Dot Product

How the dot product works

Two vectors live in space. Place them tail-to-tail at the origin. The dot product asks one question: how much of one vector lies along the other? If they point in the same direction, the answer is large and positive. If they are perpendicular, the answer is zero. If they point in opposite directions, the answer is negative.

For vectors a = (a₁, a₂, …, aₙ) and b = (b₁, b₂, …, bₙ), there are two equivalent definitions:

• Algebraic: a·b = a₁b₁ + a₂b₂ + … + aₙbₙ. Multiply matching components, add the products.
• Geometric: a·b = |a||b|cosθ, where θ is the angle between the two vectors and |a|, |b| are their lengths.

The two formulas agree. The proof comes from the law of cosines applied to the triangle formed by a, b, and a − b, which yields |a − b|² = |a|² + |b|² − 2(a·b). Both sides expand into the same component sums.

The dot product builds nearly every other geometric quantity in linear algebra. Length is √(v·v). The cosine of the angle between two vectors is (a·b) / (|a||b|). Orthogonality is just a·b = 0. The component of u along v is (u·v) / (v·v) — a quotient of two dot products.

Worked example: the angle between two vectors

Take a = (1, 2, 3) and b = (4, 5, 6). Find the angle between them.

1. Compute the dot product: a·b = 1·4 + 2·5 + 3·6 = 4 + 10 + 18 = 32.
2. Compute the lengths: |a| = √(1 + 4 + 9) = √14 ≈ 3.742; |b| = √(16 + 25 + 36) = √77 ≈ 8.775.
3. Apply the geometric formula: cosθ = 32 / (√14 · √77) = 32 / √1078 ≈ 32 / 32.833 ≈ 0.9746.
4. Take the inverse cosine: θ = arccos(0.9746) ≈ 12.93°.

The two vectors are nearly aligned — they point in almost the same direction in three-space. A cosine close to 1 confirms tight alignment; a cosine close to 0 would mean perpendicularity; a cosine close to −1 would mean opposite directions.

Algebraic properties

• Commutative. a·b = b·a, because cosθ is symmetric in its arguments and component sums don't care about order.
• Distributive. a·(b + c) = a·b + a·c. Useful for breaking complicated combinations into pieces.
• Compatible with scalar multiplication. (ka)·b = k(a·b) = a·(kb).
• Positive definite on real vectors. v·v ≥ 0, with equality only when v = 0. This is what lets √(v·v) define a length.
• Cauchy–Schwarz inequality. |a·b| ≤ |a||b|, with equality iff a and b are parallel. The cosine formula makes this obvious — |cosθ| ≤ 1.

Real dot product vs Hermitian inner product

The dot product as defined above works for vectors with real entries. Once you let entries become complex, you need a new operation that still gives a positive notion of length. The Hermitian (or complex) inner product fixes this by conjugating one factor.

	Real dot product	Hermitian inner product
Vectors	Real entries (ℝⁿ)	Complex entries (ℂⁿ)
Formula	a·b = Σ aᵢ bᵢ	⟨a,b⟩ = Σ ā ᵢ bᵢ
Symmetry	a·b = b·a (symmetric)	⟨a,b⟩ = conjugate of ⟨b,a⟩ (Hermitian)
v·v / ⟨v,v⟩	Real, ≥ 0	Real, ≥ 0 (always real)
Linearity	Linear in both arguments	Conjugate-linear in first, linear in second
Defines	Length, angle in ℝⁿ	Length, angle in ℂⁿ; quantum amplitudes
Used in	Geometry, classical physics, ML	Quantum mechanics, signal processing, FFT

Without the conjugate, the complex sum (i, 0)·(i, 0) = i² = −1, which would force a length of √(−1). The Hermitian fix makes ⟨(i,0),(i,0)⟩ = ī·i = (−i)(i) = 1, and length stays sane. Quantum mechanics depends entirely on this construction — wavefunctions live in a complex Hilbert space and ⟨ψ,ψ⟩ is a probability.

Where the dot product shows up

• Work in physics. When a constant force F moves an object through displacement d, the work done is W = F·d. Only the component of force along the motion contributes; perpendicular force does no work. Carry a heavy backpack horizontally and gravity does zero work, even though it acts on you the whole time.
• Projection. The component of u along v is (u·v / v·v) v. Least-squares regression, image compression via PCA, and Gram–Schmidt orthonormalization all build on projection.
• Lighting in 3D graphics. The brightness of a surface lit by a directional light is N·L, the dot product of the surface normal and the light direction. Negative or zero means the surface faces away. Almost every shader written since 1980 uses this single line.
• Cosine similarity. Search engines, recommendation systems, and word embeddings compare documents or items by computing the cosine of the angle between their feature vectors. Two vectors with cosine 0.9 are deemed nearly the same regardless of magnitude.
• Neural networks. Every neuron computes w·x + b, a dot product followed by a bias. A modern transformer evaluates billions of dot products per forward pass.
• Fourier coefficients. Decomposing a function f into sines and cosines is just dot products of f with each basis function in an inner-product space.

Common mistakes

• Confusing dot and cross. The dot product returns a scalar; the cross product returns a vector. Different operation, different result type. Never write "a·b" when you mean a vector.
• Assuming a non-zero result means alignment. Sign matters. A negative dot product means the vectors point more than 90° apart — they oppose, not align.
• Forgetting the conjugate in complex spaces. The real formula computes the wrong thing on complex vectors. Quantum and DSP code that drops the conjugate gives nonsense.
• Treating cosine similarity as a distance. It measures angular alignment, not Euclidean distance. Two unit vectors with cosine 0.99 are angularly close but their physical separation depends on length.
• Using the algebraic sum without normalizing. The raw dot product mixes magnitude and angle. When you only care about direction, divide out the lengths first.