Factorizability in Optimal LOCC Discrimination

• The paper establishes that even when optimal measurement operators factorize into product forms, they may not be implementable by finite-round LOCC due to geometric and operational constraints.
• It demonstrates that in restricted settings, such as two-qubit unitary discrimination, a product input and LOCC measurement can achieve global optimality, highlighting measure-dependent factorizability.
• The analysis connects separable measurements, LOCC protocol trees, and diverse discrimination criteria, underlining that optimal discrimination factorization is multifaceted and protocol-dependent.

Factorizability of optimal LOCC discrimination concerns the relation between two distinct structures in multipartite quantum discrimination: the algebraic factorization of measurement operators into product or separable forms, and the operational realizability of the same measurement by local operations and classical communication (LOCC). In the modern literature, these notions are sharply separated. The canonical negative result is that an optimal discrimination measurement may factorize completely at the level of POVM elements and still fail to be implementable by finite-round LOCC (Cohen, 2014). At the same time, there are restricted settings in which the globally optimal strategy does admit a fully local realization, sometimes with a product input and an LOCC output measurement (Li et al., 2017), and there are also settings where factorizability depends on the discrimination criterion itself (Murshid et al., 25 Jun 2025). The subject therefore lies at the intersection of separable measurements, LOCC protocol geometry, state- and channel-discrimination theory, and the resource theory of entanglement.

1. Conceptual meaning of factorization in local discrimination

In this literature, “factorizable” most often refers to the structure of the measurement operators rather than to the existence of a sequential local protocol. For a multipartite Hilbert space

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$

a POVM element is product if it has the form $K^{(1)}\otimes\cdots\otimes K^{(P)}$, and separable if it is a sum or convex combination of such product positive operators. LOCC measurements form a strictly smaller class: they must arise from an adaptive protocol of local operations and classical communication. The fundamental inclusion is

$$\text{LOCC} \subsetneq \text{separable measurements},$$

and the gap persists in operational discrimination tasks (Cohen, 2014).

A central structural lesson is that factorization of operators is weaker than factorization of protocols. The double-trine minimum-error problem makes this distinction explicit: the paper proves

$$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$

so a separable measurement can attain the global optimum while still failing to be LOCC-implementable, and within LOCC itself a two-way adaptive protocol can outperform one-way local sequencing (Chitambar et al., 2013). This suggests that “factorizability of optimal LOCC discrimination” cannot be treated as a purely operator-theoretic question. It is also a question about protocol trees, communication rounds, and geometric constraints on the local cones generated by measurement components (Cohen, 2014).

For linearly independent pure-state ensembles under minimum-error discrimination, another recurring structural theme is that optimality can be reduced to perfect discrimination of a unique orthonormal basis of the span (Chitambar et al., 2013, Chitambar et al., 2013). In that setting, LOCC-optimality becomes equivalent to whether the induced optimal detection basis is locally perfectly distinguishable. This recasts factorization as a property of the optimal detection subspaces, not merely of the signal states.

2. Separable but not LOCC: the archetypal negative result

The cleanest negative answer is given by the multipartite unambiguous discrimination construction of Cohen (Cohen, 2014). The setting is a multipartite Hilbert space

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$

with local dimensions $d_\alpha=\dim\mathcal H_\alpha$, total dimension

$D=d_1d_2\cdots d_P,$

and a prime $N\ge 5$ chosen so that

$D=N-1.$

The construction defines $N$ product states

$K^{(1)}\otimes\cdots\otimes K^{(P)}$0

with local factors

$K^{(1)}\otimes\cdots\otimes K^{(P)}$1

where

$K^{(1)}\otimes\cdots\otimes K^{(P)}$2

Their rank-1 product projectors $K^{(1)}\otimes\cdots\otimes K^{(P)}$3 satisfy

$K^{(1)}\otimes\cdots\otimes K^{(P)}$4

From any $K^{(1)}\otimes\cdots\otimes K^{(P)}$5 of these $K^{(1)}\otimes\cdots\otimes K^{(P)}$6 states, specifically

$K^{(1)}\otimes\cdots\otimes K^{(P)}$7

one forms the reciprocal set

$K^{(1)}\otimes\cdots\otimes K^{(P)}$8

defined by

$K^{(1)}\otimes\cdots\otimes K^{(P)}$9

The discrimination task is unambiguous discrimination of $$\text{LOCC} \subsetneq \text{separable measurements},$$0 with equal priors

$$\text{LOCC} \subsetneq \text{separable measurements},$$1

A structural corollary in the paper shows that the only positive operators on $$\text{LOCC} \subsetneq \text{separable measurements},$$2 that can unambiguously identify $$\text{LOCC} \subsetneq \text{separable measurements},$$3 are proportional to $$\text{LOCC} \subsetneq \text{separable measurements},$$4. Thus every conclusive measurement operator is forced to be a product operator (Cohen, 2014).

