The Cauchy-Schwarz Inequality

Suppose you want to measure how "aligned" two vectors are. You want to do this without any reference to angles drawn in a picture. Given only an abstract Inner Product Space and the number ⟨u,v⟩\langle u, v\rangle, is there a bound? Specifically, one relating this number to the sizes of uu and vv alone. The Cauchy–Schwarz inequality answers this decisively. The magnitude of the inner product of two vectors never exceeds the product of their norms. This single fact is the hinge on which the entire theory of inner product spaces turns. The inequality matters because it is the technical engine behind almost everything that follows. It lets us define the angle between two abstract vectors. This includes even vectors that are functions, sequences, or random variables. It does so by guaranteeing that a certain ratio always lies in [−1,1][-1,1]. It is also the single ingredient needed to prove the Triangle Inequality and Norm Properties. This applies for the norm induced by an inner product. In turn, this is what makes that norm a genuine norm, rather than just a convenient formula. The big picture is this: an inner product space carries more structure than a mere vector space with a norm. It carries a notion of "how much two vectors point the same way." Cauchy–Schwarz converts this qualitative idea into a hard numerical inequality. It is provable from the inner product axioms alone. There is no appeal to geometry, coordinates, or dimension. It works identically in Rn\mathbb{R}^n, in Cn\mathbb{C}^n, in spaces of polynomials. It also works in infinite-dimensional spaces of square-integrable functions. This article gives the precise statement. It includes a complete proof using orthogonal projection onto a single vector. It also gives a full analysis of when equality holds. It gives the geometric reading of the inequality as a bound on the cosine of an angle. It closes with worked examples in R3\mathbb{R}^3 and in a function space. It also gives a counterexample showing why the inner product structure is essential. Finally, it looks at where the inequality resurfaces throughout mathematics, physics, and statistics.

Key takeaways
  • For vectors u,vu,v in an inner product space, ∣⟨u,v⟩∣≤∥u∥ ∥v∥|\langle u,v\rangle| \le \|u\|\,\|v\|, always.
  • The proof projects vv onto the line spanned by uu and expands the nonnegative norm of the leftover piece.
  • Equality holds exactly when uu and vv are linearly dependent (one is a scalar multiple of the other).
  • It works verbatim over R\mathbb{R} or C\mathbb{C}, in finite or infinite dimensions.
  • It is exactly what is needed to define cos⁡θ=⟨u,v⟩∥u∥∥v∥\cos\theta = \dfrac{\langle u,v\rangle}{\|u\|\|v\|} and to prove the Triangle Inequality and Norm Properties.
  • Key idea: orthogonal projection onto a single vector turns "angle bound" into a one-line nonnegativity argument.

Where this lives today

The Cauchy–Schwarz inequality is not a curiosity confined to geometry — it is structural machinery used across mathematics and its applications.

In probability and statistics, applied to the inner product ⟨X,Y⟩=E[XY]\langle X,Y\rangle = \mathbb{E}[XY] on random variables with finite variance, it yields ∣Cov(X,Y)∣≤σXσY|\mathrm{Cov}(X,Y)| \le \sigma_X \sigma_Y, the fact underlying the bound ∣ρ∣≤1|\rho| \le 1 for the Pearson correlation coefficient ρ\rho.

In quantum mechanics, the inner product is the Hermitian inner product on a Hilbert space of quantum states, and Cauchy–Schwarz underlies the derivation of the Heisenberg uncertainty principle, bounding the product of variances of non-commuting observables.

In signal processing and machine learning, it justifies bounds on inner products of feature vectors (used constantly in kernel methods and in bounding cosine similarity scores, always guaranteed to lie in [−1,1][-1,1]).

In numerical analysis and optimization, it gives error bounds for approximation methods, including bounds used to control the convergence rate of conjugate-gradient-type iterative solvers.

In pure mathematics, it is the direct prerequisite for the Triangle Inequality and Norm Properties (which needs exactly this bound to control cross terms when expanding ∥u+v∥2\|u+v\|^2), and it motivates the study of Orthogonality and Orthogonal Sets, since the equality case identifies precisely when two vectors fail to have any "independent direction" between them.

