Residual-network loss contours

Explore fL(z) = |(1 + z/L)L + 1|² in the complex plane. Every contour is the analytic inverse image of a circle; both axes use the same scale.

1–40, 50, 75, 100, 200, 500, 1000, ∞

Logarithmic: 1/256× to 64×

Visible features
This interactive plot needs JavaScript. The article gives the critical points and their classification explicitly.

Drag to pan; scroll or pinch to zoom at the pointer. Keyboard: arrow keys pan, +/− zoom, Home resets. Changing depth keeps your view.

0.10.51251025Minimum: f = 0Saddle: f = 1

Dashed: f = 1, the saddle level when 2 ≤ L < ∞. For L ≥ 3, its 2L branches meet at a degenerate saddle, even when loss differences there are too small for floating-point evaluation. At L = ∞ the experiment uses ez exactly, with infinitely many minima and no finite saddle.