The chaotic distributed dynamics of an exothermic reaction on a catalytic wafer is shown to evolve on a low-dimensional invariant manifold. The dominant spatial patterns during CO oxidation on a Rh black/SiO2 wafer is obtained by carrying out a statistical Karhunen-Loeve (KL) analysis on the recorded IR thermograph in video images. The entire video sequence is then projected onto the phase space spanned by the 5 leading "empirical eigenfunctions" from the analysis. Statistical properties of the projected trajectories, such as the Lyapunov exponent, is of low dimensions (n≪5) and is fully captured by the KL technique. The more conventional delayed-embedding phase-space reconstruction technique from a point measurement is shown to be far less robust.