Ask the tutor: Can you show me how Cauchy–Schwarz is used, step by step, to derive the triangle inequality for the induced norm?

DefinitionCauchy–Schwarz Inequality

Let VV be a vector space over F∈{R,C}\mathbb{F} \in \{\mathbb{R}, \mathbb{C}\} equipped with an inner product ⟨⋅,⋅⟩\langle \cdot,\cdot\rangle, and let

∥v∥:=⟨v,v⟩\|v\| := \sqrt{\langle v, v\rangle}

be the induced norm. The Cauchy–Schwarz inequality states: for all u,v∈Vu, v \in V,

∣⟨u,v⟩∣≤∥u∥ ∥v∥,|\langle u, v \rangle| \le \|u\| \, \|v\|,

with equality if and only if uu and vv are linearly dependent, i.e., one of them is a scalar multiple of the other.

Intuition

Picture two vectors uu and vv in the ordinary plane. The dot product u⋅v=∥u∥∥v∥cos⁡θu\cdot v = \|u\|\|v\|\cos\theta, where θ\theta is the angle between them, and since ∣cos⁡θ∣≤1|\cos\theta|\le 1, we immediately get ∣u⋅v∣≤∥u∥∥v∥|u\cdot v|\le \|u\|\|v\|. Cauchy–Schwarz says this bound survives even when there is no picture to draw — even in spaces of functions, matrices, or infinite sequences, where "angle" is not a primitive notion but something we are about to define using this very inequality. The proof strategy makes this precise without circularity. Drop a perpendicular from vv onto the line through uu: this gives a decomposition

v=⟨v,u⟩∥u∥2u⏟projection onto u+w⏟orthogonal remainder,v = \underbrace{\frac{\langle v,u\rangle}{\|u\|^2}u}_{\text{projection onto }u} + \underbrace{w}_{\text{orthogonal remainder}},

where w⊥uw \perp u. The norm of vv is then the hypotenuse of a right triangle whose legs are the projection and the remainder — a Pythagorean relation — and since ∥w∥2≥0\|w\|^2 \ge 0, the projection's length can be no larger than ∥v∥\|v\| itself. Unwinding this single nonnegativity statement produces exactly the Cauchy–Schwarz inequality.

Figure
Geometric picture: projecting vv onto the line spanned by uu, with orthogonal remainder ww
u v w (remainder, ⟂ u) proj of v onto u θ

Central theorem

TheoremCauchy–Schwarz Inequality

Let VV be a vector space over F∈{R,C}\mathbb{F} \in \{\mathbb{R},\mathbb{C}\} with inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle (conjugate-linear in the second slot, positive definite, conjugate-symmetric), and induced norm ∥v∥=⟨v,v⟩\|v\| = \sqrt{\langle v,v\rangle}. Then for all u,v∈Vu,v \in V:

∣⟨u,v⟩∣≤∥u∥ ∥v∥.|\langle u,v\rangle| \le \|u\|\,\|v\|.

Moreover, equality holds if and only if uu and vv are linearly dependent.

Proof

Case 1: u=0u = 0. Then ⟨u,v⟩=⟨0,v⟩=0\langle u,v\rangle = \langle 0,v\rangle = 0 by linearity in the first slot, and ∥u∥=0\|u\| = 0, so both sides of the inequality equal 00 and equality holds. Also u=0u=0 is (trivially) linearly dependent with vv (since u=0⋅vu = 0\cdot v), consistent with the equality-condition claim. So assume from now on that u≠0u \neq 0.

Step 1 (Orthogonal decomposition of vv along uu). Define the scalar

c:=⟨v,u⟩⟨u,u⟩=⟨v,u⟩∥u∥2,c := \frac{\langle v, u\rangle}{\langle u,u\rangle} = \frac{\langle v,u\rangle}{\|u\|^2},

which is well-defined since ∥u∥2=⟨u,u⟩>0\|u\|^2 = \langle u,u\rangle > 0 by positive-definiteness (as u≠0u \ne 0). Set

w:=v−cu.w := v - c u.

