Informational Completeness

A POVM is informationally complete for normal states if its entire outcome distribution uniquely determines the density operator. In finite dimension this is a spanning condition on Hermitian matrices. In infinite dimension it becomes a separation condition against trace-class operators, and uniqueness of reconstruction need not provide a stable numerical inverse.

Helpful background. POVMs supplies the probability map; trace duality supplies its natural operator pairing. Experimental calibration belongs to Measurement Tomography.

Separating density operators

A POVM $E$ is informationally complete if

$$\operatorname{Tr}(\rho E(\Delta)) =\operatorname{Tr}(\sigma E(\Delta)) \quad\text{for every }\Delta \quad\Longrightarrow\quad \rho=\sigma.$$

The distributions are understood exactly in this definition; a finite dataset estimates them with error. Completeness concerns identifiability of the model, not the number of samples needed or robustness of an estimator.

An equivalent condition is that the only self-adjoint trace-class operator $T$ satisfying

$$\operatorname{Tr}(TE(\Delta))=0 \qquad\text{for every }\Delta$$

is $T=0$. One direction follows by taking $T=\rho-\sigma$. For the converse, normalization implies $\operatorname{Tr}T=0$. If $T\ne0$, its Jordan parts have equal positive trace $c$, so $\rho=T_+/c$ and $\sigma=T_-/c$ are distinct states with the same probabilities.

By dual separation, this is equivalent to the real linear span of the effects being ultraweakly dense in the self-adjoint part of $\mathcal B(\mathcal H)$. Operator-norm density is not required. The topology identifies the class of states being separated: here the predual consists of normal trace-class states.

Finite dimension and the number of outcomes

For $\dim\mathcal H=d<\infty$, Hermitian matrices form a real vector space of dimension $d^2$. A finite POVM $(E_j)_{j=1}^m$ is informationally complete exactly when the effects span that space. Necessarily $m\geq d^2$.

The same count follows from parameters: a density matrix has $d^2-1$ independent real affine parameters, while $m$ probabilities subject to one normalization constraint supply at most $m-1$ independent numbers. Positivity restricts the allowed region but does not reduce its affine dimension.

For a qubit, take unit vectors $n_j$ at the vertices of a regular tetrahedron and

$$E_j=\tfrac14(I+n_j\cdot\sigma), \qquad j=1,\ldots,4.$$

They obey $\sum_jn_j=0$ and $\sum_jn_jn_j^{\mathsf T}=\tfrac43I_3$. For $\rho=(I+r\cdot\sigma)/2$,

$$p_j=\tfrac14(1+n_j\cdot r), \qquad r=3\sum_jp_jn_j.$$

The reconstruction is explicit. It also illustrates why outcome frequencies substituted into an exact inverse can produce $|r|>1$: sampling errors need not remain in the physical probability region. A statistical estimator must handle positivity and uncertainty separately.

A continuous sharp observable can be incomplete

Position measurement on $L^2(\mathbb R)$ gives the distribution $|\psi(x)|^2dx$. Replacing $\psi(x)$ by $e^{ikx}\psi(x)$ leaves every position probability unchanged, while generally changing the state ray and momentum statistics.

Thus infinitely many outcomes do not by themselves imply informational completeness. What matters is the span of the effects, not the size of the outcome space. A single sharp observable records a commuting family of event projections; it generally misses coherences relative to that observable.

Likewise, the qubit trine POVM has three outcomes but probes only two Bloch coordinates. Nonprojective effects can supply additional information, but being a POVM is not itself an informational-completeness certificate.

A countable infinite-dimensional construction

Choose a countable basis $(e_n)$. Enumerate rank-one projectors onto the vectors

$$e_n,\qquad \frac{e_n+e_m}{\sqrt2},\qquad \frac{e_n+i e_m}{\sqrt2} \quad(n<m)$$

as $(P_j)$. Their expectations determine all diagonal, real off-diagonal, and imaginary off-diagonal matrix entries.

Choose positive weights $w_j$ with $\sum_jw_j\leq1$, and define

$$E_j=w_jP_j,\qquad E_0=I-\sum_jw_jP_j.$$

The series converges in operator norm because $\|P_j\|=1$, and its sum is at most $I$. This is a countable POVM. Since every $w_j>0$, exact probabilities recover the expectation of every $P_j$, hence every matrix element of $\rho$. The measurement is informationally complete.

However, very small weights make some reconstructed expectations sensitive to probability errors by a factor $1/w_j$. In this construction, increasingly high basis states can put almost all probability into the residual outcome $0$, while remaining mutually orthogonal states. Uniqueness is therefore compatible with poor uniform stability. Energy or regularity constraints can make practical reconstruction meaningful by controlling the relevant tail.

Exercises

1. Why can no finite-outcome POVM be informationally complete on an infinite-dimensional Hilbert space?

Solution

Restrict to a finite-dimensional subspace of dimension $d$ with $d^2$ greater than the number of outcomes. The compressed effects cannot span its Hermitian matrices. Two states supported in that subspace therefore have identical probabilities, already violating completeness on the whole space.

2. For the tetrahedral measurement, find the probabilities for the pure state with Bloch vector $n_1$.

Solution

Because $n_1\cdot n_j=-1/3$ for $j\ne1$, the probabilities are $p_1=1/2$ and $p_2=p_3=p_4=1/6$. Substitution into the reconstruction formula returns $r=n_1$, furnishing a calibration check for measurement tomography.

References

• Busch, Paul, Pekka Lahti, Juha-Pekka Pellonpää, and Kari Ylinen, Quantum Measurement, Springer, 2016 — informational completeness and operator measures. Publisher.
• Heinosaari, Teiko, and Mário Ziman, The Mathematical Language of Quantum Theory: From Uncertainty to Entanglement, Cambridge University Press, 2012 — informationally complete observables and finite-dimensional state spaces. Publisher.