This paper establishes a fundamental difference between
$\mathbb{Z}$ subshifts of finite type and
$\mathbb{Z}^{2}$ subshifts of finite type in the context of ergodic optimization. Specifically, we consider a subshift of finite type
$X$ as a subset of a full shift
$F$. We then introduce a natural penalty function
$f$, defined on
$F$, which is 0 if the local configuration near the origin is legal and
$-1$ otherwise. We show that in the case of
$\mathbb{Z}$ subshifts, for all sufficiently small perturbations,
$g$, of
$f$, the
$g$-maximizing invariant probability measures are supported on
$X$ (that is, the set
$X$ is stably maximized by
$f$). However, in the two-dimensional case, we show that the well-known Robinson tiling fails to have this property: there exist arbitrarily small perturbations,
$g$, of
$f$ for which the
$g$-maximizing invariant probability measures are supported on
$F\setminus X$.