This paper introduces a novel deep reinforcement learning framework for portfolio optimization that integrates Physics-Informed Kolmogorov-Arnold Networks (PIKANs). The core innovation lies in replacing traditional multilayer perceptrons (MLPs) with Kolmogorov-Arnold Networks (KANs) in both actor and critic components, which use learnable B-spline univariate functions for parameter-efficient and interpretable function approximation. Additionally, a physics-informed regularization loss based on Newton's second law is incorporated during actor updates. This loss enforces second-order temporal consistency by penalizing discrepancies between observed market acceleration (derived from return changes) and the acceleration implied by the agent's portfolio actions (treated as forces). The framework is evaluated on three equity markets—the United States (S&P 100), Vietnam (VN100), and China (CSI 300)—using historical daily data from 2015 to 2025. Experimental results demonstrate that PIKAN-enhanced versions of A2C, DDPG, PPO, and TD3 consistently outperform their standard counterparts and classical online portfolio selection methods. The PIKAN agents achieve higher cumulative and annualized returns, superior Sharpe and Calmar ratios, and more favorable drawdown characteristics, indicating more stable training and better risk-adjusted performance. The approach is particularly effective in highly dynamic and noisy financial markets, where conventional DRL often suffers from instability and poor generalization. The work bridges physics-informed machine learning and financial reinforcement learning, offering a structured inductive bias that improves policy learning and portfolio performance.