# POVMs & Generalized Measurement > Projective measurement is not the most general thing you can do to a quantum state. If you let the > system interact with an ancilla, measure *that* projectively, and forget the ancilla, the statistics > seen by the original system obey a richer rule — a **POVM**. This is the measurement formalism that > appears whenever the world is realistic: noisy detectors, partial information, optimal discrimination > of states that overlap. We build it from measurement operators, prove projective measurement is the > special case, and use a POVM to do something projective measurement *cannot*: cleanly handle > non-orthogonal states. ## Learning Objectives After this lesson you will be able to: 1. State the **general measurement postulate** in terms of measurement operators $\{M_m\}$ with $\sum_m M_m^\dagger M_m = I$, and compute outcome probabilities and post-measurement states. 2. Define **POVM elements** $E_m = M_m^\dagger M_m \succeq 0$ with $\sum_m E_m = I$, and explain what a POVM does and does not determine. 3. Show **projective measurement** is the special case $M_m = P_m$ (orthogonal projectors). 4. State **Naimark's dilation theorem**: every POVM is a projective measurement on a larger space. 5. Construct a **2-state discrimination POVM** for non-orthogonal states (unambiguous discrimination) and explain why projective measurement is insufficient. --- ## Intuition A projective measurement is a clean, repeatable question with orthogonal answers. But many real "measurements" are not like that. A photon detector clicks or doesn't, with efficiency $<1$. You couple a qubit weakly to a meter and read a noisy pointer. You are handed one of two non-orthogonal states and asked to guess which — and you would rather sometimes answer "I don't know" than be wrong. All of these are captured by allowing the measurement to be **a projective measurement on a bigger system**: append an ancilla in a known state, apply a joint unitary (let them interact), then measure the ancilla projectively. From the original system's point of view, the resulting outcome probabilities are *not* of the form $\langle\psi|P_m|\psi\rangle$ for orthogonal $P_m$ — they are $\langle\psi|E_m|\psi\rangle$ for a family of **positive operators** $E_m$ that sum to the identity but need not be orthogonal, need not be projectors, and can even outnumber the dimension. That family is a **POVM**, and Naimark's theorem says the "bigger system" story is always available: POVMs are exactly the shadows of projective measurements. --- ## Theory ### General measurement operators > **General measurement postulate.** A measurement is described by a collection of **measurement > operators** $\{M_m\}$ acting on the system's Hilbert space, indexed by the outcomes $m$, satisfying the > **completeness relation** > $$ \sum_m M_m^\dagger M_m = I . $$ > Measuring state $|\psi\rangle$ yields outcome $m$ with probability > $$ p_m = \langle\psi|M_m^\dagger M_m|\psi\rangle = \|M_m|\psi\rangle\|^2, $$ > and the post-measurement state, given outcome $m$, is > $$ |\psi_m\rangle = \frac{M_m|\psi\rangle}{\sqrt{p_m}} \qquad (p_m>0). $$ These are valid probabilities: $p_m = \|M_m|\psi\rangle\|^2 \ge 0$, and completeness gives $\sum_m p_m = \langle\psi|\big(\sum_m M_m^\dagger M_m\big)|\psi\rangle = \langle\psi|I|\psi\rangle = 1$. Note the post-measurement state depends on the *operator* $M_m$, not just on the probability — and two different $M_m$ can produce the same probabilities but different collapses (see below). ### POVM elements Define, for each outcome, $$ \boxed{\;E_m := M_m^\dagger M_m\;}. $$ Each $E_m$ is **positive semidefinite** ($E_m\succeq0$, since $\langle\psi|E_m|\psi\rangle = \|M_m|\psi\rangle\|^2\ge0$ — it has the form $B^\dagger B$, [0.1.6](../../term-0-foundations/c01-linear-algebra/06-special-operators.md)), and the completeness relation becomes $$ \sum_m E_m = I . $$ > **Definition (POVM).