Quantum information · Theory (AI-assisted)
Probabilistic distillation of superposition
arXiv:2611.16719v1 ↗PDF ↗Submitted 12 November 2026
- Question
- When the reference states of a superposition are not orthogonal, as for the coherent states \(\ket{\pm\alpha}\) of a cat state, the components interfere, and states with coefficients of equal modulus need not be equally useful. In the resource theory of superposition, the operations that cannot create superposition map each reference state to a multiple of a reference state; for coherent states they include photon loss, displacement and normalised photon subtraction. The optimal probability of converting one superposition into another was known only as a semidefinite program. What is it in closed form, and which superpositions are the better resource?
- Result
- The probability is bounded above by \(\tau(\psi)/\tau(\phi)\), a ratio of normalised squared coefficient norms that measure how destructively the components interfere, and below by the probability of a route through a maximally coherent state. The upper bound is exact when the source's component Gram matrix is permutation-invariant, and the optimum is a generalised eigenvalue when the target's is. Consequently an odd cat state becomes an even one with certainty, by normalised photon subtraction, while the reverse succeeds with probability at most \(\tanh^2\lvert\alpha\rvert^2\). Collective operations help: two even cats become two odd cats with probability \(\tanh^3\lvert\alpha\rvert^2\) instead of \(\tanh^4\lvert\alpha\rvert^2\). Distilling coherence gains nothing from copies. For two reference states every conversion is solved; for three or more with equal positive overlap no state reaches all others with certainty.
The paper in detail
Write a pure state as \(\ket\psi=\sum_jx_j\ket{c_j}\) over linearly independent reference states with Gram matrix \(S\). Its components \(x_j\ket{c_j}\) have their own Gram matrix \(G_\psi=X^\dagger SX\), with \(X=\mathrm{diag}(x)\), whose entries sum to 1. The off-diagonal entries are interference terms, and \(\|x\|^2-1\) is minus their total. The normalised quantity \(\tau(\psi)=(d\|x\|^2-1)/(d-1)\) exceeds 1 when the components interfere destructively and is below 1 when they interfere constructively.
Every successful branch of a free operation must undo the source's coefficients, permute the reference states and apply the target's coefficients, so the optimal probability is a semidefinite program over permutation weights: permuted copies of \(G_\phi\), weighted by branch probabilities, must fit under \(G_\psi\). In coherence theory all these matrices are diagonal and the program is linear. With overlaps, the off-diagonal entries carry phases, and different permutations test different directions.
The central result is an interference principle. Testing the constraint against the projector orthogonal to the all-ones vector gives \(P^\star\,\tau(\phi)\le\tau(\psi)\): free operations cannot make the components interfere more destructively with certainty. For the coherent states \(\ket{\pm\alpha}\), \(s=e^{-2\lvert\alpha\rvert^2}\), \(\tau(\mathrm{odd})=\coth\lvert\alpha\rvert^2\) and \(\tau(\mathrm{even})=\tanh\lvert\alpha\rvert^2\), which are the cat states' mean photon numbers in units of \(\lvert\alpha\rvert^2\). The bound is exact for these sources, so every optimal conversion among odd and even cats of any amplitudes is a product or ratio of hyperbolic tangents.
With three or more equal positive overlaps the interference is frustrated, since three phases cannot all differ by \(\pi\). The states with the largest minimal weight, the Fourier states of the Löwdin frame, still yield a maximally coherent state with certainty, but they reach some balanced states only probabilistically: for qutrits at \(s=1/2\), the real state \(\ket R\) with phases \((0,\pi,\pi)\) is reached with probability \(4/7\). No pure state reaches every balanced state with certainty.
Open questions include whether the copy bound \(r^{N+1}\) is attained for four or more copies at all overlaps, asymptotic conversion rates, mixed states, and laboratory implementations of the optimal filters beyond photon subtraction.
Transferable methods
A dual certificate that measures interference
The semidefinite program gives numbers but not structure. A single feasible dual matrix, the projector orthogonal to the all-ones vector, turns it into a closed-form bound with a physical reading in terms of interference, and permutation symmetry shows when that bound is exact.
