University Physics V · Mathematical Foundations of Quantum Physics · 1.3

Vector Spaces over C & Their Inner Products

Quantum mechanics does its accounting in a complex vector space, and this is the week you learn to audit one. Check the axioms on a candidate space, conjugate the correct slot of every inner product, and treat Cauchy–Schwarz as the inequality you will still be using when it is called the uncertainty principle.

Open the lab ↓Practice & worked examples ↓

Build the model

Connect the measurement to the mechanism.

Quantum mechanics makes a structural bet before it makes a physical one: every state is a vector in a complex vector space, and every question about two states is answered by one complex number, their inner product. The vector part is forced by superposition — the Schrödinger equation is linear, so sums and scalar multiples of solutions must be admissible states — and the scalars must be C rather than R because a relative phase between components produces observable interference. The inner product ⟨φ|ψ⟩ is the added structure: linear in the ket, conjugate-linear in the bra, conjugate-symmetric under exchange.

That one conjugate is load-bearing. It makes ⟨ψ|ψ⟩ real and positive, so a norm exists, normalisation to ⟨ψ|ψ⟩ = 1 is possible, and the squared moduli Born's rule will read as probabilities are guaranteed non-negative. From the same axioms, with no physics added, follows Cauchy–Schwarz: |⟨φ|ψ⟩|² ≤ ⟨φ|φ⟩⟨ψ|ψ⟩, the inequality that caps every overlap probability at one and returns in the uncertainty unit as the Robertson bound.

The cost of the bet appears when the vectors are wavefunctions: L² is infinite-dimensional, membership must be checked integral by integral, and completeness — every Cauchy sequence converging inside the space — is assumed this week and earned only in a functional-analysis course.

Simple definition: A complex inner-product space is a vector space over C carrying a map ⟨φ|ψ⟩ that is linear in its second argument, conjugate-linear in its first, conjugate-symmetric, and positive-definite — the minimal structure that gives each state a length and each pair an overlap.
Example: In C², ⟨u|v⟩ = u₁*v₁ + u₂*v₂: with u = (1, i) and v = (3, 2), ⟨u|v⟩ = 3 + (−i)(2) = 3 − 2i, while ⟨u|u⟩ = 1 + 1 = 2 — real and positive, as a squared length must be.

Sesquilinearity (physics convention)⟨φ|c₁ψ₁ + c₂ψ₂⟩ = c₁⟨φ|ψ₁⟩ + c₂⟨φ|ψ₂⟩⟨cφ|ψ⟩ = c*⟨φ|ψ⟩

Fixes every later manipulation: a coefficient pulled out of a bra comes out conjugated.

c, c₁, c₂ ∈ C; kets carry the linearity, bras the conjugate — most mathematics texts conjugate the opposite slot

Conjugate symmetry⟨φ|ψ⟩ = ⟨ψ|φ⟩*

Forces ⟨ψ|ψ⟩ = ⟨ψ|ψ⟩*, hence real — the prerequisite for it to be a squared length.

swapping the arguments conjugates the value, so the modulus |⟨φ|ψ⟩| is order-blind

The two working inner products⟨u|v⟩ = Σᵢ uᵢ* vᵢ on Cⁿ⟨f|g⟩ = ∫ f*(x) g(x) dx on L²

The Cⁿ form is u†v — conjugate-transpose row times column; the L² form is its continuum limit.

uᵢ, vᵢ ∈ C are components on one basis; the L² integral runs over the whole domain, so a 1D ψ carries units m^(−1/2)

Induced norm and normalisation‖ψ‖ = √⟨ψ|ψ⟩ψ̂ = ψ/‖ψ‖ so that ⟨ψ̂|ψ̂⟩ = 1

Normalisation spends the free scale of a linear equation on making total probability equal one.

positive-definiteness: ‖ψ‖ > 0 for every ψ ≠ 0, so only the zero vector has zero norm

Cauchy–Schwarz inequality|⟨φ|ψ⟩|² ≤ ⟨φ|φ⟩⟨ψ|ψ⟩

Caps every overlap probability at 1, proves ⟨f|g⟩ converges on L², and returns as the Robertson bound ΔA ΔB ≥ ½|⟨[A, B]⟩|.

equality iff φ and ψ are linearly dependent — ψ = cφ for some c ∈ C, both kets on a single ray

