Cauchy–Schwarz and the Triangle Inequality

Statement

In any inner product space (real or complex) with inner product ⟨·,·⟩ and induced norm ‖u‖ = √⟨u,u⟩, every pair of vectors satisfies the Cauchy–Schwarz inequality |⟨u,v⟩| ≤ ‖u‖‖v‖, with equality precisely when u and v are linearly dependent; from this the norm obeys the triangle inequality ‖u+v‖ ≤ ‖u‖ + ‖v‖.

Why it matters

Cauchy–Schwarz is the single inequality that makes an inner product deserve to be called a geometry. It guarantees that ⟨u,v⟩/(‖u‖‖v‖) lies in [−1,1], so an angle between vectors — and hence orthogonality, projection, and the whole apparatus of Hilbert-space geometry — is well defined. Without it, "length" would not satisfy the triangle inequality and the space would not be a metric space.

In physics the same inequality is the backbone of the Robertson–Schrödinger uncertainty relation, the bound |⟨ψ|φ⟩| ≤ 1 on normalized quantum overlaps (fidelities), the Bogoliubov inequality in statistical mechanics, and stability estimates for wave equations. It sits on the symmetry thread because it is exactly the statement that the positive-definite bilinear form ⟨·,·⟩ — the object left invariant by the space's isometry group — can never be "larger" than the lengths it measures.

Assumptions

Derivation

Result

Limiting cases

• u ∥ v (parallel): equality |⟨u,v⟩| = ‖u‖‖v‖; the residual ‖u − λv‖ = 0.
• u ⊥ v (orthogonal): ⟨u,v⟩ = 0, both sides trivially satisfied, and ‖u+v‖² = ‖u‖² + ‖v‖² — the Pythagorean theorem.
• v → 0: both sides → 0; the inequality degenerates to 0 ≤ 0 and the projection coefficient is undefined but never needed.
• Real 2-vectors: reduces to (a·b)² ≤ |a|²|b|², i.e. cos²θ ≤ 1.
• Rank-1 overlap of normalized states: |⟨ψ|φ⟩| = 1 iff |φ⟩ = e^iα|ψ⟩ — same physical state up to global phase.

Breaks when

Failure modes

• Dropping the modulus. Writing ⟨u,v⟩ ≤ ‖u‖‖v‖ in the complex case is meaningless — ⟨u,v⟩ is generally complex and cannot be compared with a real number. The bound is on |⟨u,v⟩|.
• Squaring sign errors. Confusing |⟨u,v⟩|² with ⟨u,v⟩². The former is ⟨u,v⟩⟨u,v⟩* (real, ≥ 0); the latter is complex and does not appear.
• Wrong conjugation slot. Applying "linear in the second argument" (physics/Dirac convention) while using the math convention, so λ* and λ land on the wrong terms and Step 6 fails to cancel.
• Assuming equality means "equal vectors". Equality means parallel (u = λv for some scalar λ), not u = v; in ℂ the phase and magnitude of λ are free.
• Using it on ℓ¹ or ℓ^∞. These are Banach but not Hilbert spaces; forcing a "⟨u,v⟩ ≤ ‖u‖‖v‖" statement there is wrong — the parallelogram law fails, so no inner product induces those norms.

Discussion

The proof is a disguised optimization. Step 4's choice λ = ⟨u,v⟩/‖v‖² is precisely the coefficient that makes u − λv orthogonal to v — the orthogonal projection of u onto the line spanned by v. The identity ‖u‖² = |⟨u,v⟩|²/‖v‖² + ‖u − λv‖² is Pythagoras applied to the decomposition of u into its component along v plus the perpendicular remainder. Cauchy–Schwarz is then nothing but the statement that a leg of a right triangle cannot exceed the hypotenuse. Seeing it this way makes the equality condition obvious: the perpendicular remainder vanishes exactly when u already lies along v.

The logical flow is one-directional and worth memorizing: positive-definiteness ⟹ Cauchy–Schwarz ⟹ triangle inequality ⟹ ‖·‖ is a norm ⟹ d(u,v) = ‖u−v‖ is a metric. Every downstream fact of Hilbert-space geometry rests on the first arrow. The triangle inequality is not an independent axiom of inner product spaces; it is a theorem, which is why "inner product space" is a stronger structure than "normed space."

In quantum mechanics the inequality is ubiquitous. The transition amplitude between normalized states satisfies |⟨ψ|φ⟩| ≤ 1, so transition probabilities |⟨ψ|φ⟩|² never exceed 1 — a consistency condition on the Born rule. Applied to the vectors A|ψ⟩ and B|ψ⟩ for Hermitian operators, Cauchy–Schwarz produces ⟨A²⟩⟨B²⟩ ≥ |⟨AB⟩|², whose imaginary part is the commutator and whose real part is the anticommutator; taking modest care with variances yields the full Robertson–Schrödinger uncertainty relation σ_A²σ_B² ≥ |½⟨{A,B}⟩ − ⟨A⟩⟨B⟩|² + |½ᵢ⟨[A,B]⟩|². Heisenberg's σ_xσ_p ≥ ℏ/2 is the special case.

