What the paper shows: The authors study how many copies of an unknown n-qubit quantum state you need to test or learn it when the algorithm can keep only k qubits of coherent quantum memory between rounds. They prove that testing whether a state is a stabilizer state needs Θ(n−k) copies, and that non-adaptive learning of a stabilizer state needs Θ(n2/k) copies. A striking consequence is that even with almost all qubits kept coherent (for example k = 0.99 n) you still need Θ(n) copies to test stabilizer states. This means the previously known constant-copy tester, which used six copies, really needed full coherent memory across a whole copy of the state.

What the researchers did: The work sets up a streaming model. In each round an algorithm receives a fresh copy of the unknown n-qubit state. The algorithm may perform operations on that copy together with up to k qubits it kept coherent from before. All other qubits can be measured and discarded. The authors give matching upper and lower bounds for the number of copies needed for testing and for non-adaptive learning in this model. The testing upper bound scales as O((n−k)/ε) for distinguishing exact stabilizer states from states that are ε-away, while every tester needs Ω(n−k) copies. For learning, they give an algorithm using O(n2/k) copies and prove any non-adaptive learner needs Ω(n2/k).

How the methods work at a high level: Two ideas appear repeatedly. First, Bell sampling is used as a basic primitive. Bell sampling is a two-copy measurement that pairs corresponding qubits from the two copies and measures each pair in the Bell basis; implementing it usually requires keeping an entire copy coherent. Second, the authors connect testing to the hidden-shift problem, a pattern-finding task in computer science, and use that to build the tester when memory is limited. For lower bounds they develop a new average-case argument about likelihood ratios, using combinatorics of a mathematical object called the stochastic orthogonal group. In the learning algorithm they repeatedly Bell-sample blocks of k qubits, which explains the n2/k scaling.