Square-integrabilityψ ∈ L²(R) ⇔ ∫ |ψ(x)|² dx < ∞

The admission test for wavefunctions: e^(−x²/2) passes, the plane wave e^(ikx) fails — sharp momentum is not a vector of the space.

only finiteness of the integral decides membership, not its value

Superposition forces the vector axioms

The Schrödinger equation is linear: if ψ₁ and ψ₂ solve it, so does c₁ψ₁ + c₂ψ₂ for any complex c₁, c₂. Physics therefore hands you closure under addition and scalar multiplication, and the remaining axioms — associativity, a zero vector, additive inverses, distributivity — are bookkeeping that Cⁿ and function spaces satisfy automatically. The scalars must be complex, not real: a relative phase between components, as in (|0⟩ + e^(iδ)|1⟩)/√2, changes measurable interference terms, so multiplying one component by e^(iδ) genuinely changes the state. Two cautions. The zero vector is a bookkeeping element, not a physical state — it cannot be normalised. And a global phase e^(iα)ψ, applied to the whole vector, changes no inner-product modulus and so no prediction: states are really rays, vectors up to overall complex scale.

Conjugate the bra, keep the ket linear

The inner product eats two vectors and returns one complex number, but not symmetrically. In the physics convention ⟨φ|ψ⟩ is linear in the ket, ⟨φ|cψ⟩ = c⟨φ|ψ⟩, and conjugate-linear in the bra, ⟨cφ|ψ⟩ = c*⟨φ|ψ⟩. On C² this is a recipe: conjugate the first vector's components, then multiply and add. With u = (2, 1+i) and v = (i, 1): ⟨u|v⟩ = 2*⋅i + (1+i)*⋅1 = 2i + 1 − i = 1 + i. Conjugate symmetry then says ⟨v|u⟩ = (1 + i)* = 1 − i — swapping the slots conjugates the number, and only the modulus survives the swap. Mind the literature: most mathematics texts put the conjugate on the second slot instead. Both conventions are internally consistent; mixing them mid-calculation is the actual error, and it announces itself as a sign flip on every imaginary part.

Positive-definiteness turns overlap into length

Set φ = ψ and conjugate symmetry makes ⟨ψ|ψ⟩ real; the positive-definiteness axiom demands more: ⟨ψ|ψ⟩ > 0 for every nonzero ψ. That is what entitles you to define ‖ψ‖ = √⟨ψ|ψ⟩ and call it a length. On C², ⟨ψ|ψ⟩ = Σ|ψᵢ|², a sum of squared moduli, positive term by term. Take |ψ⟩ = (1+i)|0⟩ + 2|1⟩: ⟨ψ|ψ⟩ = |1+i|² + |2|² = 2 + 4 = 6, so ψ̂ = ψ/√6 is the unit representative of the ray, and the numbers |⟨0|ψ̂⟩|² = 2/6 and |⟨1|ψ̂⟩|² = 4/6 already sum to one — Born's rule will simply read them off. The axiom has teeth: drop the conjugate from the recipe and the supposed norm of (1, i) computes to 1 + i² = 0, a nonzero vector of zero length, and the whole probability interpretation dies with the division by ‖ψ‖.

Cauchy–Schwarz from the axioms alone

For any two vectors, |⟨φ|ψ⟩|² ≤ ⟨φ|φ⟩⟨ψ|ψ⟩. The proof is one application of positive-definiteness: for φ ≠ 0, the vector ψ − (⟨φ|ψ⟩/⟨φ|φ⟩)φ — ψ with its component along φ removed — has non-negative squared norm, and expanding that squared norm and rearranging gives the inequality. The equality condition rides along free: zero headroom forces the residual vector to vanish, so ψ = cφ. Two consequences arrive immediately. For unit vectors the squared overlap is at most 1, so Born's rule can never manufacture a probability above one. And the triangle inequality ‖φ + ψ‖ ≤ ‖φ‖ + ‖ψ‖ follows, making the norm an honest notion of distance. The payoff deferred to the uncertainty unit: applied to the vectors (A − ⟨A⟩)|ψ⟩ and (B − ⟨B⟩)|ψ⟩ for Hermitian A and B, the same inequality becomes Robertson's bound ΔA ΔB ≥ ½|⟨[A, B]⟩| — the uncertainty principle is Cauchy–Schwarz wearing operators.

