Picture a canoe crossing a river. The current doesn't care where you're trying to go — it just pushes, at some angle, with some strength. Part of that push carries you straight downstream; the rest shoves you uselessly sideways, towards one bank or the other. If you only care about how much the current speeds up (or slows down) your trip downstream, you don't need the whole push — you need exactly the sliver of it that points downstream, with the sideways part thrown away.

That sliver has a name: the projection of one vector onto another. It answers a question that comes up everywhere forces and directions meet: "how much of this vector actually acts along that direction?"

The shadow picture

Shine a light straight down onto a line and a vector casts a shadow on it — as if the sun's rays fall perpendicular to the line and every point of the vector's tip drops straight down onto it. That shadow is the projection of \vec{v} onto \vec{u}: the part of \vec{v} that lies along \vec{u}'s direction, with the perpendicular part discarded entirely.

The length of the shadow (with a sign — negative if it falls backwards) is \vec{v}\cdot\hat{u}, the dot product with the unit vector in \vec{u}'s direction. Written without needing a unit vector first, the full projection vector is built straight from the dot product:

Read it in two pieces. The fraction \dfrac{\vec{u}\cdot\vec{v}}{\vec{u}\cdot\vec{u}} is just a number — how many copies of \vec{u} you need to reach the shadow's tip. Multiplying that number back onto \vec{u} turns it into a vector that actually points along \vec{u}'s direction.

Worked example 1 — a plain numeric projection

Project \vec{v} = \begin{bmatrix} 6 \\ 8 \end{bmatrix} onto \vec{u} = \begin{bmatrix} 3 \\ 4 \end{bmatrix}. First the dot products:

So the scaling number is \tfrac{50}{25} = 2, and the projection is

The shadow lands exactly on \vec{v} itself — which makes sense, because \vec{v} is already pointing exactly along \vec{u}'s direction (it's just twice as long). When two vectors already point the same way, projecting one onto the other changes nothing about its direction.

Try a case where the vectors don't line up: project \vec{v} = \begin{bmatrix} 1 \\ 7 \end{bmatrix} onto the same \vec{u} = \begin{bmatrix} 3 \\ 4 \end{bmatrix}. This time \vec{u}\cdot\vec{v} = 3+28=31, so the projection is \tfrac{31}{25}\begin{bmatrix} 3\\4 \end{bmatrix} \approx \begin{bmatrix} 3.72 \\ 4.96 \end{bmatrix}. Subtract that from \vec{v} and what's left, \begin{bmatrix} -2.72 \\ 2.04 \end{bmatrix}, is the perpendicular leftover — and if you dot it with \vec{u} you'll get (up to rounding) exactly zero. Every vector splits, without exception, into "the shadow along \vec{u}" plus "the bit perpendicular to \vec{u}," and the two pieces always add back up to the original vector.

Worked example 2 — the river current, with real numbers

Back to the canoe. Suppose the current pushes with vector \vec{c} = \begin{bmatrix} 3 \\ 4 \end{bmatrix} km/h (partly downstream, partly sideways), and "downstream" is the direction \vec{d} = \begin{bmatrix} 1 \\ 0 \end{bmatrix}. How fast does the current actually carry you downstream?

Exactly 3 km/h downstream — the sideways 4 km/h is real (it will drift you toward a riverbank), but it contributes nothing to your downstream progress. Projection is precisely the tool that separates "the part that matters for this direction" from "everything else."

Casting the shadow

Drag \vec{v}. The bold segment on the line through \vec{u} is the projection, and the dashed line dropping from \vec{v}'s tip meets that line at a right angle — projection always drops perpendicularly, exactly like the sunlight picture. When \vec{v} is itself perpendicular to \vec{u}, the shadow shrinks all the way to nothing — there's no "along \vec{u}" component left at all.

Worked example 3 — a force on a ramp

Here's the classic physics use of this idea, well before it's ever called "projection" in a physics class. A crate sits on a frictionless ramp, and gravity pulls straight down with force \vec{g} = \begin{bmatrix} 0 \\ -10 \end{bmatrix} (in some convenient units). The ramp's surface points along \vec{r} = \begin{bmatrix} 4 \\ -3 \end{bmatrix}. How much of gravity's pull actually acts along the ramp — the part that slides the crate down its surface?

Whatever is left over — \vec{g} minus that projection — presses the crate into the ramp instead of sliding it. Splitting a force this way, into "along the surface" and "into the surface," is exactly what a physics teacher means by resolving forces on an inclined plane — usually taught with triangles and trigonometry, but it's this same dot-product machinery underneath the whole time.

The workhorse of applied maths

Projection is how you find the "closest point" on a line or plane to a given target — which is exactly the question behind least-squares fitting, the engine of linear regression. It splits any vector into "the part along \vec{u}" plus "the leftover perpendicular part," and that split — signal plus residual — turns up everywhere from graphics to statistics. It closes out Stage A: with addition, scaling, length, the dot product and projection, you have the full vocabulary of vectors.

There's an even bigger payoff waiting just past this page. Chain projections together — project onto one direction, then project what's left onto a second orthogonal direction, then a third — and you can rebuild any vector as a sum of independent, non-overlapping pieces. That process is called Gram-Schmidt, and it's how a computer takes a messy set of directions and straightens them into a clean orthogonal basis, ready for the "no leakage between directions" convenience that comes with orthogonal axes. Every one of those straightening steps is nothing more than the shadow formula above, applied once per direction.