1. Understanding the fitness consequences of different life histories is critical for explaining their diversity and for predicting effects of changing environmental conditions. However, current theory on plant life histories relies on phenomenological, rather than mechanistic, models of resource production.

2. We combined a well-supported mechanistic model of ontogenetic growth that incorporates differences in the size-dependent scaling of gross resource production and maintenance costs with a dynamic optimization model to predict schedules of reproduction and prolonged dormancy (plants staying below ground for ≥ 1 growing season) that maximize lifetime offspring production.

3. Our model makes three novel predictions: First, maintenance costs strongly influence the conditions under which a monocarpic or polycarpic life history evolves and how resources should be allocated to reproduction by polycarpic plants. Second, in contrast to previous theory, our model allows plants to compensate for low survival conditions by allocating a larger proportion of resources to storage and thereby improving overwinter survival. Incorporating this ecological mechanism in the model is critically important because without it our model never predicts significant investment into storage, which is inconsistent with empirical observations. Third, our model predicts that prolonged dormancy may evolve solely in response to resource allocation tradeoffs.

4. Significance: Our findings reveal that maintenance costs and the effects of resource allocation on survival are primary determinants of the fitness consequences of different life history strategies, yet previous theory on plant life history evolution has largely ignored these factors. Our findings also validate recent arguments that prolonged dormancy may be an optimal response to costs of sprouting. These findings have broad implications for understanding patterns of plant life history variation and predicting plant responses to changing environments.

The data presented in this manuscript were generated using a stochastic dynamic programming (SDP) model. The R script containing the SDP model code is deposited here.

Some of the parameter names used in the R script differ from the names provided in the manuscript. A translation from the parameter names in the model code to those used in the manuscript is provided at the beginning of the R script.