1: \begin{abstract}

2: The Viterbi process is the limiting maximum a-posteriori estimate

3: of the unobserved path in a hidden Markov model as the length of the

4: time horizon grows. The existence of such a process suggests that

5: approximate estimation using optimization algorithms which process

6: data segments in parallel may be accurate. For models on state-space

7: $\mathbb{R}^{d}$ satisfying a new ``decay-convexity'' condition,

8: we develop an approach to existence of the Viterbi process via fixed points

9: of ordinary differential equations in a certain infinite dimensional

10: Hilbert space. Bounds on the distance to the Viterbi

11: process show that approximate estimation via parallelization can indeed

12: be accurate and scaleable to high-dimensional problems because the

13: rate of convergence to the Viterbi process does not necessarily depend

14: on $d$. The results are applied to a factor model with stochastic volatility and a model of neural population activity.

15: \end{abstract}

16: