---
title: Min Max Average Cycle Weight Optimization
url: https://www.emergentmind.com/topics/min-max-average-cycle-weight
type: topic
---

# Min Max Average Cycle Weight Optimization

“Min Max Average Cycle Weight” denotes a family of cycle-based optimization objectives centered on extremal cycle weights. In classical graph optimization, the closest standard notion is the **minimum mean-weight cycle**, which asks for a directed cycle minimizing total weight divided by length, \(\mu(C)=w(C)/|C|\) [1307.4473]. In more recent fair-allocation work, the same phrase is used more literally for the problem of choosing an allocation that minimizes the **maximum average weight** of directed cycles in an induced envy graph, that is, \(\min_A MACW(G_A)\) with \(MACW(G)=\max_C \overline w(C)\) [2507.20253]. Related literature also studies **min-max cycle covers**, where one minimizes the maximum total weight of any cycle in a cover rather than a cycle mean [2003.12134].

## 1. Core definitions and objective families

For a weighted directed graph \(G=(V,E,w)\), a cycle \(C\) has total weight
\[
w(C)=\sum_{e\in C} w(e),
\]
length \(|C|\), and mean or average weight
\[
\mu(C)=\frac{w(C)}{|C|}.
\]
The **minimum cycle mean** or **minimum average cycle weight** problem is
\[
\mu^*=\min_C \mu(C).
\]
A dual maximum version is
\[
\mu_{\max}=\max_C \mu(C),
\]
and exact minimum and maximum versions are equivalent under weight negation: if \(w'(e)=-w(e)\), then
\[
\min_C \mu_w(C)=-\max_C \mu_{w'}(C)
\]
[1307.4473].

A distinct but related line studies families of cycles rather than a single cycle. In rooted cycle-cover problems, a cycle cover \(\mathscr C\) has maximum cycle weight
\[
M(\mathscr C)=\max_i w(C_i),
\]
total cycle weight
\[
T(\mathscr C)=\sum_i w(C_i),
\]
and average cycle weight
\[
\bar w(\mathscr C)=\frac{T(\mathscr C)}{q},
\]
where \(q\) is the number of cycles. The rooted min-max cycle cover problem optimizes \(\min_{\mathscr C} M(\mathscr C)\), not \(\bar w(\mathscr C)\), although \(\bar w(\mathscr C)\le M(\mathscr C)\) for every cover [2003.12134].

A third formulation appears in envy-graph optimization. Given an allocation \(A\), the envy of agent \(i\) at agent \(j\) is
\[
E_A(i,j)=v_i(A_j)-v_i(A_i),
\]
which induces a complete directed envy graph \(G_A\). For any directed cycle \(C\), its average weight is
\[
\overline w(C)=\frac{1}{k}\sum_{(u\to v)\in C} w(u\to v),
\]
and the graph functional is
\[
MACW(G)=\max_C \overline w(C).
\]
The optimization problem is then
\[
\min_{A\in\mathcal A} MACW(G_A)
\]
or, with preexisting conditions, \(\min_A MACW(G_A-G_O)\) [2507.20253].

| Objective | Optimization form | Representative setting |
|---|---|---|
| Minimum mean-weight cycle | \(\min_C w(C)/|C|\) | weighted digraphs, random complete graphs |
| Rooted min-max cycle cover | \(\min_{\mathscr C}\max_{C\in\mathscr C} w(C)\) | multi-robot routing |
| Min Max Average Cycle Weight | \(\min_A MACW(G_A)\) | envy graphs in fair allocation |

## 2. Threshold behavior in random complete graphs

A particularly detailed probabilistic theory is available for the minimum mean-weight cycle on the complete graph or complete digraph with i.i.d. exponential edge weights of mean \(1\). In this stochastic mean-field distance model, the natural scale of the minimum average cycle weight is \(1/n\), and Mathieu and Wilson showed a sharp threshold at \(1/(en)\) [1201.3955].

