JPEG images become an important subject of research on reversible data hiding (RDH) due to the widespread application of image compression standards. To achieve better embedding performance, existing RDH approaches for JPEG images usually evaluate each frequency band based on distortion cost functions to optimize the modification order of discrete cosine transform (DCT) coefficients across inter-block regions. Additionally, many approaches perform parallel data embedding across multiple histograms to further optimize expansion coefficients, thereby enhancing embedding efficiency. However, existing approaches generally calculate the distortion cost for the entire frequency band and fail to consider the intra-band distribution differences of DCT coefficients. Moreover, the frequency band length and expansion coefficients are often optimized separately. Specifically, frequency band length determines the number of DCT coefficients available for embedding, while the expansion coefficient controls the embedding capacity and embedding distortion per coefficient. As a result, it is difficult to obtain optimal embedding performance with such separate optimization. To this end, this study constructs an adaptive multiple histogram modification (MHM) mapping framework based on joint optimization. By introducing a dynamic distortion cost function and a joint optimization strategy for frequency band length and expansion coefficients, the overall embedding performance is significantly improved. Additionally, by reconstructing the distortion cost function and accounting for the bias caused by variations in the distribution differences of different frequency band lengths and coefficients, the potential distortion costs under each combination of frequency band partitioning and expansion coefficients are dynamically calculated, thus enhancing the accuracy of distortion cost functions. Meanwhile, a joint optimization strategy is introduced to address the limitations of independent optimization of frequency bands and expansion coefficients in traditional methods. Expansion coefficients are dynamically allocated according to the local characteristics within each frequency band, and different combinations of frequency band lengths and expansion coefficients are determined through rate-distortion collaboration to realize layer-by-layer embedding and thus effectively improve embedding efficiency and minimize unit distortion cost while satisfying a given embedding capacity. Experimental results on the USC-SIPI and BOSSbase v1.01 datasets show that the proposed algorithm not only maintains a relatively high embedding capacity but also effectively improves image quality and achieves better performance in file size increase, outperforming various state-of-the-art RDH algorithms.