The main theorem states that the optimal global measurement is separable, unique on $$\text{LOCC} \subsetneq \text{separable measurements},$$5, given by

$$\text{LOCC} \subsetneq \text{separable measurements},$$6

and has failure probability

$$\text{LOCC} \subsetneq \text{separable measurements},$$7

Equivalently, the success probability is

$$\text{LOCC} \subsetneq \text{separable measurements},$$8

Because the discrimination task only concerns $$\text{LOCC} \subsetneq \text{separable measurements},$$9, the element $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$0 is the inconclusive outcome, so the optimal POVM may be written as

$$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$1

Every element is rank-1 and product. In the strongest possible POVM-level sense, the optimal measurement factorizes completely (Cohen, 2014).

Yet the same paper proves that this optimal separable measurement cannot be implemented by any finite-round LOCC protocol. The obstruction comes from a necessary condition for finite-round LOCC: if a separable measurement with $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$2 distinct product POVM elements

$$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$3

is implementable by finite-round LOCC, then

$$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$4

where $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$5 counts the distinct extreme rays in the convex cone generated by the local factors $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$6. For the optimal measurement above, each local factor is rank-1 positive and distinct, so

$$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$7

Since $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$8 and $$\mathrm{GLOBAL}=\mathrm{SEP}>\mathrm{LOCC}>\mathrm{LOCC}_\rightarrow,$$9,

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$0

so the finite-round LOCC condition is violated (Cohen, 2014).

This result directly answers the central question in the negative: optimality plus factorization of the POVM elements does not imply LOCC realizability. The paper further shows that enlarging the Hilbert space does not help. If the input states are supported only on $$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$1, then any LOCC measurement on a larger space $$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$2 is effectively identical, on those inputs, to an LOCC measurement on $$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$3 itself (Cohen, 2014). A plausible implication is that the obstruction is genuinely operational and geometric, not an artifact of representation or ancilla omission.

3. Minimum-error discrimination and the rigidity of optimal measurements

Minimum-error discrimination exhibits the same separation in a different form. For the two-qubit double-trine ensemble

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$4

with priors $$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$5, the global optimum is the Pretty Good Measurement on an entangled orthonormal basis

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$6

and has error probability

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$7

equivalently

$$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$8

(Chitambar et al., 2013). Thus the globally optimal measurement is not factorized at the level of basis states.

The same paper proves a theorem of broader structural significance: for any linearly independent pure-state ensemble, there exists a unique orthonormal basis $$\mathcal H=\mathcal H_1\otimes\cdots\otimes \mathcal H_P,$$9 of the span such that a POVM attains the optimal minimum error if and only if it perfectly distinguishes $d_\alpha=\dim\mathcal H_\alpha$0 (Chitambar et al., 2013). In the double-trine case, this basis is exactly the entangled basis $d_\alpha=\dim\mathcal H_\alpha$1. Therefore any LOCC protocol attaining the global optimum would have to perfectly, or asymptotically perfectly, distinguish these entangled basis states. Using Walgate–Hardy and asymptotic LOCC constraints, the paper rules this out, establishing that even asymptotic LOCC cannot reach the global minimum-error optimum (Chitambar et al., 2013).

At the same time, the global optimum can be reproduced by a separable measurement. The paper constructs POVM elements

$d_\alpha=\dim\mathcal H_\alpha$2

each of which is separable and expressible as a convex combination of rank-one product projectors. Since the double-trine states are orthogonal to the singlet, the added term does not alter the ensemble statistics, and therefore

$d_\alpha=\dim\mathcal H_\alpha$3

(Chitambar et al., 2013). This is a second archetypal example in which the optimal statistics are attainable by a separable, product-structured POVM, yet not by LOCC.

