithm.

- Where else does contextuality imply simulation hardness?
- What about operations that are not closed?