arXiv2026
We study the Erdős-Moser equation in arithmetic progressions, $\sum i^k=m^k$, where $i$ runs over the positive integers $i<m$ with $i\equiv b \pmod q$. We prove that a solution has at most $\lfloor k/\log 2+3/2\rfloor$ terms for all $k<10^{55}$, and at most $k/\log 2+5/2$ terms in general; this linear bound is best possible for real solutions and replaces the previously known exponential bound $(k+1)2^{2k+1}$. For $k=3$ we obtain a complete classification: solutions with exactly $n$ terms exist if and only if $n\in\{3,4,5\}$, and in each case there are infinitely many primitive solutions. For even exponents there are no solutions with three terms, and there are no solutions at all for $k=4$, for every $k=2^s\ge 4$, and for every even $k$ with $4\le k\le 24$. The key analytic input is a uniform estimate for the classical equation $1^k+\dots+(m-1)^k=m^k$: the real exponent $κ_m$ for which $1^κ+\dots+(m-1)^κ=m^κ$ satisfies $κ_m=\log 2\,(m-3/2-γ_m/m)$ with $0<γ_m<1/(8\log 2)$ for every $m\ge 3$. The proofs rest on a method based on centres of mass, Steiner's parallel-axis theorem and what we call the half-mass principle. The same framework identifies the constants of a general class of balance equations through the moments of a mass distribution: $3/2$ is a Steiner quotient, and $25/12-3\log 2=1/12-f''(\log 2)/2$ with $f(λ)=λ/(e^λ-1)$, whose Bernoulli expansion explains why this constant is so small.