Then v=cu+wv = cu + w. We claim w⊥uw \perp u, i.e., ⟨w,u⟩=0\langle w, u\rangle = 0. Compute, using linearity of the inner product in its first argument:

⟨w,u⟩=⟨v−cu,u⟩=⟨v,u⟩−c⟨u,u⟩=⟨v,u⟩−⟨v,u⟩⟨u,u⟩⟨u,u⟩=0.\langle w, u \rangle = \langle v - cu, u\rangle = \langle v,u\rangle - c\langle u,u\rangle = \langle v,u\rangle - \frac{\langle v,u\rangle}{\langle u,u\rangle}\langle u,u\rangle = 0.

Step 2 (Pythagorean expansion). Since v=cu+wv = cu + w with ⟨w,u⟩=0\langle w,u\rangle = 0 (hence also ⟨u,w⟩=⟨w,u⟩‾=0\langle u,w\rangle = \overline{\langle w,u\rangle} = 0 by conjugate symmetry), expand:

∥v∥2=⟨cu+w, cu+w⟩=∣c∣2⟨u,u⟩+c⟨u,w⟩+cˉ⟨w,u⟩+⟨w,w⟩.\|v\|^2 = \langle cu + w,\, cu+w\rangle = |c|^2\langle u,u\rangle + c\langle u,w\rangle + \bar c\langle w,u\rangle + \langle w,w\rangle.

The two cross terms vanish because ⟨u,w⟩=⟨w,u⟩=0\langle u,w\rangle = \langle w,u\rangle = 0. Hence

∥v∥2=∣c∣2∥u∥2+∥w∥2.\|v\|^2 = |c|^2\|u\|^2 + \|w\|^2.

Step 3 (Discard the nonnegative remainder). By positive-definiteness, ∥w∥2≥0\|w\|^2 \ge 0. Therefore

∥v∥2≥∣c∣2∥u∥2.\|v\|^2 \ge |c|^2 \|u\|^2.

Substitute c=⟨v,u⟩∥u∥2c = \dfrac{\langle v,u\rangle}{\|u\|^2}:

∥v∥2  ≥  ∣⟨v,u⟩∣2∥u∥4 ∥u∥2=∣⟨v,u⟩∣2∥u∥2.\|v\|^2 \;\ge\; \frac{|\langle v,u\rangle|^2}{\|u\|^4}\,\|u\|^2 = \frac{|\langle v,u\rangle|^2}{\|u\|^2}.

Multiply both sides by ∥u∥2>0\|u\|^2 > 0:

∥u∥2∥v∥2≥∣⟨v,u⟩∣2.\|u\|^2\|v\|^2 \ge |\langle v,u\rangle|^2.

Since ∣⟨v,u⟩∣=∣⟨u,v⟩‾∣=∣⟨u,v⟩∣|\langle v,u\rangle| = |\overline{\langle u,v\rangle}| = |\langle u,v\rangle| (conjugate symmetry and the fact that a complex number and its conjugate have equal modulus), this is

∥u∥2∥v∥2≥∣⟨u,v⟩∣2.\|u\|^2\|v\|^2 \ge |\langle u,v\rangle|^2.

Taking nonnegative square roots of both sides (both sides are nonnegative, and ⋅\sqrt{\cdot} is monotonic on [0,∞)[0,\infty)) gives

∥u∥ ∥v∥≥∣⟨u,v⟩∣,\|u\|\,\|v\| \ge |\langle u,v\rangle|,

which is the desired inequality.

