Improved Rounding for Parallel Repeated Unique Games

We show a tight relation between the behavior of unique games under parallel repetition and their semidefinite value. Let G be a unique game with alphabet size k. Suppose the semidefinite value of G, denoted sdp(G), is at least 1− ε. Then, we show that the optimal value opt(G) of the `-fold repetition of G is at least 1 − O( √ `ε log k). This bound confirms a conjecture of Barak et al. (2008), who showed a lower bound that was worse by √ `ε log(1/ε). A consequence of our bound is the following tight relation between the semidefinite value of G and the amortized value opt(G) := sup`∈IN opt(G`)1/`, sdp(G) k) ≤ opt(G) ≤ sdp(G) . The proof closely follows the approach of Barak et al. (2008). Our technical contribution is a natural orthogonalization procedure for nonnegative functions. The procedure has the property that it preserves distances up to an absolute constant factor. In particular, our orthogonalization avoids the additive increase in distances caused by the truncation step of Barak et al. (2008).