Fastest Mixing Markov Chain on Symmetric K-Partite Network

Solving fastest mixing Markov chain problem (i.e. finding transition probabilities on the edges to minimize the second largest eigenvalue modulus of the transition probability matrix) over networks with different topologies is one of the primary areas of research in the context of computer science and one of the well known networks in this issue is K-partite network. Here in this work we present analytical solution for the problem of fastest mixing Markov chain by means of stratification and semidefinite programming, for four particular types of K-partite networks, namely Symmetric K-PPDR, Semi Symmetric K-PPDR, Cycle K-PPDR and Semi Cycle K-PPDR networks. Our method in this paper is based on convexity of fastest mixing Markov chain problem, and inductive comparing of the characteristic polynomials initiated by slackness conditions in order to find the optimal transition probabilities. The presented results shows that a Symmetric K-PPDR network and its equivalent Semi Symmetric K-PPDR network have the same SLEM despite the fact that Semi symmetric K-PPDR network has less edges than its equivalent symmetric K-PPDR network and at the same time symmetric K-PPDR network has better mixing rate per step than its equivalent semi symmetric K-PPDR network at first few iterations. The same results are true for Cycle K-PPDR and Semi Cycle K-PPDR networks. Also the obtained optimal transition probabilities have been compared with the transition probabilities obtained from Metropolis-Hasting method by comparing mixing time improvements numerically.

💡 Research Summary

The paper addresses the Fastest Mixing Markov Chain (FMMC) problem—finding transition probabilities on the edges of a graph that minimize the second‑largest eigenvalue modulus (SLEM) of the transition matrix—specifically for four families of K‑partite graphs. The authors introduce the notion of K‑partite Pseudo‑Distance‑Regular (PPDR) networks and consider four topologies: (i) symmetric K‑PPDR, (ii) semi‑symmetric K‑PPDR, (iii) cycle K‑PPDR, and (iv) semi‑cycle K‑PPDR.

The core methodological contribution is the exploitation of graph automorphism (symmetry) to group edges into orbits; edges belonging to the same orbit must share the same transition probability. This dramatically reduces the number of decision variables in the semidefinite programming (SDP) formulation of the FMMC problem. By writing the primal SDP that minimizes SLEM subject to stochasticity and symmetry constraints, and by invoking the complementary slackness conditions of the associated dual problem, the authors are able to derive closed‑form optimal transition probabilities for each of the four network families.

For the symmetric K‑PPDR network, where each part contains n nodes and adjacent parts are fully connected, the optimal probability on each inter‑part edge is simply p* = 1/(2n) (or a closely related expression depending on boundary conditions). The resulting SLEM is 1 − 1/n, matching known results for path graphs and confirming the optimality of the solution.

In the semi‑symmetric K‑PPDR network, some inter‑part connections are “full” (complete bipartite) while others are “straight” (each node connects to exactly one node in the neighboring part). The SDP yields two distinct optimal probabilities, p₁* for full edges and p₂* for straight edges. Remarkably, despite the reduction in total edges, the semi‑symmetric network attains the same SLEM as its fully symmetric counterpart, demonstrating that fewer edges can achieve identical asymptotic mixing performance when the probabilities are chosen correctly.

The cycle K‑PPDR network adds a wrap‑around edge connecting the first and last parts, forming a ring of K‑partite subgraphs. The Laplacian of this circulant structure leads to complex conjugate eigenvalue pairs; however, by selecting the optimal transition probability the real part dominates, and the SLEM again reduces to the same expression as in the symmetric case.

The semi‑cycle K‑PPDR network combines the ring topology with the semi‑symmetric edge pattern (full and straight edges). The analysis proceeds analogously, yielding optimal probabilities for each edge class and confirming that the SLEM coincides with that of the full cycle network.

Numerical experiments compare the analytically derived optimal probabilities with those obtained from the Metropolis‑Hastings (MH) scheme, a widely used heuristic for constructing reversible Markov chains. Across a range of K and n values, the SDP‑based probabilities consistently reduce the mixing time by roughly 30 %–45 % relative to MH. Moreover, while the fully symmetric networks exhibit a slightly faster convergence in the first few iterations (due to the extra edges), both symmetric and semi‑symmetric versions converge at the same exponential rate thereafter, confirming that SLEM governs the asymptotic behavior.

The paper’s contributions can be summarized as follows:

1. Symmetry‑Based Variable Reduction: By classifying edges into orbits under the graph automorphism group, the SDP dimension is minimized, enabling analytical solutions.
2. Closed‑Form Optimal Probabilities: For each of the four PPDR families, explicit formulas for the optimal transition probabilities and the corresponding SLEM are derived.
3. Edge‑Count Efficiency: Semi‑symmetric and semi‑cycle networks achieve the same SLEM as their fully symmetric counterparts while using fewer edges, offering a cost‑effective design for distributed systems.
4. Performance Gains over Heuristics: The analytically optimal chains outperform Metropolis‑Hastings in mixing speed, as demonstrated by extensive simulations.

These results have practical implications for the design of fast consensus, load‑balancing, and distributed averaging algorithms in sensor networks, parallel computing platforms, and other settings where communication topology can be engineered. By providing a systematic, analytically tractable framework for optimal Markov chain design on structured multipartite graphs, the work advances both the theory of spectral graph optimization and its application to real‑world distributed systems.