The same paper quantifies protocol-level non-factorizability inside LOCC. The optimal one-way LOCC error is

$d_\alpha=\dim\mathcal H_\alpha$4

while an explicit two-way adaptive protocol achieves about

$d_\alpha=\dim\mathcal H_\alpha$5

Thus even within LOCC, a more “factorized” one-pass structure is suboptimal relative to multiround adaptation (Chitambar et al., 2013).

A related rigidity appears in two-qubit minimum-error state discrimination more broadly. For linearly independent ensembles, optimality is equivalent to perfect discrimination of a unique orthogonal decomposition of the span (Chitambar et al., 2013). In particular, for linearly independent two-qubit pure states, if $d_\alpha=\dim\mathcal H_\alpha$6, LOCC can be optimal only if the optimal orthonormal detection basis contains at least two product states; if $d_\alpha=\dim\mathcal H_\alpha$7, all optimal detection states must be product (Chitambar et al., 2013). The same paper proves that three randomly chosen two-qubit pure states almost surely cannot be optimally discriminated by LOCC (Chitambar et al., 2013). This suggests that exact LOCC factorization of the optimal measurement is non-generic.

4. Positive factorizability results in restricted settings

Despite these negative results, there are important regimes in which optimal discrimination does factorize operationally. A particularly sharp example is single-query discrimination of two two-qubit entangling unitaries without local parts (Li et al., 2017). The problem is reduced to state discrimination after one use of the unknown unitary, and the main optimization is over the input state

$d_\alpha=\dim\mathcal H_\alpha$8

For two such unitaries $d_\alpha=\dim\mathcal H_\alpha$9, the minimum-error quantity is

$D=d_1d_2\cdots d_P,$0

with

$D=d_1d_2\cdots d_P,$1

The paper proves that the global optimum is always achieved by a product input state, and then invokes known LOCC results for discrimination of two pure output states to conclude that LOCC attains the same distinguishability as global operations (Li et al., 2017). Within this scope, factorization holds at both layers: product probe state and LOCC output measurement.

An earlier related result proves the same conclusion for a restricted class of two-qubit entangling unitaries in the absence of local parts: if global perfect discrimination is possible, then LOCC perfect discrimination is also possible, while if global perfect discrimination is impossible, LOCC still attains the same optimal minimum-error probability (Bae, 2016). The proof again rests on product-state attainability of the optimal overlap and LOCC-optimal discrimination of the resulting pure-state pair.

There are also positive results in state discrimination under strong structural assumptions. For two bipartite mixed states of the form

$D=d_1d_2\cdots d_P,$2

with orthogonality and nonoverlapping-support conditions separating the $D=d_1d_2\cdots d_P,$3- and $D=d_1d_2\cdots d_P,$4-branches, the optimal global unambiguous discrimination success probability is exactly achieved by an explicit finite-round LOCC protocol (Zhang et al., 2020). The mechanism is an orthogonal-sector decomposition into two pure-state discrimination tasks. The paper does not prove that the optimal global POVM itself factorizes as a single product measurement, but it does prove operational attainability by adaptive LOCC (Zhang et al., 2020).

Many-copy hypothesis testing also contains exact factorizability regimes. For certain orthogonal bipartite state pairs, including a maximally entangled state versus its orthogonal complement and extremal Werner states, a sufficient condition implies

$D=d_1d_2\cdots d_P,$5

and

$D=d_1d_2\cdots d_P,$6

so the optimal many-copy test is realized by tensor powers of a single-copy LOCC measurement (Cheng et al., 2020). This is an exact multiplicative factorization across copies, but only under a specific one-copy LOCC/PPT compatibility condition.

5. Measure dependence, protocol width, and sequence-level factorization

A major development is the realization that factorization can depend on the discrimination figure of merit. The 2025 paper on “quantum nonlocality without entanglement and state discrimination measures” constructs a family of six equally probable linearly independent product states in $D=d_1d_2\cdots d_P,$7 for which

$D=d_1d_2\cdots d_P,$8

(Murshid et al., 25 Jun 2025). In minimum-error discrimination, the globally optimal measurement is the square-root measurement on an entangled orthonormal basis and cannot be implemented by LOCC (Murshid et al., 25 Jun 2025). In unambiguous discrimination, however, an explicit sequential local protocol attains the global optimum

$D=d_1d_2\cdots d_P,$9

(Murshid et al., 25 Jun 2025). This shows that optimal local factorizability is not an intrinsic property of the ensemble alone; it depends on the performance criterion.