Let \(\mu_{\min}\) denote the minimum mean cycle weight. For the directed complete graph, the limiting probability of a \(c\)-light cycle satisfies
\[
\lim_{n\to\infty}\Pr(\mu_{\min}\le c/n)=
\begin{cases}
1-\exp[-T(c)+c], & c\le 1/e,\\
1, & c>1/e,
\end{cases}
\]
where the tree function is
\[
T(z)=\sum_{k=1}^{\infty} k^{k-1}\frac{z^k}{k!},
\qquad T(z)=ze^{T(z)}.
\]
For undirected complete graphs, the corresponding limit is
\[
1-\exp\bigl[(-T(c)+c+c^2)/2\bigr]
\]
for \(c\le 1/e\), and \(1\) for \(c>1/e\). The limiting distribution is analytic for \(c<1/e\), discontinuous at \(c=1/e\), and equal to \(1\) for \(c>1/e\) [1201.3955].

The same work identified a two-regime picture for the minimizing cycle. If the minimum mean weight is at most \(1/(en)\), the minimizing cycle has length \(\Theta_p(1)\) and fixed-length probabilities \(p_k\) converge to explicit integrals. If the minimum mean weight is larger than \(1/(en)\), then conditional on that event,
\[
\mu_{\min}=\frac{1+o(1)}{en},
\qquad
|C^*|\ge \left(\frac{2}{\pi^2}-o(1)\right)\log^2 n\log\log n
\]
with probability \(1-o(1)\) [1201.3955].

Ding, Sun, and Wilson completed the supercritical analysis. Writing \(W_n\) for the minimum mean weight and \(L_n\) for the length of the minimizing cycle, and conditioning on \(nW_n>1/e\), they proved
\[
nW_n=\frac{1}{e}\left[1+\frac{\pi^2}{2\log^2 n}+O\!\left(\frac{1}{\log^3 n}\right)\right]
\]
with high probability, equivalently
\[
W_n=\frac{1}{ne}\left[1+\frac{\pi^2}{2\log^2 n}+O\!\left(\frac{1}{\log^3 n}\right)\right],
\]
and
\[
L_n\asymp (\log n)^3.
\]
Thus the supercritical minimum average cycle weight sits just above \(1/(ne)\), with an explicit \(\pi^2/(2\log^2 n)\) correction, and the minimizing cycle is polylogarithmically long rather than constant-sized [1504.00918].

The mechanism behind the threshold is a first-moment balance for \(c\)-light \(k\)-cycles together with confinement of an associated exp-minus-one random walk. In the refined supercritical analysis, the principal eigenvalue
\[
\lambda_A=\exp\left\{-\frac{\pi^2}{2A^2}+O\!\left(\frac{1}{A^3}\right)\right\}
\]
of the exp-minus-one walk killed outside \([0,A]\) produces the same \(\pi^2\) correction that appears in \(W_n\) [1504.00918].

## 3. Exact and approximate algorithms for minimum cycle mean

The algorithmic minimum cycle mean problem is standard for directed graphs with integer weights. One formulation uses
\[
G=(V,E,w),\qquad w:E\to\mathbb Z,
\]
and seeks
\[
\mu^*=\min_{C\text{ cycle}}\frac{w(C)}{|C|}.
\]
A dynamic-programming characterization is based on path weights of exact length \(t\): if \(\delta_t(x)\) is the minimum weight of a length-\(t\) path starting at \(x\), then
\[
\frac{\delta_t(x)}{t}\longrightarrow \mu(x),
\]
where \(\mu(x)\) is the minimum mean weight of a cycle reachable from \(x\) [1307.4473].

Classical exact algorithms include Karp’s \(O(mn)\) method, and the literature also includes parametric shortest-path methods such as Karp–Orlin, improved by Young–Tarjan–Orlin [1504.00918]. In the matrix-algebraic direction, the minimum cycle mean problem reduces to min-plus matrix multiplication. One result shows that exact minimum cycle mean is reducible in \(O(n^2)\) time to a logarithmic number of min-plus matrix multiplications of \(n\times n\) matrices, and that for nonnegative weights there is a \((1+\varepsilon)\)-approximation algorithm with runtime
\[
\tilde O\!\left(\frac{n^\omega \log^3(nW/\varepsilon)}{\varepsilon}\right),
\]
where \(W\) is the maximum weight and \(O(n^\omega)\) is the time for ordinary matrix multiplication [1307.4473].