L²: an infinite-dimensional space of functions

Wavefunctions live in L²(R), the functions with ∫|ψ|² dx finite, under ⟨f|g⟩ = ∫ f*(x) g(x) dx. Membership is checked integral by integral: e^(−x²/2) passes; 1/(1+x²) passes; the plane wave e^(ikx) has |ψ|² = 1 everywhere and fails, which is why sharp-momentum states are not vectors of this space and are handled later as δ-normalised distributions. That L² is closed under addition is itself a small theorem — |f+g|² ≤ 2|f|² + 2|g|² does it — and Cauchy–Schwarz guarantees ⟨f|g⟩ converges whenever both members pass the test. Two fine points are assumed rather than proved this week. Functions equal except on a set of measure zero are identified, otherwise positive-definiteness would fail on functions nonzero only at isolated points. And the space is complete — Cauchy sequences converge inside it — which is exactly what the name Hilbert space adds to inner-product space, and what expansion in an infinite basis will silently rely on.

Auditing a candidate inner product

Given a proposed ⟨⋅|⋅⟩, test the axioms in order of how often they fail. Conjugate symmetry first: on C², the unconjugated form Σuᵢvᵢ is symmetric but not conjugate-symmetric, and its self-overlaps are not even real — feed it u = (1, 1+i) and it returns 1 + (1+i)² = 1 + 2i, a complex number no square root can turn into a length. Linearity in the ket second: it usually holds by construction. Positive-definiteness last, because it fails quietly. The weighted product ⟨f|g⟩ = ∫ f* g w(x) dx is a legitimate inner product whenever w(x) > 0 — the Sturm–Liouville problems closing this unit use exactly this — but let w vanish on an interval and any function supported there acquires zero norm: the form degrades to a seminorm, and normalisation divides by zero. The habit to build: name the space, name the inner product, and verify positivity before trusting any norm, probability or orthogonality claim computed from it.

Change one variable at a time

Make the relationship visible.

Interactive model

Amplitude a on |0⟩1.0

Amplitude b on |1⟩0.6

Relative phase δ60 °

Set a = b = 1 and drag δ to 0: the marker climbs onto the dashed ceiling, because |ψ⟩ is then exactly √2⋅|φ⟩ and Cauchy–Schwarz saturates only for kets on one ray. Unbalance a and b and no phase will close the gap — watch the headroom readout refuse to reach zero.

NORM ‖ψ‖1.17

OVERLAP |⟨φ|ψ⟩|²0.98

C–S CEILING ⟨φ|φ⟩⟨ψ|ψ⟩1.36

HEADROOM0.38

Live interpretationNORM ‖ψ‖: 1.17. OVERLAP |⟨φ|ψ⟩|²: 0.98. C–S CEILING ⟨φ|φ⟩⟨ψ|ψ⟩: 1.36. HEADROOM: 0.38

Catch the common trap

Explain before calculating.

Misconception check

The inner product on a complex space is not the dot product with i's carried along: one slot must be conjugated. Students write ⟨u|v⟩ = Σuᵢvᵢ because it worked on Rⁿ, where conjugation is invisible. Over C it is fatal: the length it assigns to (1, i) is √(1 + i²) = 0, a nonzero state with zero norm, so normalisation divides by zero — and self-overlaps can come out negative or complex, which no Born rule can read as a probability. What is convention is only which slot carries the conjugate: physics picks the bra, most mathematics texts pick the second argument. That some slot carries it is not convention at all — it is forced by demanding ⟨ψ|ψ⟩ real and positive. A related slip is writing ⟨φ|ψ⟩ = ⟨ψ|φ⟩. The true statement is conjugate symmetry, ⟨φ|ψ⟩ = ⟨ψ|φ⟩*; the two orders agree only when the overlap happens to be real.

In C² with the physics convention ⟨u|v⟩ = Σᵢ uᵢ* vᵢ, let |u⟩ = (1, 2i)ᵀ and |v⟩ = (3, 1+i)ᵀ. What is ⟨u|v⟩?

Choose an answer to test the model.

Practice & worked examples

Reason from the model, then test the result.

EasyIn C² with orthonormal basis (|0⟩, |1⟩), a device prepares |ψ⟩ = 3|0⟩ + 4i|1⟩. Normalise the state, read off the Born probabilities of the two basis outcomes, and show that the global phase e^(iπ/7) changes neither.

