Part VII · Foundations Revisited — 20

Generalized Measurement

Chapter 03’s measurement gave one of a fixed set of sharp answers. But that’s only the tidiest case. The full theory — the POVM — allows softer questions, including the wisest answer of all: “I don’t know.”

↩ before you start · keep these handy

·From Ch. 3: a projective measurement forces one of a fixed set of sharp answers, every time.

·From Ch. 15: two states overlap by s = cos(φ/2) — 0 = opposite, 1 = identical.

·From No-Cloning: non-orthogonal states can never be told apart with perfect reliability.

🔑 symbol decoder · every new mark, in plain words

POVMPositive Operator-Valued Measure — the general kind of measurement, sharp ones included. Eᵢone outcome piece — a positive operator (its "probability detector" for outcome i). Σ Eᵢ = Ithe pieces must cover everything — total probability stays 1. Tr(Eᵢρ)the probability of outcome i on state ρ — the general Born rule. Naimarkthe theorem: any POVM = a sharp projective measurement on a bigger space.

feel

When the best answer is “don’t know”

Two non-orthogonal states overlap, so no measurement can always tell them apart (no-cloning, chapter 7a, is a cousin of this). A projective measurement forces a verdict every time — and so it must sometimes be wrong. A generalized measurement has another option: hold three outcomes instead of two, and let the third be an honest “inconclusive.” The reward for occasionally admitting defeat is that when it does commit, it is never mistaken.

🍷 everyday picture

A careful wine taster is handed a glass that’s either Merlot or Cabernet. A reckless taster always blurts a guess — and is wrong a fair bit when the two are similar. A wise one allows a third response: “definitely Merlot,” “definitely Cabernet,” or “I’d need another sip.” Whenever they commit, they’re right; the price is occasionally abstaining. A POVM is exactly that extra, honest “not sure” outcome — unavailable to a forced two-way guess.

recapAllowing an “inconclusive” outcome buys a verdict that’s never wrong when it commits.

play

Tell two overlapping states apart

Alice sends one of two equally-likely states ψ₀, ψ₁. The 3-outcome POVM either declares a state — always correctly — or returns “?”. Widen the angle and the states pull apart, shrinking the inconclusive slice; narrow it and honest doubt takes over.

▸ unambiguous discriminationΣ Eᵢ = I

→ “definitely ψ₀”{{ p0Pct }}
→ “definitely ψ₁”{{ p1Pct }}
→ “?” inconclusive{{ pIncPct }}

              angle φ
              
            
overlap s{{ overlap }}
error rate0%
if forced to guess{{ projErr }} err
{{ desc }}

recapThe more the states overlap, the bigger the “?” slice — but a committed answer is always correct.

math

A POVM is a set of positive pieces

Drop the demand that outcomes be sharp projectors. A POVM is just a set of positive operators E_i — one per outcome — that add to the identity, so probabilities still sum to one:

Ei ≥ 0,   Σi Ei = I,   P(i) = Tr(Ei ρ)
E₀ ⊥ ψ₁ ⇒ a click means “ψ₀” with certainty  // never fires on ψ₁
E₁ ⊥ ψ₀ ⇒ a click means “ψ₁” with certainty
E₂ = I − E₀ − E₁ ⇒ the “?” that mops up the rest, prob = s

Projective measurement is the special case where every E_i is a sharp projector and there are exactly as many as the dimension. By enlarging the toolkit — or, equivalently, measuring projectively on a bigger space (Naimark’s theorem) — you buy options projective measurement never had, like a guaranteed-correct verdict at the price of sometimes abstaining.

✎ worked example · two states 120° apart

overlap s = cos(120°/2) = cos60° = 0.5.
optimal inconclusive rate = s = 50% → half the shots return “?”.
each confident verdict = (1 − s)/2 = 25% — and those are never wrong.
a forced projective guess would instead be wrong about ½(1−sin60°) ≈ 7% of the time — you trade 0 errors for 50% abstentions. ✓

recapA POVM is positive pieces Eᵢ summing to I; projective measurement is just the sharp special case.

⚠ common misconception

“Measurement always gives one of N sharp, orthogonal answers.” That’s the textbook starter case, not the rule. A POVM can have more outcomes than dimensions and answers that aren’t orthogonal — which is exactly what let us add a third “?” outcome to a 2-dimensional qubit. The projective picture from chapter 03 is true, just incomplete.

With this, the toolkit Part VII set out to recover is complete: interference, interaction-free measurement, the partial trace, the Schmidt view of entanglement, and now general measurement. The final part assembles them into the one language — quantum channels — that holds gates, noise, and measurement all at once.

✓ you can now

✓define a POVM as positive operators Eᵢ with Σ Eᵢ = I and P(i) = Tr(Eᵢρ)

✓use a 3-outcome POVM to tell overlapping states apart with zero error, at the cost of abstaining

✓see projective measurement as the special case, and why more outcomes than dimensions is allowed

← 19 Schmidt Decomposition end of Part VII · roadmap →