Step 4 (Equality conditions). (⇐\Leftarrow) Suppose u,vu,v linearly dependent. If u=0u = 0, equality was shown above. Otherwise u≠0u \ne 0 and dependence forces v=λuv = \lambda u for some scalar λ∈F\lambda \in \mathbb{F} (if instead u=μvu = \mu v with v≠0v \ne 0 this is the symmetric case; if v=0v = 0 both sides are 00). Then

⟨u,v⟩=⟨u,λu⟩=λˉ⟨u,u⟩=λˉ∥u∥2,\langle u,v\rangle = \langle u,\lambda u\rangle = \bar\lambda \langle u,u\rangle = \bar\lambda \|u\|^2,

so ∣⟨u,v⟩∣=∣λ∣ ∥u∥2|\langle u,v\rangle| = |\lambda|\,\|u\|^2. Also ∥v∥=∥λu∥=∣λ∣ ∥u∥\|v\| = \|\lambda u\| = |\lambda|\,\|u\|, so

∥u∥∥v∥=∣λ∣ ∥u∥2=∣⟨u,v⟩∣.\|u\|\|v\| = |\lambda|\,\|u\|^2 = |\langle u,v\rangle|.

Equality holds. (⇒\Rightarrow) Suppose equality holds and u≠0u \ne 0 (the case u=0u=0 is already linearly dependent with anything). Retracing Steps 2–3, equality ∥u∥∥v∥=∣⟨u,v⟩∣\|u\|\|v\| = |\langle u,v\rangle| forces equality at every non-strict inequality used, in particular

∥v∥2=∣c∣2∥u∥2i.e.∥w∥2=0.\|v\|^2 = |c|^2\|u\|^2 \quad\text{i.e.}\quad \|w\|^2 = 0.

By positive-definiteness of the inner product, ∥w∥2=0  ⟹  w=0\|w\|^2 = 0 \implies w = 0. Hence v=cu+w=cuv = cu + w = cu, so vv is a scalar multiple of uu: u,vu,v are linearly dependent. ■\blacksquare

∎

Worked examples

Example 1 — asymmetric real vectors in R3\mathbb{R}^3 with the standard dot product. Let u=(2,−1,3)u = (2,-1,3) and v=(1,4,−2)v = (1,4,-2) in R3\mathbb{R}^3 with ⟨x,y⟩=x1y1+x2y2+x3y3\langle x,y\rangle = x_1y_1+x_2y_2+x_3y_3. Compute the inner product:

⟨u,v⟩=(2)(1)+(−1)(4)+(3)(−2)=2−4−6=−8.\langle u,v\rangle = (2)(1) + (-1)(4) + (3)(-2) = 2 - 4 - 6 = -8.

So ∣⟨u,v⟩∣=8|\langle u,v\rangle| = 8. Compute the norms:

∥u∥2=4+1+9=14,∥u∥=14,\|u\|^2 = 4+1+9 = 14, \qquad \|u\| = \sqrt{14},
∥v∥2=1+16+4=21,∥v∥=21.\|v\|^2 = 1+16+4 = 21, \qquad \|v\| = \sqrt{21}.

Then

∥u∥∥v∥=14⋅21=294≈17.146.\|u\|\|v\| = \sqrt{14}\cdot\sqrt{21} = \sqrt{294} \approx 17.146.

Indeed 8≤2948 \le \sqrt{294}, confirming the inequality — and strictly, since u,vu,v are visibly not parallel (no scalar λ\lambda satisfies 2=λ⋅12 = \lambda\cdot 1, −1=λ⋅4-1=\lambda\cdot 4, 3=λ⋅(−2)3 = \lambda\cdot(-2) simultaneously: the first gives λ=2\lambda = 2, the second gives λ=−1/4\lambda = -1/4, contradiction). We can also extract the angle between uu and vv:

cos⁡θ=⟨u,v⟩∥u∥∥v∥=−8294≈−0.4666,\cos\theta = \frac{\langle u,v\rangle}{\|u\|\|v\|} = \frac{-8}{\sqrt{294}} \approx -0.4666,

so θ≈117.8∘\theta \approx 117.8^\circ, a well-defined obtuse angle — legitimate precisely because Cauchy–Schwarz guarantees cos⁡θ∈[−1,1]\cos\theta \in [-1,1].

