The exact zero constraint requires real 0.0 to map to an integer zero point Z without quantization error. In affine quantization, r = S(q − Z), so q = Z reconstructs exactly to zero. section 2.2
During calibration, include zero in the real range and choose an integer zero point. For example, extend an observed range of [1.0, 5.0] to [0.0, 5.0]. Widening the range can reduce resolution for the observed values when the integer range is fixed. section 2.2
Use Z when padding an integer tensor to obtain real-zero padding. Exact zeros also preserve ReLU sparsity: hardware that supports zero-skipping can skip them, while small nonzero reconstruction errors can prevent skipping and leak signal. section 2.2
Quantizing ReLU outputs
For unsigned 8-bit quantization of [0.0, 0.0, 0.25, 1.5, 6.0], calibration gives scale 0.023529 and zero point 0. The quantized values are [0, 0, 11, 64, 255]; both zeros reconstruct exactly, while 0.25 reconstructs to about 0.258824. section 2.2
