TernaryKAN: Kolmogorov-Arnold Network with Discrete Ternary Control Points and Regularization by Quantization

We present TernaryKAN, the first Kolmogorov-Arnold Network (KAN) with discrete ternary control points {-1, 0, +1}, stored in 2-bit format. Our 4-phase quantization-aware training (QAT) pipeline yields a counterintuitive result: quantization to ternary control points improves accuracy over the float baseline (94.78±0.25% to 95.35±0.14%) on MNIST, demonstrating a regularization-by-quantization effect not previously reported in KAN literature. The resulting model stores 79,800 ternary control points in 19.95 KB, fitting within the SRAM of mainstream microcontrollers while requiring no floating-point parameter storage.