Example 2 — polynomials with an L2L^2 inner product on [0,1][0,1]. Let V=P2(R)V = \mathcal{P}_2(\mathbb{R}), polynomials of degree ≤2\le 2, with inner product

⟨f,g⟩=∫01f(x)g(x) dx.\langle f,g\rangle = \int_0^1 f(x)g(x)\,dx.

Take f(x)=1+xf(x) = 1 + x and g(x)=x2g(x) = x^2 (asymmetric, unrelated shapes). Compute the inner product:

⟨f,g⟩=∫01(1+x)x2 dx=∫01(x2+x3) dx=13+14=712.\langle f,g\rangle = \int_0^1 (1+x)x^2\,dx = \int_0^1 (x^2+x^3)\,dx = \frac{1}{3}+\frac{1}{4} = \frac{7}{12}.

Compute ∥f∥2\|f\|^2:

∥f∥2=∫01(1+x)2 dx=∫01(1+2x+x2) dx=1+1+13=73.\|f\|^2 = \int_0^1 (1+x)^2\,dx = \int_0^1 (1+2x+x^2)\,dx = 1 + 1 + \frac13 = \frac{7}{3}.

Compute ∥g∥2\|g\|^2:

∥g∥2=∫01x4 dx=15.\|g\|^2 = \int_0^1 x^4\,dx = \frac15.

So

∥f∥∥g∥=73⋅15=715≈0.4667≈0.6831.\|f\|\|g\| = \sqrt{\frac73}\cdot\sqrt{\frac15} = \sqrt{\frac{7}{15}} \approx \sqrt{0.4667} \approx 0.6831.

Compare to ⟨f,g⟩=7/12≈0.5833\langle f,g\rangle = 7/12 \approx 0.5833. Indeed 0.5833≤0.68310.5833 \le 0.6831, and the inequality is strict because ff and gg are not proportional (one is degree 1, the other degree 2, so no scalar multiple of one equals the other). This numerically confirms Cauchy–Schwarz in a genuinely infinite-dimensional-flavored space (a function space), not just in Rn\mathbb{R}^n.

CounterexampleWhy an inner product — not just any bilinear/sesquilinear form — is needed

The proof relies critically on positive-definiteness: ⟨w,w⟩≥0\langle w,w\rangle \ge 0 for all ww, with equality only at w=0w=0. Drop this and the inequality can fail. Let V=R2V = \mathbb{R}^2 and define the symmetric bilinear form

B(x,y)=x1y1−x2y2,B(x,y) = x_1y_1 - x_2y_2,

which satisfies bilinearity and symmetry but is not positive-definite (it is an indefinite form, the same type that appears in special-relativistic Minkowski geometry). Define ∥x∥B:=∣B(x,x)∣\|x\|_B := \sqrt{|B(x,x)|} purely as a formal analogue of the norm (this is not an honest norm, precisely because BB is not positive-definite). Take u=(1,1)u = (1,1) and v=(1,−1)v = (1,-1). Then

B(u,v)=(1)(1)−(1)(−1)=1+1=2,B(u,v) = (1)(1) - (1)(-1) = 1+1 = 2,

while

B(u,u)=1−1=0,B(v,v)=1−1=0,B(u,u) = 1 - 1 = 0, \qquad B(v,v) = 1-1=0,

so the naive right-hand side is ∣B(u,u)∣⋅∣B(v,v)∣=0\sqrt{|B(u,u)|}\cdot\sqrt{|B(v,v)|} = 0. The claimed inequality ∣B(u,v)∣≤∥u∥B∥v∥B|B(u,v)| \le \|u\|_B\|v\|_B would assert 2≤02 \le 0, which is false. The failure traces directly to Step 3 of the proof: the decomposition v=cu+wv = cu+w still makes sense algebraically, but B(w,w)B(w,w) can be negative, so the step "∥w∥2≥0\|w\|^2 \ge 0" collapses, and with it the entire chain of inequalities. This shows positive-definiteness is not a decorative hypothesis — it is exactly the load-bearing wall of the proof.