** A **Positive Operator-Valued Measure** is a set of positive operators > $\{E_m\}$ with $E_m\succeq0$ and $\sum_m E_m = I$. The outcome probabilities are > $$ p_m = \langle\psi|E_m|\psi\rangle \quad\text{(pure)} \qquad\text{or}\qquad p_m = \operatorname{Tr}(\rho\,E_m)\ \text{(general)}. $$ **What a POVM determines, and what it doesn't.** The POVM $\{E_m\}$ fixes the *outcome statistics* for every state — that is all probabilities ever depend on. It does **not** fix the post-measurement state: given $E_m$, any $M_m = U_m\sqrt{E_m}$ with $U_m$ unitary satisfies $M_m^\dagger M_m = \sqrt{E_m}\,U_m^\dagger U_m\sqrt{E_m} = E_m$ (polar freedom). So *if you only care about outcome probabilities* — as in state discrimination, tomography, and estimation — the POVM $\{E_m\}$ is the complete description, and you never need to name the $M_m$. When you also need the post-measurement state (e.g. feed-forward), you must specify the $\{M_m\}$. ### Projective measurement is the special case Take $M_m = P_m$, a complete set of orthogonal projectors ([Lesson 1](01-projective-measurement.md)). Then since $P_m^\dagger = P_m$ and $P_m^2 = P_m$, $$ E_m = M_m^\dagger M_m = P_m^\dagger P_m = P_m^2 = P_m, $$ so the POVM elements are the projectors themselves, $p_m = \langle\psi|P_m|\psi\rangle$, and the collapse $M_m|\psi\rangle/\sqrt{p_m} = P_m|\psi\rangle/\sqrt{p_m}$ is exactly the projection postulate. Projective measurements are precisely the POVMs whose elements are **orthogonal projectors** ($E_jE_k = \delta_{jk}E_k$). The defining extra structure that general POVMs drop is *orthogonality and idempotence* of the elements — a general $E_m$ need be neither a projector nor orthogonal to the others. Two structural freedoms a POVM has that a projective measurement does not: - **More outcomes than dimensions.** A qubit ($d=2$) admits projective measurements with at most $2$ outcomes (rank-1) or $1$ (the trivial $\{I\}$). A POVM on a qubit can have $3$, $4$, … outcomes — e.g. the symmetric "trine" of three sub-normalized projectors at $120^\circ$, each $E_m = \tfrac23|\phi_m\rangle\langle\phi_m|$, summing to $I$. This is impossible projectively. - **Non-repeatability.** $E_m$ need not be idempotent, so re-measuring need not reproduce the outcome. Generalized measurements are not, in general, repeatable. ### Why POVMs are the most general: Naimark dilation The claim that POVMs are the *most general* measurement is a theorem, not a definition. > **Naimark's dilation theorem.** Every POVM $\{E_m\}$ on a system $\mathcal H_S$ can be realized as a > **projective measurement on a larger space**: there is an ancilla $\mathcal H_A$ in a fixed state > $|a_0\rangle$, a joint unitary $U$ on $\mathcal H_S\otimes\mathcal H_A$, and an orthogonal projective > measurement $\{\Pi_m\}$ on the enlarged space such that > $$ \langle\psi|E_m|\psi\rangle = \big(\langle\psi|\otimes\langle a_0|\big)\,U^\dagger \Pi_m U\,\big(|\psi\rangle\otimes|a_0\rangle\big)\quad\forall\,|\psi\rangle,m. $$ *Idea of the construction.* Choose $M_m=\sqrt{E_m}$. Define an isometry $V:\mathcal H_S\to\mathcal H_S\otimes\mathcal H_A$ by $V|\psi\rangle = \sum_m \big(M_m|\psi\rangle\big)\otimes|m\rangle$, where $\{|m\rangle\}$ is an orthonormal ancilla basis. Then $V^\dagger V = \sum_m M_m^\dagger M_m = I$, so $V$ is an isometry and extends to a unitary $U$ on the joint space (with $V|\psi\rangle = U(|\psi\rangle\otimes|a_0\rangle)$). Measuring the ancilla with $\Pi_m = I\otimes|m\rangle\langle m|$ gives, by the mixed-product rule, $\| \Pi_m V|\psi\rangle\|^2 = \|M_m|\psi\rangle\|^2 = \langle\psi|E_m|\psi\rangle$. ∎ (sketch) Conversely — and this is the half that makes "most general" precise — **any** physical measurement (append ancilla in a known state, apply a unitary, measure projectively, and read the marginal statistics on the system) produces probabilities of the form $\langle\psi|E_m|\psi\rangle$ for some POVM. *Proof of the converse.