Accepted

Binary neural networks (1-bit) lose significant expressivity. Floating-point (32-bit) is energy-prohibitive for edge deployment.

Use ternary state space {+1, 0, -1} as the fundamental computation unit.

Metric	Binary	Ternary	FP32
Values	2	3	2^32
Bits/trit	1	1.58	32
Energy/op	~0.1 pJ	<1 pJ	~3.7 pJ
Expressivity	Low	High	Very High

Ternary achieves the best energy-expressivity tradeoff for edge AI inference.

• All arithmetic must be GF(3) (mod 3)
• Hardware needs ternary ALU or software emulation
• 5-trit packing achieves 99.06% entropy efficiency