On the log-Sobolev constant for the simple random walk on the \(n\)-cycle: The even cases (Q1406355)

The authors consider a simple random walk on the \(n\)-cycle \(Z_n=\{1,2,\dots ,n\}\) with Markov kernel \(K\) given by \(K(x,x\pm 1)=1/2.\) Its unique stationary distribution \(\pi\) is the uniform distribution on \(Z_n\). They show that the exact value of the log-Sobolev constant \(\alpha\), which is the infimum of the quotient \({\mathcal E}(f,f)/{\mathcal L}(f)\) of the Dirichlet form \({\mathcal E}(f,f)\) associated with \((K,\pi)\) and some entropy-like quantity \( {\mathcal L}(f)\), taken over all real functions \(f\) on \(Z_n\), is equal to \(\alpha = (1-\cos(2\pi/n))/2\) if \(n\) is even and \(n\geq 4\). (The exact value of the log-Sobolev constant is well-known for \(n\leq 4\).) For this, they use the well-known inequality \( 2\alpha\leq \lambda\), which relates the log-Sobolev constant \(\alpha\) to the spectral gap \(\lambda f(K,\pi)\), and another result of \textit{O. S. Rothaus} [J. Funct. Anal. 42, 110-120 (1981; Zbl 0471.58025) and ibid. 39, 42-56 (1980; Zbl 0472.47024)] which states that \(2\alpha <\lambda\) implies the existence of a certain non-constant positive function \(f\) such that the infimum in the definition of the log-Sobolev constant is attained at \(f\). The authors show that such a function cannot exist in the above setting.