* Let the system start in $|\psi\rangle$, ancilla in $|a_0\rangle$, joint unitary $U$, and ancilla projective measurement $\{\Pi_m = I\otimes Q_m\}$ with $\{Q_m\}$ orthogonal projectors on $\mathcal H_A$. The outcome-$m$ probability is $$ p_m = \big\| \Pi_m\,U\,(|\psi\rangle\otimes|a_0\rangle)\big\|^2 = \big(\langle\psi|\otimes\langle a_0|\big)U^\dagger\Pi_m U\big(|\psi\rangle\otimes|a_0\rangle\big) = \langle\psi|E_m|\psi\rangle, $$ where $E_m := \big(I\otimes\langle a_0|\big)U^\dagger\Pi_m U\big(I\otimes|a_0\rangle\big)$ acts on $\mathcal H_S$ alone. Each $E_m$ is positive (it is a "compression" of the projector $U^\dagger\Pi_m U\succeq0$: $\langle\psi|E_m|\psi\rangle = \|\,\Pi_m U(|\psi\rangle\otimes|a_0\rangle)\|^2\ge0$) and they sum to $I$ because $\sum_m\Pi_m = I\otimes I$ and $\langle a_0|a_0\rangle=1$. So the marginal statistics are exactly a POVM. ∎ Together the two directions say: **the POVMs are exactly the achievable measurements** — no more, no less. This is the operational reason the formalism stops at POVMs. (The same enlarge-and-forget logic gives Kraus operators for noisy channels in [Term 4.1](../../term-4-noise-hardware/c01-open-systems/02-kraus-operators-channels.md); a measurement is a channel with a classical output.) ### Application: discriminating non-orthogonal states [Lesson 1](01-projective-measurement.md) proved orthogonal states are perfectly distinguishable and non-orthogonal states are not. POVMs let us still do something *useful and error-controlled* with non-orthogonal states. We treat **unambiguous state discrimination (USD)**: never answer wrongly, at the price of sometimes answering "don't know". Suppose we are given either $|\psi_0\rangle$ or $|\psi_1\rangle$ (equiprobable, both known to us, with $0<|\langle\psi_0|\psi_1\rangle|<1$). Use a 3-outcome POVM $\{E_0, E_1, E_?\}$: - $E_0$ fires **only** when the state is $|\psi_1\rangle$ — wait, we orient it the standard way: outcome $0$ means "definitely $|\psi_0\rangle$", outcome $1$ means "definitely $|\psi_1\rangle$", and $?$ means "inconclusive". To never err, $E_0$ must be **orthogonal to $|\psi_1\rangle$** (so it can never fire on $|\psi_1\rangle$), and $E_1$ orthogonal to $|\psi_0\rangle$. Let $|\psi_1^\perp\rangle$ be a unit vector with $\langle\psi_1|\psi_1^\perp\rangle = 0$, and likewise $|\psi_0^\perp\rangle$. Set $$ E_0 = c_0\,|\psi_1^\perp\rangle\langle\psi_1^\perp|,\qquad E_1 = c_1\,|\psi_0^\perp\rangle\langle\psi_0^\perp|,\qquad E_? = I - E_0 - E_1, $$ with constants $c_0,c_1\ge0$ chosen as large as possible while keeping $E_?\succeq0$ (it must remain a valid POVM element). By construction $\langle\psi_1|E_0|\psi_1\rangle = 0$ and $\langle\psi_0|E_1|\psi_0\rangle = 0$ — **no error is ever made**. The probability of a *conclusive, correct* identification of $|\psi_0\rangle$ is $\langle\psi_0|E_0|\psi_0\rangle = c_0|\langle\psi_1^\perp|\psi_0\rangle|^2$. For the symmetric real case with overlap $s := |\langle\psi_0|\psi_1\rangle|$, optimizing $c_0=c_1$ subject to $E_?\succeq0$ gives the celebrated **Ivanovic–Dieks–Peres (IDP) bound**: the optimal inconclusive probability is exactly the overlap, $$ \Pr[?]