Measure-Instance EF ↔ Real-Valued Cardinal EF
Theorem. For a divisible measure instance $I : \mathrm{MeasureInstance}\ N\ \Omega$ ([[social_choice.fair_division.divisible.measure_instance]]) with every $I.\mathrm{measure}\ i$ finite, and any allocation $A$, the $\mathrm{ENNReal}$-valued envy-freeness predicate agrees with the real-valued cardinal envy-freeness obtained from $.toCardinalInstance$: $$ I.\mathrm{IsEnvyFree}\ A \;\iff\; (I.\mathrm{toCardinalInstance}).\mathrm{IsEnvyFree}\ A. $$
In Lean:
SocialChoice.FairDivision.Divisible.MeasureInstance.isEnvyFree_iff_toCardinalInstance_isEnvyFree.
Proof
The two predicates compare values of the form $\mu_i(A(j))$ vs $\mu_i(A(i))$:
- The $\mathrm{ENNReal}$ form compares directly: $\mu_i(A(j)) \le \mu_i(A(i))$ in $\mathrm{ENNReal}$.
- The real form compares after
.toReal: $(\mu_i(A(j))).\mathrm{toReal} \le (\mu_i(A(i))).\mathrm{toReal}$.
Both quantities are finite ([IsFiniteMeasure (I.measure i)]) so
measure_ne_top lets ENNReal.toReal_le_toReal translate between the
two forms. Each direction of the iff is a short calculation.
Why this matters
MeasureInstance is the natural input for measure-theoretic algorithms
(cut-and-choose, Dubins–Spanier, Stromquist), which all reason with
$\mathrm{ENNReal}$-valued measures. But the bundled real-valued cardinal
interface
([[social_choice.fair_division.divisible.cardinal_instance]]) is what
downstream call sites prefer (algebra over $\mathbb{R}$ is much easier
than over $\mathrm{ENNReal}$).
This equivalence lets a single theorem stated against either form discharge both — pick the convenient one when authoring, then convert at the boundary.
References
- [AGT Chapter 13] Nisan, Roughgarden, Tardos, and Vazirani, Algorithmic Game Theory.