Compute the squared norm with the conjugate on the bra side: ⟨ψ|ψ⟩ = 3*⋅3 + (4i)*⋅(4i) = 9 + (−4i)(4i) = 9 + 16 = 25, so ‖ψ‖ = 5.
Divide by the norm: ψ̂ = (3/5)|0⟩ + (4i/5)|1⟩. Check: (3/5)² + |4i/5|² = 9/25 + 16/25 = 1.
The Born weights are squared moduli of the coefficients: P(0) = |3/5|² = 0.36 and P(1) = |4i/5|² = 0.64 — the i contributes nothing, since |4i|² = 16.
Multiply by e^(iπ/7): every coefficient gains a modulus-one factor, so ⟨ψ|ψ⟩ = |e^(iπ/7)|²⋅25 = 25 and each |cₙ|² is untouched. A global phase moves the vector but not the ray, so no prediction changes.

Answerψ̂ = (3/5)|0⟩ + (4i/5)|1⟩, with P(0) = 0.36 and P(1) = 0.64; the global phase leaves the norm at 5 and both probabilities fixed.

MediumFor |φ⟩ = (1, i)ᵀ and |ψ⟩ = (2, 1)ᵀ in C², verify the Cauchy–Schwarz inequality explicitly, find the overlap probability |⟨φ̂|ψ̂⟩|² of the normalised states, and confirm from the equality condition that the inequality had to be strict.

Overlap first: ⟨φ|ψ⟩ = 1*⋅2 + i*⋅1 = 2 − i, so |⟨φ|ψ⟩|² = 2² + (−1)² = 5.
Norms: ⟨φ|φ⟩ = |1|² + |i|² = 2 and ⟨ψ|ψ⟩ = |2|² + |1|² = 5, so the Cauchy–Schwarz ceiling is ⟨φ|φ⟩⟨ψ|ψ⟩ = 10.
Compare: 5 ≤ 10 — the inequality holds, with headroom 5.
Normalising divides the squared overlap by both squared norms: |⟨φ̂|ψ̂⟩|² = 5/10 = 0.50, a legitimate probability.
Equality would require ψ = cφ for one complex c: the first components force c = 2, the second force c = 1/i = −i. No single c exists, the kets are not on one ray, and strict inequality was guaranteed before any arithmetic.

Answer|⟨φ|ψ⟩|² = 5 ≤ 10 = ⟨φ|φ⟩⟨ψ|ψ⟩; overlap probability 0.50; strict because no c ∈ C gives ψ = cφ.

HardOn L²(R), take ψ₁(x) = e^(−x²/2) and ψ₂(x) = x⋅e^(−x²/2), given ∫e^(−x²)dx = √π and ∫x²e^(−x²)dx = √π/2 over the real line. Show the two are orthogonal, normalise both, and build the normalised equal-weight superposition — then say what this pair becomes later in the course.

Membership first: both functions decay like a Gaussian, so every integral below is finite and both belong to L² — unlike the plane wave e^(ikx).
Orthogonality: ⟨ψ₁|ψ₂⟩ = ∫ x⋅e^(−x²) dx. The integrand is odd and the domain symmetric, so the integral is 0 — orthogonal by parity, no computation needed.
Norms: ⟨ψ₁|ψ₁⟩ = ∫e^(−x²)dx = √π ≈ 1.772, and ⟨ψ₂|ψ₂⟩ = ∫x²e^(−x²)dx = √π/2 ≈ 0.886.
Normalise: ψ̂₁ = π^(−1/4)⋅e^(−x²/2) and ψ̂₂ = √2⋅π^(−1/4)⋅x⋅e^(−x²/2). Check the second: (√2⋅π^(−1/4))² × √π/2 = (2/√π)(√π/2) = 1.
Superpose: χ = (ψ̂₁ + ψ̂₂)/√2 has ⟨χ|χ⟩ = ½(1 + 0 + 0 + 1) = 1 — orthogonality kills the cross terms, so the weights ½ and ½ are already Born probabilities.
These are the n = 0 and n = 1 harmonic-oscillator eigenfunctions up to scale; their orthogonality by parity previews the theorem that eigenvectors of a Hermitian operator with distinct eigenvalues are always orthogonal.

Answer⟨ψ₁|ψ₂⟩ = 0; ψ̂₁ = π^(−1/4)e^(−x²/2), ψ̂₂ = √2⋅π^(−1/4)⋅x⋅e^(−x²/2); χ = (ψ̂₁ + ψ̂₂)/√2 is normalised with weight ½ on each mode.