We consider the problem of reconstructing a sparse signal $x^0\in\R^n$ from a limited number of linear measurements. Given $m$ randomly selected samples of $U x^0$, where $U$ is an orthonormal matrix, we show that $\ell_1$ minimization recovers $x^0$ exactly when the number of measurements exceeds \[ m\geq \mathrm{Const}\cdotμ^2(U)\cdot S\cdot\log n, \] where $S$ is the number of nonzero components in $x^0$, and $μ$ is the largest entry in $U$ properly normalized: $μ(U) = \sqrt{n} \cdot \max_{k,j} |U_{k,j}|$. The smaller $μ$, the fewer samples needed. The result holds for ``most'' sparse signals $x^0$ supported on a fixed (but arbitrary) set $T$. Given $T$, if the sign of $x^0$ for each nonzero entry on $T$ and the observed values of $Ux^0$ are drawn at random, the signal is recovered with overwhelming probability. Moreover, there is a sense in which this is nearly optimal since any method succeeding with the same probability would require just about this many samples.