The widespread adoption of public-key cryptography has enabled the secure transmission of information, particularly over the Internet. However, with the potential development of a quantum computer powerful enough to break current encryption standards, research into post-quantum cryptography—cryptographic protocols resistant to quantum attacks—has flourished over the past two decades. Meanwhile, side-channel attacks, which exploit physical information such as power consumption or electromagnetic emissions, have proven both practical and effective against implementations that lack proper countermeasures. As post-quantum cryptography nears the final stages of its standardization process, addressing the need for secure key exchange mechanisms and digital signature algorithms, many important questions continue to drive ongoing research. Notably, since the initial design of standardized post-quantum schemes did not prioritize side-channel security, the systematic study of this issue remains an active and evolving area of research. In this work, we first conduct a side-channel analysis of several post-quantum schemes currently undergoing standardization. Through this analysis, we observe that while some schemes have structural vulnerabilities to leakage attacks, others demonstrate promising properties for secure implementation. We then present insights on how to design post-quantum protocols that are more efficient to implement with protection against side-channel attacks. We then examine the side-channel security of post-quantum schemes at the operation level, showing that the security of some of the most critical operations can be reduced to the hardness of certain physical learning problems. We provide several examples of this synergy, including the use of inexact computing to implement key operations as physical functions and the application of existing learning problems for side-channel analysis of post-quantum schemes. Finally, after extending our hardness analysis of various physical learning problems, we take an initial step towards a generalized approach for learning problem reductions. We introduce a new generalized unstructured learning problem and provide the first reduction from this problem to the classical Learning with Errors (LWE) problem.