Scope, edge cases & common traps

SituationWhat happens
u=0u = 0 or v=0v = 0Both sides are 00; equality holds trivially, consistent with linear dependence (zero vector is dependent with anything)
F=C\mathbb{F} = \mathbb{C}Inequality is on ∣⟨u,v⟩∣|\langle u,v\rangle| (modulus), since ⟨u,v⟩\langle u,v\rangle itself may be non-real
Infinite-dimensional VVProof is unchanged; only uses algebra of the inner product, not dimension
Non-positive-definite formsInequality can fail entirely (see counterexample); positive-definiteness is essential
Real inner product, u,v≠0u,v \ne 0Equality iff v=λuv = \lambda u for some real λ\lambda (can further split into λ>0\lambda>0: same direction, or λ<0\lambda<0: opposite direction)

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789–1857 · French mathematician, Ecole Polytechnique and Sorbonne

Cauchy was one of the most prolific mathematicians in history, producing foundational work in real and complex analysis, the theory of permutation groups, elasticity theory, and the rigorous definition of limits and continuity that underlies modern calculus. He first proved a discrete version of the inequality that bears his name for finite sums of real numbers in 1821, in his textbook Cours d'Analyse, as a lemma supporting his study of series and inequalities. The integral form of the inequality was established later by Viktor Bunyakovsky (1859) and independently, in a broader functional-analytic setting, by Hermann Schwarz (1888); in many countries the result is accordingly called the Cauchy–Bunyakovsky–Schwarz inequality.

History
- 1821 — Cauchy proves the discrete inequality (∑aibi)2≤(∑ai2)(∑bi2)\left(\sum a_ib_i\right)^2 \le \left(\sum a_i^2\right)\left(\sum b_i^2\right) for real numbers in Cours d'Analyse.
- 1859 — Bunyakovsky extends the inequality to integrals, establishing it for continuous functions.
- 1885–1888 — Schwarz independently rediscovers and applies the integral form while studying minimal surfaces and harmonic functions.
- Early 1900s — Hilbert and others formalize abstract inner product spaces, recognizing the inequality as a purely axiomatic consequence of positive-definiteness, valid in any dimension.
- 20th century — The inequality becomes a standard tool across functional analysis, probability, and quantum mechanics, cited in its modern abstract form in every treatment of inner product spaces.
Does Cauchy–Schwarz require VV to be finite-dimensional?

No. The proof given uses only the algebraic axioms of an inner product — linearity, conjugate symmetry, and positive-definiteness — and never invokes a basis, dimension count, or coordinate representation. It therefore holds verbatim in infinite-dimensional spaces such as L2([0,1])L^2([0,1]), the space of square-summable sequences ℓ2\ell^2, or spaces of continuous functions equipped with an L2L^2-type inner product. This dimension-independence is part of why the inequality is so widely applicable in analysis, where infinite-dimensional function spaces are the norm rather than the exception.

Why is the proof done by projecting onto a single vector rather than using an orthonormal basis?

Projecting onto a single vector uu is the most economical route: it needs no basis, no Gram–Schmidt process, and no assumption about the dimension or even finiteness of VV. Using an orthonormal basis to prove Cauchy–Schwarz would work in finite dimensions but is circular in spirit, since orthonormal expansions are usually developed using Cauchy–Schwarz-adjacent machinery (like orthogonal projections onto subspaces), and would require extra machinery unnecessarily. The one-vector projection isolates exactly the needed inequality with minimal apparatus.

How does Cauchy–Schwarz let us define the angle between abstract vectors?

