Entangled States and Maximal Product Overlap

A pure bipartite vector is entangled when it is not a simple tensor for the declared subsystem split. Its largest Schmidt coefficient gives a quantitative version of that statement: it determines the best product overlap and constructs a bounded test separating the state from every separable mixture. The canonical Entangled States article supplies the physical introduction; here the Hilbert spaces may be infinite-dimensional.

Helpful background. Schmidt Decomposition in Hilbert Space supplies the countable expansion. Fidelity and Trace Distance supplies operational distance formulas. Separable and Entangled Mixed States supplies trace-norm closure for separable mixtures.

The largest Schmidt coefficient is the best product overlap

Let $\Psi=\sum_ns_nu_n\otimes v_n$ be normalized, with $s_1\geq s_2\geq\cdots>0$. For unit vectors $a,b$,

$$|\langle a\otimes b,\Psi\rangle| =|\langle a,T_\Psi\bar b\rangle| \leq\|T_\Psi\|=s_1,$$

where $T_\Psi$ is the compact operator in the Schmidt proof. The bound is attained by $a=u_1,b=v_1$. Therefore

$$\sup_{\|a\|=\|b\|=1} |\langle a\otimes b,\Psi\rangle|^2=s_1^2.$$

Compactness matters: it supplies an actual largest singular value and maximizing product vectors, rather than merely a formal supremum over an uncontrolled infinite list.

Normalization implies $s_1=1$ exactly when all other coefficients vanish, that is, exactly when $\Psi$ is product. Every entangled pure vector has $s_1<1$, even if its Schmidt rank is infinite.

Optimizing the irrelevant overall product phase gives the nearest unit product-vector distance

$$\inf_{a,b,\theta}\|\Psi-e^{i\theta}a\otimes b\| =\sqrt{2(1-s_1)}.$$

For the corresponding pure-state density operators, the minimum trace distance to a pure product is $\sqrt{1-s_1^2}$. These are different metrics on different representations of a state.

A bounded witness excludes all separable mixtures

Define the bounded self-adjoint operator

$$W=s_1^2I-|\Psi\rangle\langle\Psi|.$$

The product-overlap bound gives nonnegative expectation of $W$ in every pure product state. By convexity it is nonnegative on finite mixtures, and by trace-norm continuity of bounded expectations it remains nonnegative on their closed convex hull, the full separable set.

On the target state, however,

$$\langle\Psi,W\Psi\rangle=s_1^2-1<0.$$

Thus this is an entanglement witness for any entangled pure vector. The proof remains valid for infinite Schmidt rank, because $W$ is bounded and no unbounded moment is used.

The effect $|\Psi\rangle\langle\Psi|$ also yields a quantitative separation. For every separable $\sigma$,

$$D(|\Psi\rangle\langle\Psi|,\sigma) \geq1-s_1^2, \qquad D(\rho,\sigma)=\tfrac12\|\rho-\sigma\|_1.$$

Indeed the two states’ probabilities for that effect differ by at least $1-s_1^2$. This lower bound is not a claim that the nearest separable state is pure or that the bound is always optimal.

If an approximate preparation $\rho$ has $D(\rho,|\Psi\rangle\langle\Psi|)<1-s_1^2$, its fidelity with the target projector remains above the separable threshold $s_1^2$. The same bounded test certifies entanglement despite that preparation error.

A geometric pair illustrates robust infinite rank

For the normalized state

$$\Psi_q=\sqrt{1-q}\sum_{n=0}^\infty q^{n/2}|n,n\rangle, \qquad 0<q<1,$$

the largest squared coefficient is $1-q$. It has infinite Schmidt rank but is separated from every separable state by trace distance at least $q$. The closest pure product has distance $\sqrt q$ and is $|0,0\rangle$.

These two bounds leave room for mixed separable approximants; they should not be identified. As $q\downarrow0$, the state approaches a product even though its exact Schmidt rank remains infinite for every positive $q$. Rank and distance capture different aspects of entanglement.

The witness need not be easy to measure in a laboratory: its projector can be a genuinely joint observable. Mathematical existence of a certificate does not itself supply an experimentally efficient protocol.

Local filtering can increase conditional entanglement

A selected local operation changes the normalized branch state and its success probability together. Choose two nonzero Schmidt coefficients $s_1\geq s_2$ and the bounded contraction on $A$

$$F=\frac{s_2}{s_1}|u_1\rangle\langle u_1| +|u_2\rangle\langle u_2|,$$

with zero action on the orthogonal complement. Then

$$(F\otimes I)\Psi =s_2(u_1\otimes v_1+u_2\otimes v_2).$$

The successful normalized branch is an equal two-mode entangled pair, with success probability $2s_2^2$. Completing $F$ with the failure operator $\sqrt{I-F^*F}$ makes a valid local two-outcome operation.

This does not imply deterministic entanglement gain under local processing. Discarded failures and branch probabilities are part of the operation; a conditional state alone omits them. For the geometric example, the displayed success probability is $2q(1-q)$.

Exercises

1. An entangled pure state has largest reduced-state eigenvalue $0.9$. Give its best pure-product trace distance and a guaranteed distance from every separable state.

Solution

Here $s_1^2=0.9$. The best pure-product distance is $\sqrt{0.1}$, while every separable state is at distance at least $0.1$. The second statement applies to mixtures as well and is a lower bound, not generally an equality.

2. Why does a global change of tensor-factor identification require reevaluating these criteria?

Solution

The optimization is over products for a specified pair of subsystem algebras. A general global unitary need not preserve that product set. Local unitaries preserve Schmidt coefficients and all the displayed criteria, whereas changing the subsystem structure changes the question being asked.

References

• Holevo, Alexander S., Quantum Systems, Channels, Information: A Mathematical Introduction, second revised and expanded edition, De Gruyter, 2019 — bipartite Hilbert-space states and local operations. Publisher.
• Watrous, John, The Theory of Quantum Information, Cambridge University Press, 2018 — Schmidt coefficients, distinguishability, separability, and local transformations. Author’s text.