For multipartite quantum sequences, the notion of factorization is made explicit at the level of success probabilities. If

$N\ge 5$0

are multipartite ensembles and

$N\ge 5$1

is the corresponding sequence ensemble, the paper defines optimal LOCC discrimination to be factorizable when

$N\ge 5$2

(Ha et al., 7 Aug 2025). It proves this equality under several exact conditions, including the case where the optimal strategy on the full sequence is merely to guess one most probable sequence label, characterized by the block-positivity condition

$N\ge 5$3

(Ha et al., 7 Aug 2025). It also gives counterexamples where

$N\ge 5$4

so sequence-level LOCC optimization can exploit genuinely global structure across positions (Ha et al., 7 Aug 2025). A plausible implication is that “factorizability of optimal LOCC discrimination” itself has layers: POVM-element factorization, protocol factorization, and temporal or sequence factorization are distinct questions.

The protocol side of the subject also has its own normal-form results. Any LOCC protocol for discriminating a finite multipartite ensemble can be replaced, without loss in success probability, by one in which every local measurement on a $N\ge 5$5-dimensional system has at most $N\ge 5$6 outcomes (Leung et al., 2019). More generally, any fine-grained LOCC protocol can be decomposed into a convex combination of “slim protocols” with that bounded branching width, and for convex objectives such as discrimination success probability, one such slim protocol is itself optimal (Leung et al., 2019). This does not imply product factorization of the final POVM, but it does show that optimal LOCC discrimination admits a bounded-width deterministic protocol normal form.

6. Resources, limitations, and open structural boundaries

Several results show that even when signal states are weakly entangled or separable, optimal discrimination may still fail to factorize into LOCC without additional resources. For the uniform noisy Bell ensemble

$N\ge 5$7

with priors $N\ge 5$8, the exact local and global success probabilities are

$N\ge 5$9

so unassisted LOCC is strictly suboptimal for every $D=N-1.$0 (Bandyopadhyay et al., 2021). With a two-qubit pure resource state

$D=N-1.$1

the assisted LOCC optimum becomes

$D=N-1.$2

which equals the global optimum if and only if $D=N-1.$3, that is, if and only if the resource is a Bell state (Bandyopadhyay et al., 2021). The paper concludes that the entanglement cost of optimal LOCC discrimination is exactly 1 ebit. In the white-noise example

$D=N-1.$4

all four states are separable for $D=N-1.$5, yet the entanglement cost remains 1 ebit (Bandyopadhyay et al., 2021). This shows that factorization properties of the signal ensemble do not control factorization properties of the optimal local discrimination protocol.

Channel discrimination under restricted measurements displays the same non-uniformity. There exist channel pairs for which

$D=N-1.$6

so entangled inputs remain fully useful despite an LOCC restriction on the final measurement (Matthews et al., 2010). There are also channel pairs for which

$D=N-1.$7

so the LOCC restriction erases all advantage of entangled inputs (Matthews et al., 2010). This again suggests that no universal factorization principle governs optimal LOCC discrimination.

The finite- versus infinite-round boundary also remains important. In the multipartite unambiguous discrimination example of Cohen, the impossibility result excludes all finite-round LOCC protocols but does not rule out infinite-round LOCC, which is conjectured to fail as well (Cohen, 2014). In the $D=N-1.$8-copy trine problem, no finite $D=N-1.$9 allows exact optimal LOCC discrimination, yet LOCC becomes asymptotically optimal as $N$0 (Chitambar et al., 2013). These cases indicate that exact factorizability can fail at every finite stage while re-emerging only as an asymptotic approximation.

Taken together, these results define the modern understanding of the subject. Factorizability of optimal LOCC discrimination is not a single property but a family of non-equivalent properties concerning product inputs, separable POVM elements, LOCC protocol trees, copywise tensorization, and sequence-level decomposition. The strongest negative statement established so far is that even a unique optimal measurement built entirely from rank-1 product operators can fail to be finite-round LOCC (Cohen, 2014). The strongest positive statements are correspondingly restricted: some unitary, mixed-state, many-copy, and sequence settings do admit exact factorization or exact LOCC/global equality (Li et al., 2017, Zhang et al., 2020, Cheng et al., 2020, Ha et al., 7 Aug 2025). A plausible overall synthesis is that factorization is best understood as a problem in geometry and protocol complexity rather than in tensor-product algebra alone.