Foundations — Gradient vector ∇f — definition, properties

Before you can trust that one arrow, you must be able to read every mark on the page. Below is every symbol the parent note uses, built from the ground up, each one earning the next.

0. The stage: a function $f(x,y)$

Picture a flat table (the $x$–$y$ plane). Above every point on the table there is a height. Sweep over all points and those heights trace out a hill — a curved surface floating above the table.

• $x$ and $y$ ::: the two coordinates of where you stand on the flat map (east–west and north–south).
• $z=f(x,y)$ ::: the height of the hill directly above where you stand.

Why we need this: the whole topic is about slopes of this hill, and you cannot talk about a slope until you have a surface to slope.

1. The point $\mathbf{p}$ and bold letters

So when the parent writes $f(\mathbf{p})$ it just means "the height at the spot $\mathbf{p}$." Bold letters like $\mathbf{p}$, $\mathbf{x}$, $\mathbf{v}$ each secretly hold two numbers.

• $\mathbf{p}$ ::: where you are standing (a point).
• $\mathbf{x}$ ::: a nearby general point you might move to.
• $\mathbf{x}-\mathbf{p}$ ::: the arrow from $\mathbf{p}$ to $\mathbf{x}$ — subtract coordinate by coordinate.

Why we need this: writing $(p_1,p_2)$ everywhere is noisy. One bold letter keeps formulas readable.

2. The limit $\underset{h\to 0}{\lim }$

Picture the fraction $\frac{f(x+h)-f(x)}{h}$ as a stick touching the hill at two points $h$ apart. As $h$ shrinks, the stick pivots and settles onto the true tangent slope at one point.

• $h$ ::: a tiny step size, heading to zero.
• $\to$ ::: "approaches."

Why we need this: slope is "rise over run," and true slope needs an infinitesimally small run. The limit is the only honest way to say "small run."

3. Partial derivatives $f_x,f_y$

Now hold one coordinate frozen so the two-variable hill becomes a simple one-variable curve.

The curly $\partial$ (say "partial dee") is a warning flag: "other variables are being held still." Contrast with the straight $d$ of ordinary one-variable calculus, where there's nothing else to hold.

• $\partial$ ::: "derivative, but other variables held constant."
• $f_x$ ::: how fast the height climbs if you walk due east only.
• $f_y$ ::: how fast the height climbs if you walk due north only.

Why the topic needs this: $f_x$ and $f_y$ are literally the two components of the gradient. Build these and the gradient is half-built. See Partial derivatives.

4. A unit direction $\hat{u}$

You can walk in any compass direction, not just east or north. We name a direction with an arrow.

Why length must be 1: a "rate of climb per step" only makes sense if every step is the same size. A length-1 arrow guarantees one unit of walking, so the rate is honest. Feed in a length-2 arrow and every rate doubles — a classic mistake the parent warns about.

• $|\mathbf{v}|$ (the bars) ::: the length of the arrow $\mathbf{v}$, from $\sqrt{v_1^2+v_2^2}$.
• Normalise ::: shrink an arrow to length 1 by dividing by its own length: $\hat{u}=\mathbf{v}/|\mathbf{v}|$.

5. The dot product $\mathbf{a}\cdot \mathbf{b}$

This is the machine that turns "gradient + direction" into "climb rate."

Geometrically, the dot product measures how much one arrow points along the other — the length of the shadow of $\mathbf{b}$ on $\mathbf{a}$, scaled by $|\mathbf{a}|$.

• Same direction ($\theta =0$): $\cos \theta =1$, dot is biggest.
• Perpendicular ($\theta =90^{\circ }$): $\cos \theta =0$, dot is zero.
• Opposite ($\theta =180^{\circ }$): $\cos \theta =-1$, dot is most negative.

Why the topic lives on this: the directional derivative is a dot product. So "which way is steepest?" becomes "which direction makes $\cos \theta =1$?" — and the answer is forced: point the same way as the gradient. This single fact powers steepest ascent, descent, and the perpendicular-to-level-curve property. See Directional derivative.

6. The angle $\theta$ and $\cos \theta$

$\theta$ (theta) is just the Greek letter we use for the angle between two arrows. Here it is the angle between your walking direction $\hat{u}$ and the gradient. $\cos \theta$ is the number, between $-1$ and $+1$, that says how aligned they are (Section 5). That's all the parent's $|\nabla f|\cos \theta$ means: gradient length, dimmed by how off-direction you are.

7. The gradient symbol $\nabla f$

Now every part exists, so we can assemble the star.

• $\nabla f$ ::: the arrow of all partial slopes at a point.
• $|\nabla f|$ ::: its length = the steepest possible climb rate.
• $\nabla f\cdot \hat{u}$ ::: your actual climb rate when walking in direction $\hat{u}$.

Why one symbol: packing $f_x,f_y$ into $\nabla f$ lets the whole directional-derivative story shrink to one clean line, $D_{\hat{u}}f=\nabla f\cdot \hat{u}$.

8. The chain rule (the glue)

The parent derives the dot-product formula using the chain rule. In one variable it says: if $z$ depends on $u$ and $u$ depends on $h$, then $\frac{dz}{dh}=\frac{dz}{du}\frac{du}{dh}$ — rates multiply along a chain. In many variables, walking in direction $\hat{u}$ changes both $x$ and $y$ at once, so the rates add up: $\frac{d}{dh}f(p_1+h{u}_1,p_2+h{u}_2)=f_x{u}_1+f_y{u}_2.$ Why we need it: it is the bridge that shows the climb rate is a dot product. See Chain rule (multivariable).

9. Level curve $f=c$

Along a contour the height never changes, so climb rate is zero, so $\nabla f\cdot \hat{u}=0$ for the along-contour direction $\hat{u}$ — meaning $\nabla f$ sits perpendicular to the contour. That is the parent's third big property. See Level curves and surfaces.

Prerequisite map

Equipment checklist

Test yourself — you are ready for the parent note if you can answer each without peeking.

Connections

• Yeh note Hinglish mein padho →
• Partial derivatives — the components $f_x,f_y$ built here.
• Directional derivative — the dot product $\nabla f\cdot \hat{u}$ assembled here.
• Chain rule (multivariable) — the glue behind the dot-product formula.
• Level curves and surfaces — where the perpendicularity picture lives.
• Tangent plane and linearisation — next step once the gradient is built.
• Gradient descent — uses $-\nabla f$.
• Lagrange multipliers — uses gradient normals.