A later development revisited minimum mean cycle on low-diameter graphs. Altschuler and Parrilo gave an approximation algorithm based on a linear programming relaxation, entropic regularization, and reduction to matrix balancing. For graphs with polylogarithmic diameter, the runtime is near-linear in the number of edges, and for complete graphs it is the first algorithm whose dependence on \(n\) is \(\tilde O(n^2)\). The method uses only \(O(n)\) memory beyond reading the input, and returns a cycle \(\sigma\) with
\[
\mu(\sigma)\le \mu^*(G)+\varepsilon
\]
[2004.03114].

The algorithmic distinction between **minimum total cycle weight** and **minimum mean cycle weight** remains important. Strongly polynomial \((1+\varepsilon)\)-approximation schemes are known for minimum-weight cycle via approximate APSP and min-plus methods, but those schemes target total cycle weight, not the normalized mean objective [1907.11078]. This suggests that mean-cycle and sum-cycle problems share algebraic infrastructure but are not interchangeable.

## 4. Min-max cycle covers and worst-cycle objectives

In rooted multi-robot routing, the central optimization target is often not a cycle mean but a worst-route criterion. One such formulation is the rooted min-max cycle cover problem on a complete undirected graph
\[
G=(V,E)
\]
with nonnegative metric edge weights \(w:E\to\mathbb R_+\). The vertex set contains depots \(D\subseteq V\) and inspection sites \(V^- = V\setminus D\). A rooted cycle cover \(\mathscr C=\{C_1,\dots,C_q\}\) must satisfy: \(q\le k\), the cycles are edge-disjoint, each cycle contains exactly one depot, and the union of cycle vertices is all of \(V\) [2003.12134].

For a cycle \(C\),
\[
w(C)=\sum_{e\in\mathcal E(C)} w(e),
\]
and the optimization problem is
\[
\min_{\mathscr C\in\mathcal C^k}\max_{C\in\mathscr C} w(C).
\]
In travel-time applications this minimizes makespan; in energy applications it minimizes the maximum energy used by any robot. The paper explicitly distinguishes this objective from minimizing
\[
T(\mathscr C)=\sum_i w(C_i)
\quad\text{or}\quad
\bar w(\mathscr C)=\frac{1}{q}\sum_i w(C_i),
\]
although every bound on \(\max_i w(C_i)\) automatically bounds \(\bar w(\mathscr C)\) because \(\bar w(\mathscr C)\le M(\mathscr C)\) [2003.12134].

The proposed approximation algorithm constructs a rooted spanning forest, enumerates \(2^{m-1}\) combinations of inter-forest edges when there are \(m\) depots, decomposes heavy trees using a lemma of Khani and Salavatipour, and converts each resulting tree into a rooted cycle by doubling edges and shortcutting via the triangle inequality. The final guarantee is a \((5+\varepsilon)\)-approximation:
\[
\max_{C\in \mathscr C^{\mathrm{alg}}} w(C)\le (5+\varepsilon)\lambda^*,
\]
where \(\lambda^*\) is the optimum max-cycle weight. The complexity is
\[
O\bigl(n^2 + 2^{m-1} n \log(n+k)\bigr),
\]
and for fixed \(m\) and \(\varepsilon\) this is \(O(n^2)\) [2003.12134].

This formulation is sometimes conflated with average-cycle objectives because it controls the worst route among several cycles. The distinction is structural: the paper optimizes **maximum total cycle weight**, not cycle means, and the presence of depots and cover constraints places it closer to min-max routing than to minimum mean cycle theory.

## 5. Envy graphs and the named Min Max Average Cycle Weight problem

A recent paper gives the phrase “The Min Max Average Cycle Weight Problem” a specific meaning in fair allocation. Let \(N\) be a set of agents and \(M\) a set of objects, with \(|N|=|M|=n\). Each agent \(i\) assigns a positive real value \(v_i(o)\) to each object \(o\). An allocation \(A\) is a perfect matching in the complete bipartite graph \(H_v\), and \(A_i\) denotes the object assigned to agent \(i\) [2507.20253].