_{\min} = s = |\langle\psi_0|\psi_1\rangle|,\qquad \Pr[\text{conclusive}]_{\max} = 1 - s . $$ As $s\to0$ (nearly orthogonal) we almost always succeed; as $s\to1$ (nearly identical) we almost always abstain — exactly the right limiting behavior. A 2-outcome *projective* measurement cannot achieve zero error here: its two rank-1 projectors are orthogonal, so at most one of them can be orthogonal to a given $|\psi_i\rangle$, and the other necessarily has nonzero overlap with the "wrong" state, forcing a nonzero error probability. The **third ("don't know") outcome** — unavailable projectively on a qubit — is what buys zero error. We construct this POVM explicitly in code below. --- ## Worked Examples ### Example 1 — A non-projective qubit POVM (the "trine") Let $|\phi_m\rangle = \cos\tfrac{2\pi m}{3}|0\rangle + \sin\tfrac{2\pi m}{3}|1\rangle$ for $m=0,1,2$ — three real states at $120^\circ$ on the Bloch sphere's $XZ$ great circle. Define $$ E_m = \tfrac23\,|\phi_m\rangle\langle\phi_m|,\qquad m=0,1,2. $$ **Positivity.** Each $E_m$ is $\tfrac23$ times a rank-1 projector, hence $\succeq0$. ✓ **Completeness.** Using $|\phi_m\rangle\langle\phi_m| = \tfrac12(I + \cos\tfrac{4\pi m}3 Z + \sin\tfrac{4\pi m}3 X)$ and $\sum_{m=0}^2\cos\tfrac{4\pi m}3 = \sum_{m=0}^2\sin\tfrac{4\pi m}3 = 0$ (three cube-roots-of-unity directions sum to zero): $$ \sum_{m=0}^2 E_m = \tfrac23\sum_m |\phi_m\rangle\langle\phi_m| = \tfrac23\cdot\tfrac32 I = I . \checkmark $$ This is a legitimate **3-outcome measurement on a 2-dimensional system** — impossible for a projective measurement, which on a qubit has at most $2$ outcomes. (The trine is the optimal POVM for several qubit estimation tasks; verified numerically below.) ### Example 2 — Unambiguous discrimination of $|0\rangle$ and $|+\rangle$ Take $|\psi_0\rangle = |0\rangle$ and $|\psi_1\rangle = |+\rangle = \tfrac1{\sqrt2}(|0\rangle+|1\rangle)$, with overlap $s = |\langle 0|+\rangle| = \tfrac1{\sqrt2}\approx0.707$. The orthogonal complements: $|\psi_1^\perp\rangle = |-\rangle$ (since $\langle+|-\rangle=0$) and $|\psi_0^\perp\rangle = |1\rangle$ (since $\langle0|1\rangle=0$). The USD POVM is $$ E_0 = c\,|-\rangle\langle-|,\qquad E_1 = c\,|1\rangle\langle1|,\qquad E_? = I - E_0 - E_1, $$ with $c$ as large as possible keeping $E_?\succeq0$. Working out the constraint (or quoting IDP), the optimum is $c = \dfrac{1}{1+s} = \dfrac{1}{1+1/\sqrt2}\approx 0.586$, giving an inconclusive probability $\Pr[?] = s = 1/\sqrt2\approx0.707$ and success $\Pr[\text{conclusive}] = 1-1/\sqrt2\approx0.293$. **Zero error, by construction.** If the state is $|\psi_1\rangle=|+\rangle$: $\langle+|E_0|+\rangle = c|\langle-|+\rangle|^2 = 0$, so we *never* report "$0$" — outcome $0$ unambiguously means $|0\rangle$. Symmetrically, outcome $1$ never fires on $|0\rangle$. The price of certainty is the $\sim71\%$ of runs that come back "don't know". The NumPy code below builds these $E_m$, checks $\sum_m E_m = I$ and $E_m\succeq0$, and simulates the statistics to confirm zero error. --- ## Hands-on (Python) We implement a general POVM, verify completeness and positivity, and simulate outcome statistics — first the trine, then unambiguous discrimination. ```python import numpy as np ket0 = np.array([1, 0], dtype=complex) ket1 = np.array([0, 1], dtype=complex) plus = (ket0 + ket1) / np.sqrt(2) minus = (ket0 - ket1) / np.sqrt(2) def is_povm(elements, tol=1e-9): """Check E_m >= 0 (Hermitian, nonneg eigenvalues) and sum_m E_m = I.""" d = elements[0].shape[0] total = np.zeros((d, d), dtype=complex) for E in elements: assert np.allclose(E, E.conj().T), "POVM element not Hermitian" w = np.linalg.eigvalsh((E + E.conj().T) / 2) assert np.all(w >= -tol), f"POVM element not PSD: min eig {w.min():.2e}" total += E return np.allclose(total, np.eye(d), atol=tol) def povm_probabilities(psi, elements): """p_m = .""" return