Résumé
Interval arithmetic enables rigorous bounding of rounding errors, but standard representations require storing two floating-point numbers per interval, which increases memory costs and data transfer in large-scale computations. We propose a compressed interval representation inspired by the FP-ANR format, encoding both the center and the radius within a single floating-point word while preserving the strict inclusion of the original interval. We present an efficient algorithm to convert center-radius intervals into this format with minimal over-approximation. Its applications to the interval matrix product and the interval Newton method demonstrate the practical benefits of this representation. Additionally, the proposed approach enables the integration of mixed-precision computations, paving the way for scalable and memory-efficient interval arithmetic in scientific computing.