## 1. Precise restatement The claim is an **existence statement about a parameterized family**, not a claim about arbitrary pairs of Markov chains. For every \(0<\varepsilon<1\), there are two strictly positive, symmetric, row-stochastic matrices \(P_A,P_B\) on the same four labeled states such that: - Both have stationary distribution \(\pi=(1/4,1/4,1/4,1/4)\). - They have identical eigenvalues and identical **stationary Shannon entropy rates**. - For every mean-zero state function \(f\), meaning \(\sum_i f_i=0\), \[ \|P_Af\|_2,\ \|P_Bf\|_2\le a_\varepsilon\|f\|_2, \qquad a_\varepsilon=\frac{1-\varepsilon}{2+2\varepsilon}<\frac12. \] Thus “contract by at least a factor of two” refers to the Euclidean norm, equivalently \(L^2(\pi)\). - Their equal-weight arithmetic average is \[ M=\frac{P_A+P_B}{2}=\frac14\mathbf1\mathbf1^\top. \] Every initial distribution therefore becomes uniform after one step. - Their equal-weight **softmax-logit average** is \[ G_{ij} =\frac{\sqrt{(P_A)_{ij}(P_B)_{ij}}} {\sum_k\sqrt{(P_A)_{ik}(P_B)_{ik}}}. \] This is normalized geometric averaging, not averaging binary log-odds. - For any fixed total-variation tolerance \(0<\delta<1/2\), the worst-case mixing time of \(G\) tends to infinity as \(\varepsilon\downarrow0\). The constituent spectra and entropy rates agree **within each pair**; they need not remain constant as \(\varepsilon\) changes. ## 2. Check of the argument and independent derivation ### Normalization, positivity, and stationarity Set \[ c=\frac{1+\varepsilon}{2},\qquad d=2+2\varepsilon. \] Every row of either weight matrix sums to \[ 2c+1+\varepsilon=d. \] All entries are positive. The normalized matrices remain symmetric, so they are doubly stochastic and have uniform stationary distribution. These assertions are correct. ### Permutation equivalence and entropy rates Swapping states \(3\) and \(4\), in both rows and columns, transforms \(P_A\) into \(P_B\). They therefore have the same spectrum and stationary entropy rate. More directly, every row of either matrix has the same multiset of probabilities: \[ \left\{\frac14,\frac14, \frac{1}{2(1+\varepsilon)}, \frac{\varepsilon}{2(1+\varepsilon)}\right\}. \] Consequently their row entropies—and hence stationary entropy rates—are identical. Using natural logarithms, the common rate is \[ h=\frac32\log 2+ \frac12 H_{\mathrm b}\!\left(\frac{\varepsilon}{1+\varepsilon}\right), \] where \(H_{\mathrm b}(x)=-x\log x-(1-x)\log(1-x)\). ### Eigenvalues and contraction Use the orthogonal basis \[ u=(1,1,1,1),\quad v=(1,1,-1,-1),\quad w=(1,-1,1,-1),\quad z=(1,-1,-1,1). \] Direct multiplication gives, with \(a=(1-\varepsilon)/d\), \[ \begin{array}{c|rrrr} &u&v&w&z\\ \hline P_A&1&0&a&-a\\ P_B&1&0&-a&a \end{array} \] where each entry denotes the corresponding eigenvalue. The mean-zero subspace is spanned by \(v,w,z\). Symmetry therefore gives the exact operator norm \[ \left\|P_A\big|_{\mathbf1^\perp}\right\|_{2\to2} = \left\|P_B\big|_{\mathbf1^\perp}\right\|_{2\to2} =a<\frac12. \] The supporting text’s spectral and contraction claims are correct. ### Arithmetic average Within either block, entries are already \(c/d=1/4\). Between blocks, averaging gives \[ \frac{1+\varepsilon}{2d}=\frac14. \] Thus \(M=\mathbf1\mathbf1^\top/4\), exactly as claimed. ### Logit average Averaging \(\log(P_A)_{ij}\) and \(\log(P_B)_{ij}\), exponentiating, and normalizing gives normalized geometric means. The common factor \(1/d\) cancels. Writing \(s=\sqrt{\varepsilon}\), \[ G=\frac{1}{2(c+s)} \begin{pmatrix} c&c&s&s\\ c&c&s&s\\ s&s&c&c\\ s&s&c&c \end{pmatrix}. \] Its eigenvalues are \[ 1,\quad \frac{c-s}{c+s},\quad 0,\quad 0. \] The nonconstant nonzero eigenvalue is therefore \[ r=\frac{c-s}{c+s} =\frac{1+\varepsilon-2\sqrt{\varepsilon}} {1+\varepsilon+2\sqrt{\varepsilon}} =\frac{(1-\sqrt{\varepsilon})^2} {(1+\sqrt{\varepsilon})^2}. \] This calculation is also correct. ### Independent reconstruction To verify the mechanism without relying on the proposed weights, define \[ \Pi=\frac{uu^\top}{4}, \qquad D=\frac{ww^\top-zz^\top}{4}. \] For any \(0