np.array([np.real(psi.conj() @ (E @ psi)) for E in elements]) def sample_povm(psi, elements, shots, rng): """Sample outcome labels 0..M-1 from the POVM distribution.""" p = povm_probabilities(psi, elements) return rng.choice(len(elements), size=shots, p=p) # ---- The trine POVM: 3 outcomes on a qubit (impossible projectively) ----- trine_states = [ np.array([np.cos(2*np.pi*m/3), np.sin(2*np.pi*m/3)], dtype=complex) for m in range(3) ] trine = [(2/3) * np.outer(s, s.conj()) for s in trine_states] print("Trine is a valid POVM:", is_povm(trine)) # True print("Sum E_m =\n", np.round(sum(trine).real, 6)) # identity print("Probs on |0>:", np.round(povm_probabilities(ket0, trine), 4)) # sums to 1 ``` ```python # ---- Unambiguous discrimination of |0> and |+> --------------------------- s = abs(ket0.conj() @ plus) # overlap = 1/sqrt(2) c = 1 / (1 + s) # IDP-optimal weight E0 = c * np.outer(minus, minus.conj()) # fires only for |0> (orthogonal to |+>) E1 = c * np.outer(ket1, ket1.conj()) # fires only for |+> (orthogonal to |0>) E_inc = np.eye(2) - E0 - E1 # "don't know" usd = [E0, E1, E_inc] print("USD is a valid POVM:", is_povm(usd)) # True print("E_? eigenvalues (must be >=0):", np.round(np.linalg.eigvalsh(E_inc), 6)) # Zero error: outcome 0 never fires on |+>, outcome 1 never fires on |0> print("P(say '0' | state |+>) =", np.round(povm_probabilities(plus, usd)[0], 6)) # 0 print("P(say '1' | state |0>) =", np.round(povm_probabilities(ket0, usd)[1], 6)) # 0 # Inconclusive probability hits the IDP bound = overlap s print(f"P(?) = {povm_probabilities(ket0, usd)[2]:.4f}, overlap s = {s:.4f}") # equal ``` ```python # ---- Monte-Carlo: confirm zero error over many shots --------------------- rng = np.random.default_rng(7) shots = 100_000 # Prepare |0> half the time, |+> the other half; record (truth, outcome) errors = 0 labels = {0: "say-0", 1: "say-1", 2: "say-?"} for truth, state in [(0, ket0), (1, plus)]: outcomes = sample_povm(state, usd, shots, rng) # an ERROR is a confident-but-wrong report: say-1 when truth=0, or say-0 when truth=1 wrong = (outcomes == 1) if truth == 0 else (outcomes == 0) errors += int(wrong.sum()) print("Confident-but-wrong reports over 200k shots:", errors) # 0 -> unambiguous ``` > The empirical error count is **exactly zero** because the orthogonality constraints > $\langle\psi_1|E_0|\psi_1\rangle = \langle\psi_0|E_1|\psi_0\rangle = 0$ are built into the operators — > the simulation confirms the analytic guarantee rather than discovering it. All probability and > completeness checks are done in NumPy; a POVM is implemented on hardware via Naimark dilation > (ancilla + joint unitary + computational-basis read-out), which we revisit with Braket in Term 2. --- ## Exercises **E1 (easy).** Verify directly that the trine $\{E_m = \tfrac23|\phi_m\rangle\langle\phi_m|\}$ from Example 1 satisfies $\sum_m E_m = I$ for the explicit states $|\phi_0\rangle=|0\rangle$, $|\phi_1\rangle = -\tfrac12|0\rangle+\tfrac{\sqrt3}2|1\rangle$, $|\phi_2\rangle = -\tfrac12|0\rangle-\tfrac{\sqrt3}2|1\rangle$.
Solution Compute outer products: $|\phi_0\rangle\langle\phi_0| = \begin{psmallmatrix}1&0\\0&0\end{psmallmatrix}$, and $|\phi_1\rangle\langle\phi_1| = \begin{psmallmatrix}1/4 & -\sqrt3/4\\ -\sqrt3/4 & 3/4\end{psmallmatrix}$, $|\phi_2\rangle\langle\phi_2| = \begin{psmallmatrix}1/4 & \sqrt3/4\\ \sqrt3/4 & 3/4\end{psmallmatrix}$. Their sum is $\begin{psmallmatrix}3/2 & 0\\ 0 & 3/2\end{psmallmatrix} = \tfrac32 I$. Multiply by $\tfrac23$: $\sum_m E_m = \tfrac23\cdot\tfrac32 I = I$. ✓
**E2 (easy).** Show that if all $E_m$ in a POVM are orthogonal projectors then the post-measurement update (for $M_m=\sqrt{E_m}=E_m$) is repeatable, but that a general POVM need not be.