Theorem 2: \[P^\star(\psi\to\phi)=\max_{w\ge0}\Bigl\{\sum_\pi w_\pi:\ \sum_\pi w_\pi\,\Pi_\pi^TG_\phi\Pi_\pi\le G_\psi\Bigr\}.\] Editorial derivation of the upper bound: take \(Y=\mathbb 1-J/d\). Since \(\mathbf 1^TG\mathbf 1=1\), \(\mathrm{Tr}(YG_\psi)=\|x\|^2-1/d\), and the same value \(\|t\|^2-1/d\) for every permuted target. The dual program then gives \(P^\star\le(\|x\|^2-1/d)/(\|t\|^2-1/d)=\tau(\psi)/\tau(\phi)\).
If \(G_\psi\) is permutation-invariant, averaging any feasible \(w\) over a doubly transitive group keeps it feasible, and the averaged target Gram matrix is a combination of \(\mathbb 1\) and \(J\), which makes the bound attainable. For the qubit with \(s>0\), \(P^\star(\Psi_+\to\Psi_-)=[(1-s)/(1+s)]^2\) and \(P^\star(\Psi_-\to\Psi_+)=1\); with \(s=e^{-2\lvert\alpha\rvert^2}\), \((1-s)/(1+s)=\tanh\lvert\alpha\rvert^2\).
Editorial checks, solving the program numerically: at \(s=1/2\) it gives \(1/9\) and 1 for the two qubit conversions; it matches the closed form of Corollary 1 for three random complex sources and both targets; it gives \(\tanh^2(0.7)=0.3653\) for \(\lvert\alpha\rvert^2=0.7\); for two copies it gives \(r^3=1/27\), against \(r^4=1/81\) copy by copy; and it reproduces Table 3 at \(s=1/2\): 0.100, 0.143, 0.571, 0.804 and 0.300.
How to reuse it
For conversion programs constrained by a matrix inequality between source and target, test with projectors that are invariant under the free symmetry group. Each gives a monotone, a scalar that cannot increase on average, and symmetry of the source or target often makes it tight.
Löwdin frame and photon subtraction as a free filter
To use what is known from coherence theory, the paper moves to the orthonormal basis closest to the reference states. There, superposition-free Kraus operators are incoherent ones conjugated by \(S^{1/2}\), and the optimal cat-state operation turns out to be the annihilation operator.
The Löwdin vectors are \(\ket{l_j}=\sum_k(S^{-1/2})_{kj}\ket{c_k}\), and a superposition-free Kraus operator becomes \([K]_{\mathrm L}=S^{1/2}\Pi_hD_\xi S^{-1/2}\), a permutation times a diagonal, conjugated. For the qubit, \(\mathcal T=S^{1/2}\) stretches the plane into an ellipse with semi-axes \(\sqrt{1\pm s}\), and the balanced states \(\ket{\Psi_\pm}\) lie on its axes.
On the span of \(\ket{\pm\alpha}\), \(\hat a\ket{\pm\alpha}=\pm\alpha\ket{\pm\alpha}\) gives \[\hat a=\alpha\bigl(\ket{c_0}\langle c_0^\perp\rvert-\ket{c_1}\langle c_1^\perp\rvert\bigr),\] a superposition-free phase flip. It exchanges the long and short axes of the ellipse, so after normalisation to unit operator norm it maps the odd cat to the even cat with certainty and the even cat to the odd cat with probability \((\lambda_1/\lambda_0)^2=\tanh^2\lvert\alpha\rvert^2\), the ratio of the two mean photon numbers.
How to reuse it
When a resource theory is defined relative to a nonorthogonal basis, Löwdin orthogonalisation maps it to the orthonormal theory with a fixed non-unitary similarity. The free operations, the monotones and the optimal protocols can then be imported, and the similarity's norm measures what the overlap costs.
Context & limitations
Why it matters. The bounds apply to every physical operation that maps coherent states to coherent states, so they limit cat-state engineering by loss, subtraction and displacement, and they explain why odd cat states can be enlarged with certainty but even ones only probabilistically. They also show that the equal-modulus states, which coincide in ordinary coherence theory, are inequivalent once the reference states overlap.
Limitations. Results concern pure states and finitely many linearly independent reference states, and many closed forms assume equal real overlaps. The free operations are the superposition-free operations, a proper subset of all operations that preserve free states; under the larger class, states reaching every other state exist for every overlap. Several questions are open, including multi-copy optimality beyond three copies, asymptotic rates, and mixed states. The author states that Claude was used under his direction to explore and check derivations and proof strategies, to write and run the numerical tests, to prepare the figures and to improve the wording.