Let
\unicode[STIX]{x1D707} be the projection on
[0,1] of a Gibbs measure on
\unicode[STIX]{x1D6F4}=\{0,1\}^{\mathbb{N}} (or more generally a Gibbs capacity) associated with a Hölder potential. The thermodynamic and multifractal properties of
\unicode[STIX]{x1D707} are well known to be linked via the multifractal formalism. We study the impact of a random sampling procedure on this structure. More precisely, let
\{{I_{w}\}}_{w\in \unicode[STIX]{x1D6F4}^{\ast }} stand for the collection of dyadic subintervals of
[0,1] naturally indexed by the finite dyadic words. Fix
\unicode[STIX]{x1D702}\in (0,1), and a sequence
(p_{w})_{w\in \unicode[STIX]{x1D6F4}^{\ast }} of independent Bernoulli variables of parameters
2^{-|w|(1-\unicode[STIX]{x1D702})}. We consider the (very sparse) remaining values
\widetilde{\unicode[STIX]{x1D707}}=\{\unicode[STIX]{x1D707}(I_{w}):w\in \unicode[STIX]{x1D6F4}^{\ast },p_{w}=1\}. We study the geometric and statistical information associated with
\widetilde{\unicode[STIX]{x1D707}}, and the relation between
\widetilde{\unicode[STIX]{x1D707}} and
\unicode[STIX]{x1D707}. To do so, we construct a random capacity
\mathsf{M}_{\unicode[STIX]{x1D707}} from
\widetilde{\unicode[STIX]{x1D707}}. This new object fulfills the multifractal formalism, and its free energy is closely related to that of
\unicode[STIX]{x1D707}. Moreover, the free energy of
\mathsf{M}_{\unicode[STIX]{x1D707}} generically exhibits one first order and one second order phase transition, while that of
\unicode[STIX]{x1D707} is analytic. The geometry of
\mathsf{M}_{\unicode[STIX]{x1D707}} is deeply related to the combination of approximation by dyadic numbers with geometric properties of Gibbs measures. The possibility to reconstruct
\unicode[STIX]{x1D707} from
\widetilde{\unicode[STIX]{x1D707}} by using the almost multiplicativity of
\unicode[STIX]{x1D707} and concatenation of words is discussed as well.