LOCC protocols with bounded width per round optimize convex functions

We start with the task of discriminating finitely many multipartite quantum states using LOCC protocols, with the goal to optimize the probability of correctly identifying the state. We provide two different methods to show that finitely many measurement outcomes in every step are sufficient for approaching the optimal probability of discrimination. In the first method, each measurement of an optimal LOCC protocol, applied to a $d_{\rm loc}$-dim local system, is replaced by one with at most $2d_{\rm loc}^2$ outcomes, without changing the probability of success. In the second method, we decompose any LOCC protocol into a convex combination of a number of "slim protocols" in which each measurement applied to a $d_{\rm loc}$-dim local system has at most $d_{\rm loc}^2$ outcomes. To maximize any convex functions in LOCC (including the probability of state discrimination or fidelity of state transformation), an optimal protocol can be replaced by the best slim protocol in the convex decomposition without using shared randomness. For either method, the bound on the number of outcomes per measurement is independent of the global dimension, the number of parties, the depth of the protocol, how deep the measurement is located, and applies to LOCC protocols with infinite rounds, and the "measurement compression" can be done "top-down" -- independent of later operations in the LOCC protocol. The second method can be generalized to implement LOCC instruments with finitely many outcomes: if the instrument has $n$ coarse-grained final measurement outcomes, global input dimension $D_0$ and global output dimension $D_i$ for $i=1,...,n$ conditioned on the $i$-th outcome, then one can obtain the instrument as a convex combination of no more than $R=\sum_{i=1}^n D_0^2 D_i^2 - D_0^2 + 1$ slim protocols; in other words, $\log_2 R$ bits of shared randomess suffice.