Stacking multiple small convolutions can achieve an equal or larger receptive field compared to a single large convolution, while improving parameter efficiency and increasing network depth. For example, applying two successive $3 \times 3$ convolutions covers the same receptive field as a single $5 \times 5$ convolution. While a single $5 \times 5$ convolution requires $25 \cdot c^2$ parameters (where $c$ is the number of channels), three successive $3 \times 3$ convolutions provide an even larger receptive field while using a comparable number of parameters ($3 \cdot 9 \cdot c^2 = 27 \cdot c^2$). This efficiency demonstrates that deep and narrow networks significantly outperform shallow and wide counterparts, establishing stacked $3 \times 3$ convolutions as a standard architectural design.