The Dot Product Is a Length-Weighted Shadow

The coordinate formula for the dot product looks like bookkeeping:

$$x\cdot y=x_1{y}_1+\cdots +x_d{y}_d.$$

Geometrically, it answers a more focused question: how much of one vector points in the direction of the other? The answer is a signed projection, scaled by the length of the vector providing the direction.

The algebraic score

The similarity interpretation is already present in the coordinate formula. Regard $x$ as a row of weights applied to the coordinates of $y$:

$$x\cdot y=\left[\begin{array}{ccc}{x}_1& \dots & x_d\end{array}\right]\left[\begin{array}{c}{y}_1\\ ⋮\\ y_d\end{array}\right]=\sum _i{x}_i{y}_i.$$

Each coordinate contributes according to two things: how large the two components are, and whether their signs agree. For a single coordinate,

$$x_i{y}_i=|x_i||y_i|\mathrm{sign}(x_i{y}_i),$$

so the whole score can be read as

$$x\cdot y=\sum _i\underset{\text{size of the shared component}}{\underbrace{|x_i||y_i|}}\underset{\text{agreement or opposition}}{\underbrace{\mathrm{sign}(x_i{y}_i)}}.$$

When $x_i$ and $y_i$ have the same sign, that coordinate adds to the score. When they have opposite signs, it subtracts from the score. A coordinate contributes little when either component is small. The dot product is therefore an algebraic agreement score: it rewards large components occurring in the same signed directions and penalises components pointing against one another.

For example, the three vectors below all have length $5$:

$$\begin{array}{rlrl}x& =(3,4),& x\cdot x& =25,\\ y& =(4,-3),& x\cdot y& =12-12=0,\\ z& =(-3,-4),& x\cdot z& =-9-16=-25.\end{array}$$

The first comparison agrees in both coordinates, the second has equal positive and negative contributions that cancel, and the third disagrees in both coordinates. The numbers $25$, $0$, and $-25$ are not arbitrary labels: they record reinforcement, cancellation, and opposition in the coordinate-wise products.

The coordinate picture depends on choosing perpendicular axes, but the score itself is the same geometric quantity in every orthonormal coordinate system. The projection formula below explains why this algebraic agreement is directional similarity rather than merely a pattern of matching signs.

The shadow on a direction

Take two non-zero vectors $x$ and $y$, and let $\theta$ be the angle between them. Turn $x$ into a unit vector:

$$e_x=\frac{x}{\Vert x\Vert }.$$

The scalar $e_x\cdot y$ is the signed length of $y$‘s shadow on the line pointing along $x$. A shadow in the same direction is positive; one in the opposite direction is negative; a perpendicular vector has no shadow at all.

The notation matters here. ${\mathrm{proj}}_x(y)$ means the projection of $y$ onto the line determined by $x$. The target is the infinite line

$$\mathrm{span}(x)=\{tx:t\in \mathbb{R}\},$$

not merely the finite arrow from the origin to the endpoint of $x$. The projection is a vector parallel to $x$. When we draw it from the origin, its endpoint is the foot of the perpendicular dropped from the endpoint of $y$ to that line. It can lie beyond the endpoint of $x$, on the opposite ray, or at the origin.

The shadow has length

$$e_x\cdot y=\Vert y\Vert \cos \theta .$$

The ordinary dot product uses $x$ rather than the unit direction $e_x$. It therefore scales this shadow by $\Vert x\Vert$:

$$\begin{array}{rl}x\cdot y& =(\Vert x\Vert e_x)\cdot y\\ & =\Vert x\Vert (e_x\cdot y)\\ & =\Vert x\Vert \Vert y\Vert \cos \theta .\end{array}$$

So the dot product is not just an angle measurement. It is a length-weighted signed shadow.

Three directions, three signs

The angle controls the sign and the amount of shared direction:

• $0\le \theta <\frac{\pi }{2}$: the vectors point partly in the same direction, so $x\cdot y>0$;
• $\theta =\frac{\pi }{2}$: the shadow disappears, so $x\cdot y=0$;
• $\frac{\pi }{2}<\theta \le \pi$: the shadow points backwards, so $x\cdot y<0$.