For nonzero u,vu,v, Cauchy–Schwarz guarantees

−1≤⟨u,v⟩∥u∥∥v∥≤1-1 \le \frac{\langle u,v\rangle}{\|u\|\|v\|} \le 1

when the inner product is real-valued (in the complex case, one takes the real part or the modulus, depending on convention). Since cosine is a bijection from [0,π][0,\pi] onto [−1,1][-1,1], there is a unique θ∈[0,π]\theta \in [0,\pi] with cos⁡θ=⟨u,v⟩∥u∥∥v∥\cos\theta = \dfrac{\langle u,v\rangle}{\|u\|\|v\|}. This θ\theta is defined to be the angle between uu and vv; without the inequality, this ratio could exceed 11 in absolute value and no such angle would exist.

Is the complex case genuinely different from the real case?

The proof is essentially identical, but care is needed with conjugates: the inner product is conjugate-linear in one slot, and the scalar c=⟨v,u⟩/∥u∥2c = \langle v,u\rangle/\|u\|^2 used in the projection may be a complex number. The final inequality is stated for ∣⟨u,v⟩∣|\langle u,v\rangle|, the complex modulus, since ⟨u,v⟩\langle u,v\rangle itself need not be real. All the algebra in the proof — the cross-term cancellation, the Pythagorean expansion — goes through with ∣c∣2|c|^2 replacing c2c^2, exactly as written above.

Does Cauchy–Schwarz hold for seminorms, where ∥v∥=0\|v\|=0 does not force v=0v=0?

Yes, in the following weaker sense: if the form is only positive semi-definite (i.e., ⟨w,w⟩≥0\langle w,w\rangle \ge 0 for all ww, but possibly ⟨w,w⟩=0\langle w,w\rangle=0 for some w≠0w \ne 0), the inequality ∣⟨u,v⟩∣≤∥u∥∥v∥|\langle u,v\rangle| \le \|u\|\|v\| still holds, because Step 3 of the proof only used ∥w∥2≥0\|w\|^2 \ge 0, not strict positive-definiteness. What changes is the equality analysis: equality no longer forces linear dependence, only that the "difference vector" w=v−cuw = v-cu lies in the null space {x:⟨x,x⟩=0}\{x : \langle x,x\rangle = 0\} of the seminorm, which can be nontrivial.

What is the discrete version of Cauchy–Schwarz that Cauchy originally proved?

For real numbers a1,…,ana_1,\dots,a_n and b1,…,bnb_1,\dots,b_n,

(∑i=1naibi)2≤(∑i=1nai2)(∑i=1nbi2).\left(\sum_{i=1}^n a_ib_i\right)^2 \le \left(\sum_{i=1}^n a_i^2\right)\left(\sum_{i=1}^n b_i^2\right).

This is exactly the abstract inequality applied to V=RnV = \mathbb{R}^n with the standard dot product ⟨a,b⟩=∑aibi\langle a,b\rangle = \sum a_ib_i, so it is a special case, not a separate result. Cauchy's original 1821 proof used a discrete algebraic identity (a special case of the Lagrange identity) rather than the projection argument, but the two proofs are secretly the same computation viewed through different lenses.

Why does the Cauchy–Schwarz inequality matter for the triangle inequality?

When expanding ∥u+v∥2=∥u∥2+2 Re⟨u,v⟩+∥v∥2\|u+v\|^2 = \|u\|^2 + 2\,\mathrm{Re}\langle u,v\rangle + \|v\|^2 (in the real case, ∥u+v∥2=∥u∥2+2⟨u,v⟩+∥v∥2\|u+v\|^2 = \|u\|^2+2\langle u,v\rangle+\|v\|^2), the cross term ⟨u,v⟩\langle u,v\rangle must be controlled to show ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\|+\|v\|. Cauchy–Schwarz supplies exactly the bound ⟨u,v⟩≤∣⟨u,v⟩∣≤∥u∥∥v∥\langle u,v\rangle \le |\langle u,v\rangle| \le \|u\|\|v\|, which converts the expansion into ∥u+v∥2≤(∥u∥+∥v∥)2\|u+v\|^2 \le (\|u\|+\|v\|)^2, and taking square roots gives the triangle inequality. Without Cauchy–Schwarz, there is no known elementary way to control that cross term in general inner product spaces.

Before this

Where this leads