When a Markov chain runs, sliding windows of length n produce occupation-count vectors; these define a projected count kernel. The paper asks how much the spectral gap shrinks when we look at these window counts, relative to the original chain's gap. The authors study the infimum of the normalized ratio n·Gap(\widetilde P_n)/Gap(P) over all m-state irreducible reversible kernels and all window lengths n≥2.
Their main result pins this optimal constant c_m^* between 1/(1080m) and q_{m-2}, where q_0=1/4 and q_{r+1}=q_r(1-q_r). Since q_{m-2} behaves like 1/(m+log m), the constant is exactly of order 1/m, meaning the reciprocal optimal comparison cost grows linearly with m. In the two-state case they determine the exact values: c_2^*=1/4 and C_2^\mathrm{opt}=4.
The lower bound holds even for n=1 and is uniform over sparse or periodic kernels. The proof combines short-window decorrelation with an averaged anchor-excursion decomposition, a Green-kernel hitting estimate, and a stopped Carleson–Hardy inequality. A nested rare-state construction produces finite-state witnesses whose normalized Rayleigh quotients approach the upper bound. For every fixed finite irreducible reversible aperiodic kernel, the window-count gap is also shown to decay as \Theta_P(n^{-1}).