For non-zero vectors, a zero dot product therefore means orthogonality, not that the vectors are unrelated. It says that neither vector has any component in the other’s direction.

The projection itself is the shadow turned back into a vector:

$${\mathrm{proj}}_x(y)=\frac{x\cdot y}{x\cdot x}x.$$

The scalar $\frac{x\cdot y}{x\cdot x}$ tells us how far to travel along $x$. The remaining vector

$$y-{\mathrm{proj}}_x(y)$$

is perpendicular to $x$. The dot product is the number that separates these two pieces.

A small catalogue of cases

The target line and the direction of the result become clearer in concrete cases. “Same direction” means $y=\lambda x$ for some $\lambda >0$; “opposite directions” means $y=-\lambda x$ for some $\lambda >0$.

Identical vectors

Let $y=x$. Since $y$ is already on the target line,

$${\mathrm{proj}}_x(y)=x=y,x\cdot y=x\cdot x=\Vert x{\Vert }^2.$$

The projection is the whole vector, and the dot product is its squared length.

Same direction, different length

Let $x=(1,0)$ and $y=(3,0)$. The projection is $y$ itself because $y$ lies on the line through $x$:

$${\mathrm{proj}}_x(y)=(3,0),x\cdot y=3.$$

The endpoint of $y$ lies beyond the endpoint of $x$; the target is an infinite line, not the finite arrow $[0,x]$.

Opposite directions

Let $x=(1,0)$ and $y=(-3,0)$. The projection is still $y$ itself, but it lies on the opposite ray:

$${\mathrm{proj}}_x(y)=(-3,0),x\cdot y=-3.$$

The negative sign records that the shadow points backwards along $x$.

Perpendicular vectors

Let $x=(1,0)$ and $y=(0,2)$. The perpendicular from $y$ reaches the origin, so

$${\mathrm{proj}}_x(y)=(0,0),x\cdot y=0.$$

The zero vector

If $y=(0,0)$, then its projection and dot product are also zero. But the zero vector has no direction, so its angle with $x$ is undefined. The statement “zero dot product means perpendicular” is only valid for non-zero vectors.

An oblique pair

Let $x=(2,0)$ and $y=(1,2)$. Then

$${\mathrm{proj}}_x(y)=(1,0),y-{\mathrm{proj}}_x(y)=(0,2),$$

so $y$ splits into a part along $x$ and a perpendicular part. The dot product is $x\cdot y=2$.

The subscript also tells us which vector supplies the target line. In the oblique example,

$${\mathrm{proj}}_x(y)=(1,0),{\mathrm{proj}}_y(x)=\frac{2}{5}(1,2)=\left(\frac{2}{5},\frac{4}{5}\right).$$

These projections are different because they land on different lines, even though the dot product itself is symmetric: $x\cdot y=y\cdot x$.

Why the coordinate sum works

Choose perpendicular unit directions $e_1,\dots ,e_d$. Write

$$x=\sum _i{x}_i{e}_i,y=\sum _i{y}_i{e}_i.$$

Expanding the dot product gives

$$x\cdot y=\sum _{i,j}{x}_i{y}_j(e_i\cdot e_j).$$

Perpendicular unit directions satisfy

$$e_i\cdot e_j=\{\begin{array}{ll}1& i=j,\\ 0& i\ne j.\end{array}$$

Every cross-direction term vanishes. Only matching components remain:

$$x\cdot y=\sum _i{x}_i{y}_i.$$

The familiar coordinate formula is therefore a geometric rule in disguise. It pairs the parts of $x$ and $y$ that lie along the same perpendicular axes, then adds their contributions. Components pointing along different axes do not contribute to the shared direction.

The length hidden in similarity

Two vectors can have a large dot product because they point similarly, because they are long, or both. If only direction matters, remove the length factors:

$$\cos \theta =\frac{x\cdot y}{\Vert x\Vert \Vert y\Vert }.$$

This is the basis of cosine similarity. It compares shadows after both vectors have been scaled to length one. The ordinary dot product keeps the same geometric comparison but lets longer vectors count more heavily.

The dot product is therefore best pictured as a shadow cast onto a direction: signed by whether the vectors agree, measured by how much of one lies along the other, and enlarged by the length of the measuring vector.