Given an allocation \(A\), the envy of agent \(i\) at agent \(j\) is
\[
E_A(i,j)=v_i(A_j)-v_i(A_i).
\]
This defines a complete directed envy graph \(G_A\) on \(N\), with arc weights
\[
w_A(i\to j)=E_A(i,j).
\]
For a directed cycle \(C=(v_1\to \cdots \to v_k\to v_1)\), the average weight is
\[
\overline w(C)=\frac{1}{k}\sum_{(u\to v)\in C} w(u\to v),
\]
and the graph parameter is
\[
MACW(G)=\max_C \overline w(C).
\]
The optimization problem is
\[
\min_{A\in\mathcal A} MACW(G_A).
\]

In the **clean-slate** case, where no preexisting conditions are present, the problem collapses to maximum-value bipartite matching. If
\[
\sum_{i\in N} v_i(A_i)
\]
is maximal over all allocations, then every directed cycle in \(G_A\) has nonpositive total weight, hence \(MACW(G_A)\le 0\). Conversely, if an allocation is not a maximum-value matching, then some directed cycle has positive average weight, so \(MACW(G_A)>0\). Therefore an allocation minimizes \(MACW(G_A)\) if and only if it is a maximum-value matching in \(H_v\), yielding polynomial-time solvability via the assignment problem and, for example, the Hungarian algorithm in \(O(n^3)\) time [2507.20253].

The generalization with **preexisting conditions** introduces a fixed envy graph \(G_O\) for the original allocation. The new objective is
\[
\min_{A\in\mathcal A} MACW(G_A-G_O),
\]
where each arc weight in \(G_A-G_O\) is
\[
E_A(i,j)-E_O(i,j).
\]
A concrete \(3\times 3\) example shows that the allocation minimizing \(MACW(G_A-G_O)\) need not be a maximum-value matching once even a single arc of \(G_O\) has nonzero weight. The corresponding polynomial-time solvability question is left open:
\[
\text{Given }G_O,\text{ is there a polynomial-time algorithm for minimizing }MACW(G_A-G_O)\,?
\]
[2507.20253]

This fair-allocation formulation is the most literal instance of a **min-max average cycle weight** objective in the supplied literature: the inner extremum is a maximum average directed-cycle weight, and the outer extremum is a minimization over allocations.

## 6. Complexity frontiers and open directions

The fixed-graph problem \(MACW(G)\) is algorithmically well behaved: the fair-allocation paper notes that there exist strongly polynomial-time algorithms for computing \(MACW(G)\) for a given directed graph [2507.20253]. The harder part arises when one must optimize over a structured family of graphs, whether over allocations, over cycle covers, or over random graph instances with delicate asymptotics.

For **minimum total weight cycle** rather than mean cycle, recent distributed work gives a nearly tight approximation tradeoff in the \(\mathsf{CONGEST}\) model. For undirected weighted graphs, there is a randomized \((k+1)\)-approximation algorithm, for any real \(k\ge 1\), whose round complexity yields a smooth tradeoff between approximation ratio and complexity; when \(k\ge 2\) and \(D=\tilde O(n^{1/4})\), the bound simplifies to
\[
\tilde O\!\left(n^{\frac{k+1}{2k+1}}\right).
\]
Assuming the Erdős girth conjecture, any randomized \((k+1-\varepsilon)\)-approximation requires
\[
\tilde\Omega\!\left(n^{\frac{k+1}{2k+1}}\right)
\]
rounds even on graphs of diameter \(\Theta(\log n)\) [2603.25368]. These results concern minimum total cycle weight, not average cycle weight, but they delineate the current distributed complexity frontier for cycle-weight approximation.

Two persistent conceptual separations run through the literature. First, **total** cycle weight, **mean** cycle weight, and **maximum cycle weight within a family** are mathematically different objectives, even when they are all described informally as cycle-weight optimization. Second, the most refined asymptotic results are highly model-specific: the threshold \(1/(ne)\) and the \(\pi^2/(2\log^2 n)\) correction belong to the stochastic mean-field complete graph [1504.00918], whereas the fair-allocation formulation derives its structure from envy graphs and bipartite matchings [2507.20253].

The main unresolved direction in the explicitly named Min Max Average Cycle Weight problem is the preexisting-conditions case \( \min_A MACW(G_A-G_O)\), where the clean-slate reduction to maximum-value matching fails [2507.20253]. A plausible implication is that future progress will require combining classical cycle-mean computation with optimization over allocations or covers, rather than treating cycle computation and outer optimization as separable tasks.

Source: https://www.emergentmind.com/topics/min-max-average-cycle-weight