Abstract
<jats:p>Climate impact assessment models require meteorological inputs that are consistent with reference data in both marginal statistics and multivariate dependence (e.g., relationships between temperature, humidity, and precipitation). Some bias adjustment methods correct historical biases in dependence or capture projected changes in dependence, but not both. To address this gap, a multivariate linear correction based on Gaussian optimal transport theory is proposed as a post-processing step after univariate or stationary multivariate bias adjustment. The resulting Monge–Kantorovich mapping adjusts model correlations toward a specified target while allowing dependence to evolve between historical and future periods. By working in the log-Euclidean geometry of symmetric positive-definite matrices, the method represents changes in dependence as additive displacements and then maps the adjusted structure back to the space of valid correlation matrices. The framework is evaluated using a “perfect model” experiment designed to emulate realistic biases in marginal distributions and dependence. Pseudo-observations are drawn from regional climate model output at individual grid cells, while biased pseudomodel inputs are formed from neighbouring spatial means. Results are first demonstrated in historically snow-dominated regions, where warming alters snow cover and drives clear changes in inter-variable relationships, before extending the evaluation to a broader North American domain. Overall, the post-processing step improves upon methods that either inherit biased model dependence or impose stationary dependence, particularly when the model projects robust changes in inter-variable relationships.</jats:p>