Decisions with unequal prior probabilities
Suppose a source produces states on a finite-dimensional Hilbert space with prior probabilities and , where . A two-outcome measurement consists of effects and , where ; guess on the first outcome and on the second. With
the success probability is
Any finite-outcome POVM followed by a possibly randomized decision rule can be coarse-grained into two effects with the same success probability, so it suffices to optimize over a single effect . The optimum is expressed through the trace norm.
Definition 1.10.1
(Trace norm)
For a square matrix , the trace norm is . If is Hermitian with eigenvalues , then .
Theorem 1.10.2
(Optimal success with unequal priors)
For states with priors and , the largest success probability over all binary decision rules is
It is attained by the projector onto the positive eigenspace of .
Proof
(Maximizing over effects by diagonalizing the difference)
Diagonalize the Hermitian . Then with , and the condition forces . The sum is therefore at most the sum of the positive eigenvalues, with equality for the projector onto the positive spectral subspace, which is a valid effect. Since , the positive and negative parts satisfy and , so . Adding gives the stated formula.
Proof X-ray
Why the spectral measurement is optimal
The proof uses two hypotheses in two different places. Hermiticity of is what allows a diagonal basis, and is what confines each to . The one non-obvious move is rewriting the optimization over a matrix as an optimization over numbers : for each positive eigenvalue, setting increases the objective; for each negative eigenvalue, setting avoids a loss. A single spectral projector makes all these choices simultaneously and is a valid effect, which is why the bound is attained and the argument closes. A zero eigenspace may be assigned to either decision. If were allowed to leave the interval , the objective would be unbounded, so the effect condition is exactly what makes the optimum finite.
The result includes and . When the states are identical it gives , achieved by always guessing the more likely state. The decision threshold depends on the priors, so the equal-prior measurement need not remain optimal when the priors change.
Common mistake
(Using the equal-prior distance for unequal priors)
It is tempting to write for every prior. For and this gives , yet always guessing succeeds with probability . The prior weights must sit inside the norm: the correct quantity is .
Trace distance is an operational difference
Definition 1.10.3
(Trace distance)
For density matrices the trace distance is
Proposition 1.10.4
(Variational formula for trace distance)
For density matrices ,
Consequently , and with equal priors the optimal success probability is .
Proof
(Equal positive and negative parts)
The Hermitian difference has trace zero, so its positive and negative parts have equal trace, and each equals . The maximum of is the trace of the positive part, by the diagonalization argument above, which gives the formula. Taking shows . For the upper bound let be the positive spectral projector; then . The equal-prior value follows by setting in the previous theorem: and .
Equivalently, is the largest difference in probability that a single measurement event can see between the two states. Orthogonal pure states have distance one, since then has eigenvalues ; identical states have distance zero.
Concept lensGeometric
Trace distance on the Bloch ball
For a qubit write and with Bloch vectors , where is the vector of Pauli matrices. Then , whose eigenvalues are . Hence , half the Euclidean distance between the two points of the ball. The same quantity that was a maximum over measurement events is, for one qubit, a length.
Worked example 1.10.5
(Optimal equal-prior discrimination of and )
Take and , where . In the computational basis
which has trace and determinant , so its eigenvalues are . Thus and the optimal equal-prior success probability is . A computational-basis measurement succeeds with probability only under the natural decision rule, so an achievable measurement need not be optimal. The Bloch vectors are and , at distance , which agrees with .
Worked example 1.10.6
(The same pair with prior )
Keep and but let have prior . Then
The trace is as required, and the determinant is . The characteristic equation gives , so . Therefore
which exceeds the trivial value . The optimal measurement projects onto the eigenvector of the positive eigenvalue , not onto the equal-prior eigenvector.
A channel cannot make states more distinguishable
Definition 1.10.7
(Channel and adjoint)
A finite-dimensional quantum channel is a completely positive trace-preserving linear map , possibly between different spaces. Its adjoint maps output operators to input operators and is defined by the trace pairing
For Hermitian observables the displayed trace pairing agrees with the usual Hilbert–Schmidt adjoint.
Theorem 1.10.8
(Data processing for trace distance)
For every quantum channel and density matrices ,
More generally, for any priors the best success probability for discriminating does not exceed that for .
Proof
(Pulling output events back to input events)
Trace preservation gives . If then for every , so . Hence if , both and are positive: is itself an effect. Pulling every output event back to an input event and applying the variational formula gives
For unequal priors the difference need not have trace zero, but the same pullback bounds by , and .
This is the data-processing inequality for trace distance; the proof in fact requires only positivity and trace preservation, though a quantum channel is completely positive. A measurement on a channel's outputs could have been implemented as an effective measurement on its inputs; the channel cannot create a new binary distinction. Thus a channel cannot improve the best success probability for a fixed prior distribution. This statement concerns one channel acting without knowing which state was supplied; an operation given the correct label in advance is a different scenario. Equality is possible, for example under a unitary channel, and strict loss is also possible.
Worked example 1.10.9
(Complete dephasing halves the distance)
Let with ; it is completely positive and trace preserving. It fixes , because . It sends to . The output difference is , so , compared with before the channel. The inequality is strict, and the Bloch picture shows why: the channel projects the ball onto the -axis, shortening the distance between and to that between and the origin.
Counterexample mode
A channel need not preserve distinguishability
The reset map is a valid trace-preserving completely positive channel, but it sends both orthogonal input states and to . Their input trace distance is one and output distance is zero. Data processing is an inequality; it does not say all physical processing is reversible.
Common mistake
(Reading equal distance as reversibility)
If for one pair of states, that does not make invertible. The dephasing channel fixes and , so the distance between them survives, yet it sends both and to and cannot be undone. Data processing bounds the distance for the specific pair; it says nothing about other inputs.
Practice and solutions
Quick check
(If and , what is the optimal success probability?)
Use the identical-states case of the optimal-success theorem.
Solution · Answer
: always guess . Indeed is positive, so and .
Quick check
(Find for orthogonal pure states and for identical states.)
Compute the eigenvalues of in each case.
Solution · Answer
The values are and , respectively. For orthogonal pure states has eigenvalues , so ; for identical states the difference is zero.
Quick check
(What happens to trace distance under a unitary channel ?)
Compare the eigenvalues of the Hermitian difference before and after conjugation.
Solution · Answer
It is unchanged: unitary conjugation preserves the eigenvalues of the Hermitian difference, hence its trace norm.
Quick check
(Can a reset channel increase the best equal-prior success probability for any pair of inputs?)
Compare the output success probability with what random guessing achieves before the reset.
Solution · Answer
No. Both outputs are identical, so the output success probability is ; it was at least before the reset.