Solution If $E_m=P_m$ is a projector, $\sqrt{E_m}=P_m$, and we are back in the projective case: re-measuring gives the same outcome with certainty (Lesson 1 repeatability proof). For a general element such as a trine $E_m = \tfrac23|\phi_m\rangle\langle\phi_m|$, $E_m$ is not idempotent ($E_m^2 = \tfrac49|\phi_m\rangle\langle\phi_m|\ne E_m$), so $\langle\phi'|E_m|\phi'\rangle\ne1$ for the collapsed state in general — outcomes are not guaranteed to repeat. ∎
**E3 (medium).** Prove that any POVM element satisfies $0\preceq E_m\preceq I$ (i.e. its eigenvalues lie in $[0,1]$). Why does this make $p_m=\langle\psi|E_m|\psi\rangle$ a probability for *every* state?
Solution $E_m\succeq0$ by definition. For the upper bound, $I - E_m = \sum_{m'\ne m}E_{m'}\succeq0$ (a sum of positive operators is positive), so $E_m\preceq I$. Hence every eigenvalue $\mu$ of $E_m$ satisfies $0\le\mu\le1$. Then for any unit $|\psi\rangle$, $0\le\langle\psi|E_m|\psi\rangle\le\langle\psi|I|\psi\rangle=1$, so $p_m\in[0,1]$; and $\sum_m p_m=1$ by completeness. ∎
**E4 (medium).** Given a POVM $\{E_m\}$, show that infinitely many measurement-operator sets $\{M_m\}$ realize it, and characterize them.
Solution We need $M_m^\dagger M_m = E_m$. One solution is $M_m=\sqrt{E_m}$ (the positive square root). For any unitaries $U_m$, $M_m' := U_m\sqrt{E_m}$ also works: $M_m'^\dagger M_m' = \sqrt{E_m}U_m^\dagger U_m\sqrt{E_m} = \sqrt{E_m}\sqrt{E_m} = E_m$. Conversely, if $M_m^\dagger M_m=E_m$ then by polar decomposition $M_m = U_m\sqrt{E_m}$ for some unitary (or partial isometry) $U_m$. So the realizations are exactly $\{U_m\sqrt{E_m}\}$ — they share all outcome *probabilities* but differ in the post-measurement *state*. ∎ (This is why a POVM, not a set of $M_m$, is the right object when only statistics matter.)
**E5 (hard).** For unambiguous discrimination of two equiprobable states with real overlap $s=\langle\psi_0|\psi_1\rangle\in(0,1)$, prove that the inconclusive probability cannot be below $s$, so the IDP value $\Pr[?]_{\min}=s$ is optimal.
Solution A USD POVM $\{E_0,E_1,E_?\}$ requires $\langle\psi_1|E_0|\psi_1\rangle=0$ and $\langle\psi_0|E_1|\psi_0\rangle=0$; positivity then forces $E_0 = a|\psi_1^\perp\rangle\langle\psi_1^\perp|$ and $E_1 = b|\psi_0^\perp\rangle\langle\psi_0^\perp|$ (rank ≤ 1, supported on the orthogonal complements), with $a,b\ge0$. The success probabilities are $P_0 = a|\langle\psi_1^\perp|\psi_0\rangle|^2 = a(1-s^2)$ and likewise $P_1=b(1-s^2)$. The constraint $E_? = I-E_0-E_1\succeq0$ caps $a,b$: a $2\times2$ positivity computation (or Lagrange optimization with the equiprobable objective) gives $a=b=\tfrac1{1+s}$, so each conclusive probability is $\tfrac{1-s^2}{1+s}=1-s$ averaged appropriately, and $\Pr[?] = 1 - \tfrac12(P_0+P_1)\big|_{\text{avg}} = s$. Lower $\Pr[?]$ would require $a$ or $b$ beyond the positivity cap, breaking $E_?\succeq0$. Hence $\Pr[?]_{\min}=s$. ∎ (Full optimization: [NC, §2.2.6 / Box 2.5] and the IDP references.)
