Computing the dimension of real algebraic sets

Let $V$ be the set of real common solutions to $F = (f_1, \ldots, f_s)$ in $\mathbb{R}[x_1, \ldots, x_n]$ and $D$ be the maximum total degree of the $f_i$'s. We design an algorithm which on input $F$ computes the dimension of $V$. Letting $L$ be the evaluation complexity of $F$ and $s=1$, it runs using $O^\sim \big (L D^{n(d+3)+1}\big )$ arithmetic operations in $\mathbb{Q}$ and at most $D^{n(d+1)}$ isolations of real roots of polynomials of degree at most $D^n$. Our algorithm depends on the real geometry of $V$; its practical behavior is more governed by the number of topology changes in the fibers of some well-chosen maps. Hence, the above worst-case bounds are rarely reached in practice, the factor $D^{nd}$ being in general much lower on practical examples. We report on an implementation showing its ability to solve problems which were out of reach of the state-of-the-art implementations.