The Hessian matrix: Visual walkthrough — AI-ML deep dive

Step 1 — A height that depends on two knobs

WHAT. Imagine a landscape. At every spot $(x, y)$ on the flat ground there is a height $f (x, y)$ — how high the land is above that spot. Turning the $x$ -knob walks you east–west; turning the $y$ -knob walks you north–south.

WHY. Before we can talk about curvature we must picture the surface. Every symbol on this page is ultimately a fact about this hill.

PICTURE. The green sheet is $f$ . The flat grid below is the $(x, y)$ ground. A point on the ground pushes straight up to a point on the sheet.

Step 2 — Slope in one direction: the partial derivative

WHAT. Freeze $y$ and wiggle only $x$ . The rate at which height changes is the partial derivative $\frac{\partial f}{\partial x}$ , read "how fast $f$ climbs as I step east". Do the same for $y$ to get $\frac{\partial f}{\partial y}$ .

WHY. A hill's steepness depends on which way you face. To measure it we must pick a single direction and ask the plain 1D question "rise over run" — which is exactly what a partial derivative does.

PICTURE. The red curve is the slice through the surface at fixed $y$ (walking east only). Its tangent's steepness is $\partial f / \partial x$ . The blue curve slices the other way.

Step 3 — Curvature: the slope of the slope

WHAT. A slope can itself change as you walk. Differentiate the east-slope again with respect to $x$ : that is $\frac{\partial ^{2} f}{\partial x ^{2}}$ , the curvature along $x$ . Positive ⇒ the slice bends up like a cup; negative ⇒ bends down like a cap.

WHY. Slope alone cannot distinguish a peak from a valley — at both the slope is zero. What differs is how the slope turns around. Curvature is the smallest tool that sees a bend, so it is exactly the tool we need.

PICTURE. Same east-slice from Step 2, now with three tangent lines. Their steepness increases as we move right — the change in that steepness is the curvature.

Step 4 — Directions interact: the off-diagonal term

WHAT. There is a fourth number: as you step north ( $y$ ), does the east-slope change? That is the mixed partial $\frac{\partial ^{2} f}{\partial y \partial x}$ . A tilted, twisting surface has this term non-zero.

WHY. In 2D directions can be coupled — the east steepness can depend on how far north you are. Ignoring this misses saddles and tilted bowls. So we must record it.

PICTURE. Two east-slices taken at two different $y$ values. Their tangents differ — that difference is the mixed partial. The purple arrow shows the slope changing as $y$ increases.

Step 5 — The curvature of an arbitrary direction: the quadratic form

WHAT. Walk in a general direction $Δ x = (Δ x, Δ y)$ , not just east or north. The bend you feel is $Q (Δ x) = Δ x^{⊤} H Δ x = a Δ x^{2} + 2 b Δ x Δ y + c Δ y^{2} .$

WHY. The four numbers of $H$ let us predict curvature in every direction with one formula — instead of measuring infinitely many slices. This single number $Q$ is what decides the shape.

PICTURE. From the flat spot at the centre, arrows fan out in all directions. Each arrow is coloured by the sign of $Q$ along it — up (mint) or down (coral).

Step 6 — Turn to the natural axes: eigenvalues

WHAT. A symmetric $H$ has two special perpendicular directions — its eigenvectors — along which the twist vanishes. Along eigenvector $i$ the curvature is a pure number, the eigenvalue $λ_{i}$ . In those axes: $Q = λ_{1} u_{1}^{2} + λ_{2} u_{2}^{2} .$

WHY. In the raw $(x, y)$ frame the twist term $2 b Δ x Δ y$ mixes the directions and hides the shape. Rotating to the eigen-axes cancels the mixing, leaving two clean, independent bends. This is exactly why we reach for Eigenvalues and eigenvectors.

PICTURE. The same bowl, but drawn along its own axes $u_{1}, u_{2}$ . Two independent parabolas, one per eigenvalue — no cross-term left.

Step 7 — The verdict: all four cases

WHAT. Combine the signs of $λ_{1}, λ_{2}$ :

$λ_{1}$	$λ_{2}$	every direction	shape	name
$+$	$+$	curves up	cup	minimum
$-$	$-$	curves down	cap	maximum
$+$	$-$	up one way, down the other	Pringle	saddle
$0$	any	flat one way	trough/ridge	inconclusive

WHY. These are all the sign patterns two numbers can have. No hidden fifth case exists — the classification is complete, including the degenerate zero.

PICTURE. Four mini-surfaces side by side: bowl, dome, saddle, and a flat-bottomed trough (the zero-eigenvalue case).

Step 8 — One full loop on a coupled example

WHAT. Take $f = x^{2} + x y + y^{2}$ . Gradient $(2 x + y, x + 2 y)$ is zero only at $(0, 0)$ . Hessian: $H = [2112], D = 2 \cdot 2 - 1^{2} = 3 > 0, a = 2 > 0 \Rightarrow minimum .$ Eigenvalues: $λ = 2 \pm 1 = {1, 3}$ — both positive. ✅

WHY. This is Steps 4–7 in one shot: a real twist ( $b = 1$ ), forecast by the determinant, confirmed by eigenvalues. Two roads, same answer — that agreement is your safety check.

PICTURE. The tilted bowl of $f = x^{2} + x y + y^{2}$ with its two eigen-axes drawn: the steep one ( $λ = 3$ ) and the gentle one ( $λ = 1$ ).

The one-picture summary

The whole pipeline: height → slopes (gradient) → curvatures (Hessian) → quadratic form → eigenvalues → shape verdict. Each arrow is one of our steps.

Recall Feynman retelling — say it back in plain words

I start with a hill whose height depends on two knobs. I ask "how steep, facing east? facing north?" — those two slopes are the gradient, and at a flat spot they're both zero. But flat could be a peak, a pit, or a saddle, so I ask the deeper question: as I walk, does the slope itself bend up or down? Those bends — three distinct ones because east and north can twist into each other — I stack into the Hessian. To know the bend in any direction I sandwich a direction between two copies of the Hessian; that number is the quadratic form. To read it cleanly I rotate to the surface's own two axes, where the twist disappears and only two pure bends remain: the eigenvalues. Both up means a bowl (minimum), both down a dome (maximum), one each a saddle, any zero and I can't tell. For $2 \times 2$ I skip the eigenvalues: a positive determinant says the bends agree in sign, and a single diagonal entry tells me whether that shared sign is up or down.

Recall Quick checks

Why is $H$ symmetric? ::: The twist "east-then-north" equals "north-then-east" (Schwarz), so $b = b$ . At a critical point, what happened to the linear Taylor term? ::: The gradient is $0$ , so it vanishes and only the curvature term survives. $D = det H > 0$ and $a < 0$ — what shape? ::: A maximum (both eigenvalues negative). One zero eigenvalue means? ::: Inconclusive — a flat direction, need higher-order terms.

Related: Newton's method uses $H^{- 1}$ to jump to the bottom of this quadratic bowl; Convex functions are exactly those whose Hessian is positive semidefinite everywhere; when eigenvalues differ wildly, Gradient descent zig-zags down the thin valley.