1. We start with a simple image of size 800 by 600, which we interpret as a matrix with entries one (black) or zero (white). This is a prototypical 'geometrical' image, expected to have an extremely sparse representation in curvelets.

---

2. We compute the Curvelet transform of the image and put to zero all the coefficients except the 5000 largest ones (in modulus) at the finest scale. This corresponds to a fraction of 0.14 percent of all the coefficients (take this info with a grain of salt since we started with a mostly empty image.) The inverse Curvelet transform is then applied and, after taking the real part, gives the above image. The most negative values are in white and the most positive values in black.