The deepest content is that Cauchy–Schwarz and the parallelogram law ‖u+v‖² + ‖u−v‖² = 2‖u‖² + 2‖v‖² together characterize inner-product norms among all norms (Jordan–von Neumann): a norm arises from some inner product if and only if it satisfies the parallelogram law, in which case the inner product is recovered by polarization, ⟨u,v⟩ = ¼(‖u+v‖² − ‖u−v‖²) in the real case (with imaginary terms added in the complex case). Thus the entire rigid geometry — angles, orthogonality, projections, and Cauchy–Schwarz itself — is encoded in a single quadratic identity on lengths. On the symmetry thread this is the statement that Euclidean/Hilbert geometry is exactly the geometry whose isometry group preserves a positive-definite quadratic form; loosening the form to indefinite signature swaps the orthogonal group for the Lorentz group and flips the inequality.

Common misconceptions. (i) That completeness is needed — it is not; the inequality is purely algebraic. (ii) That the triangle inequality is an axiom — for inner product spaces it is derived. (iii) That "equality ⟹ equal vectors" — it means parallel. (iv) That the inequality holds for every bilinear form — it needs positive-definiteness, and reverses for Lorentzian signatures.

Worked examples

Problems

1. Verify Cauchy–Schwarz for u = (3, 4), v = (4, 3) in ℝ² and find the angle between them.
   Solution

   ⟨u,v⟩ = 12 + 12 = 24. ‖u‖ = √25 = 5, ‖v‖ = √25 = 5, product = 25. Then 24 ≤ 25 ✓. cos θ = 24/25 = 0.96, so θ = arccos(0.96) ≈ 16.26°. Strict because the vectors are not parallel.
2. Two normalized qubit states |ψ⟩ = |0⟩ and |φ⟩ = cos(θ/2)|0⟩ + sin(θ/2)|1⟩. Compute |⟨ψ|φ⟩| and confirm it is ≤ 1; when is it equal to 1?
   Solution

   ⟨ψ|φ⟩ = cos(θ/2), so |⟨ψ|φ⟩| = |cos(θ/2)| ≤ 1 ✓ since both states are unit vectors (‖ψ‖ = ‖φ‖ = 1). Equality holds when cos(θ/2) = ±1, i.e. θ = 0 (or 2π) — the two states coincide up to phase and are linearly dependent. The transition probability is |⟨ψ|φ⟩|² = cos²(θ/2).
3. Find the exact value of λ that minimizes ‖u − λv‖² for u = (1,1,1), v = (1,2,3) in ℝ³, and evaluate the minimum.
   Solution

   λ = ⟨u,v⟩/‖v‖² = (1+2+3)/(1+4+9) = 6/14 = 3/7. Minimum = ‖u‖² − |⟨u,v⟩|²/‖v‖² = 3 − 36/14 = 3 − 18/7 = 3/7 ≈ 0.4286. Since 3/7 > 0 the inequality is strict: |⟨u,v⟩| = 6 < ‖u‖‖v‖ = √3·√14 = √42 ≈ 6.48.
4. Use Cauchy–Schwarz on a = (a₁,…,a_n) and b = (1,…,1) to prove the RMS–AM inequality (Σaᵢ)² ≤ n Σaᵢ². State when equality holds.
   Solution

   With the dot product, ⟨a,b⟩ = Σaᵢ·1 = Σaᵢ, ‖a‖² = Σaᵢ², ‖b‖² = n. Cauchy–Schwarz |⟨a,b⟩|² ≤ ‖a‖²‖b‖² gives (Σaᵢ)² ≤ n Σaᵢ². Dividing by n² yields (mean)² ≤ mean of squares. Equality iff a ∥ b, i.e. all aᵢ equal.
5. In L²[−π,π] with ⟨f,g⟩ = ∫f*g dx, take f(x) = sin x, g(x) = cos x. Compute both sides of Cauchy–Schwarz and interpret the result.
   Solution

   ⟨f,g⟩ = ∫₋π^π sin x cos x dx = ½∫₋π^π sin 2x dx = 0 (odd, full period). ‖f‖² = ∫₋π^π sin²x dx = π, likewise ‖g‖² = π, so ‖f‖‖g‖ = π. Cauchy–Schwarz: 0 ≤ π ✓, satisfied with strict inequality. The zero overlap means sin x and cos x are orthogonal on [−π,π] — the foundational orthogonality relation underlying Fourier series. Here Cauchy–Schwarz is loose precisely because the functions are perpendicular, the opposite extreme from parallel.