**E6 (hard).** Carry out the Naimark construction for the trine: build the isometry $V|\psi\rangle = \sum_{m=0}^2 (\sqrt{E_m}|\psi\rangle)\otimes|m\rangle$ into a qubit⊗qutrit space, verify $V^\dagger V=I$, and explain how a projective measurement of the qutrit ancilla reproduces the trine statistics.
Solution With $\sqrt{E_m} = \sqrt{2/3}\,|\phi_m\rangle\langle\phi_m|$, define $V:\mathbb C^2\to\mathbb C^2\otimes\mathbb C^3$, $V = \sum_m \sqrt{E_m}\otimes|m\rangle$ (where $|m\rangle$ are orthonormal qutrit kets). Then $V^\dagger V = \sum_{m,m'}\sqrt{E_m}\sqrt{E_{m'}}\langle m'|m\rangle = \sum_m E_m = I$, so $V$ is an isometry (extend to a unitary $U$ on the 6-dim space). Measuring the ancilla with $\Pi_m = I\otimes|m\rangle\langle m|$ gives $\|\Pi_m V|\psi\rangle\|^2 = \|\sqrt{E_m}|\psi\rangle\|^2 = \langle\psi|E_m|\psi\rangle$ — the trine probabilities. The verification (numerically: assemble $V$, check `V.conj().T @ V ≈ I_2`, and compare $\|\Pi_m V\psi\|^2$ to `povm_probabilities`) is a direct extension of the Hands-on code. This *is* how a POVM runs on hardware: dilate to a projective measurement on system+ancilla. ∎
--- ## Checkpoint 1. State the general measurement postulate: the operators, the completeness relation, the probability, and the post-measurement state. 2. Define a POVM and explain precisely what it determines and what it does not. 3. Show that projective measurement is the special case of a POVM, and name the extra structure projectors have. 4. What does Naimark's theorem assert, and why does it make "POVM" the most general measurement? 5. Why can a POVM but not a projective measurement give a qubit measurement with $3$ outcomes? 6. In unambiguous discrimination, what role does the third ("don't know") outcome play, and what is the minimum inconclusive probability for overlap $s$?
Answers 1. Operators $\{M_m\}$ with $\sum_m M_m^\dagger M_m=I$; outcome $m$ with $p_m=\langle\psi|M_m^\dagger M_m|\psi\rangle$; post-measurement state $M_m|\psi\rangle/\sqrt{p_m}$. 2. A POVM is $\{E_m\}$ with $E_m\succeq0$, $\sum_m E_m=I$. It determines all outcome *probabilities* $p_m=\langle\psi|E_m|\psi\rangle$ (or $\operatorname{Tr}(\rho E_m)$) but **not** the post-measurement state — that needs a choice of $M_m=U_m\sqrt{E_m}$. 3. Set $M_m=P_m$ (orthogonal projectors): $E_m=P_m^\dagger P_m=P_m$, giving $p_m=\langle\psi|P_m|\psi\rangle$ and projective collapse. The extra structure is orthogonality/idempotence: $E_jE_k=\delta_{jk}E_k$. 4. Every POVM is realizable as a projective measurement on system+ancilla (dilation), and conversely every ancilla-assisted projective measurement yields a POVM on the system — so POVMs are exactly the achievable measurements. 5. A qubit's projective measurements have at most $2$ (rank-1) orthogonal outcomes; POVM elements need not be orthogonal, so you can have $3+$ positive operators (e.g. the trine) summing to $I$. 6. The "don't know" outcome absorbs the ambiguity so the other two outcomes can be made orthogonal to the "wrong" state — enabling zero error. The minimal inconclusive probability is $\Pr[?]_{\min}=s=|\langle\psi_0|\psi_1\rangle|$.
--- ## Further Reading - **[NC]** Nielsen & Chuang, §2.2.3–2.2.6 — general measurements, POVMs, and the discrimination / distinguishability discussion (Box 2.5). - **[Pre]** Preskill, *Ph219*, Ch. 3 — generalized measurements, the Naimark/dilation picture, and the link to quantum channels. - **[Wat]** Watrous, *The Theory of Quantum Information*, Ch. 2–3 — measurements as a special case of channels, rigorous treatment of POVMs and Naimark's theorem. --- ← Prev: [Expectation & Uncertainty](02-expectation-uncertainty.md) · Up: [Term 1](../README.md) · Next: [Multi-Qubit States](../c04-entanglement/01-multi-qubit-states.md) →