Quantum information · Theory
Catalytic compression of classical communication in exact entanglement transformations
arXiv:2611.21349v1 ↗PDF ↗Submitted 16 November 2026
- Question
- Nielsen's theorem says when one bipartite pure state can be turned into another exactly by local operations and classical communication, but not how much communication the conversion needs. A catalyst, an extra entangled state returned unchanged at the end, is known to make some impossible conversions possible. Can it also reduce the classical communication needed for conversions that are already possible, and by how much?
- Result
- The smallest message alphabet with an optimal catalyst is \[C^*(p\to q)=\bigl\lfloor2^{\Delta_R(p,q)}\bigr\rfloor+1,\] where \(\Delta_R\) is the largest Rényi entropy difference between the source and target Schmidt vectors; equivalently, an \(m\)-message protocol exists exactly when \(\Delta_R<\log_2m\). For every \(d\ge3\) an explicit pair needs \(d\) messages without a catalyst, the maximum possible, but only 2 (one bit) with one. For \(n\) copies processed together the rates with and without a catalyst coincide, both equal to \(\Delta_R\) per copy.
The paper in detail
Alice and Bob share \(\ket{\psi_p}=\sum_i\sqrt{p_i}\ket{ii}\) and want \(\ket{\phi_q}\). Nielsen's theorem says an exact deterministic conversion by local operations and classical communication exists precisely when \(p\) is majorized by \(q\), and one-way communication from Alice to Bob suffices. The authors count communication by the number of distinct messages Alice may send. A result of Daftuar and Hayden turns this into a matrix question: an \(m\)-message protocol exists exactly when Alice's reduced state \(P=\mathrm{diag}(p)\) is an equal-weight average of \(m\) unitary conjugates of \(Q=\mathrm{diag}(q)\). Minimising communication means finding the fewest rotated copies of the target that average to the source.
A catalyst \(\ket{\chi_r}\) may be borrowed and must be returned exactly and uncorrelated. The main theorem gives the optimum over all finite catalysts: \(\lfloor2^{\Delta_R}\rfloor+1\) messages, where \(\Delta_R=\max_{\alpha\in[0,\infty]}[H_\alpha(p)-H_\alpha(q)]\). The strict inequality matters: if \(2^{\Delta_R}\) is an integer \(M\), then \(M\) messages are not enough. In particular a single bit suffices exactly when every Rényi entropy drops by less than one bit.
An explicit family shows how large the saving can be. With \(p=(\tfrac12,\tfrac14,\tfrac1{4(d-2)},\dots)\) and \(q=(\tfrac34,\tfrac1{4(d-1)},\dots)\), the target reduced state is a multiple of the identity plus a rank-one piece, so every rotated copy contributes one direction while the source minus that multiple of the identity has full rank: \(d\) messages are needed, and \(d\) suffice. With a catalyst, a majorization between \(q\otimes(\tfrac12,\tfrac12)\) and \(p\) padded with zeros shows that all Rényi gaps are below one bit, so one bit suffices. The ratio \(d/2\) is unbounded.
For many copies the authors compare rates. With a catalyst the cost at block length \(n\) is \(\lfloor2^{n\Delta_R}\rfloor+1\), so the rate is \(\Delta_R\) per copy. Without one the rate is the same, because a single catalyst can be prepared once from a maximally entangled state, used across many blocks and removed at the end, for a cost that does not grow with the number of blocks. Catalysis therefore compresses individual conversions without changing the leading cost per copy.
Transferable methods
A pinching argument that cancels the catalyst
A lower bound must hold for every catalyst, of any size. Writing the equal-weight decomposition as a Gram matrix and pinching it to its diagonal blocks produces a majorization in which the catalyst appears as a tensor factor on both sides, so additivity of Rényi entropies removes it.
Suppose \(P\otimes R=\frac1m\sum_jX_j\) with each \(X_j\) unitarily equivalent to \(Q\otimes R\). Let \(V=m^{-1/2}[X_1^{1/2}\cdots X_m^{1/2}]\). Then \(VV^\dagger=P\otimes R\), so \(G=V^\dagger V\) has spectrum \((p\otimes r)\oplus0\). Its diagonal blocks are \(X_j/m\), so pinching away the off-diagonal blocks gives a matrix with spectrum \(u_m\otimes q\otimes r\), where \(u_m\) is uniform on \(m\) points.
Pinching is an average of unitary conjugations by \(W^k\), with \(W=\mathrm{diag}(1,\omega,\dots,\omega^{m-1})\otimes I\), so the pinched spectrum is majorized by the original: \[u_m\otimes q\otimes r\prec(p\otimes r)\oplus0.\] The two sides have different numbers of nonzero entries, so strict Schur concavity gives, for \(0<\alpha<\infty\), \(\log_2m+H_\alpha(q)+H_\alpha(r)>H_\alpha(p)+H_\alpha(r)\), and the catalyst cancels. At \(\alpha=\infty\), positivity of the remaining \(X_j\) gives \(p_1r_1>q_1r_1/m\); at \(\alpha=0\) the gap is zero. Hence \(\Delta_R<\log_2m\).
Editorial check: the condition must already hold without a catalyst. For 300 random averages of \(m=2\) to 4 random unitary conjugates of random spectra, \(\Delta_R(p,q)<\log_2m\) every time. In the explicit family the maximising order is not the Shannon entropy: at \(d=3\) the Shannon gap is 0.439 and the order-2 gap 0.663, while \(\Delta_R=0.727\) is reached near \(\alpha=3.5\).
How to reuse it
When a resource that must be returned intact appears on both sides of a convertibility statement, look for a majorization in which it enters as a tensor factor; additive monotones such as Rényi entropies then cancel it and give bounds valid for every catalyst. Pinching is a convenient way to produce such majorizations from decompositions into unitary orbits.
Rank counting for decompositions into unitary orbits
For the explicit family, the lower bound without a catalyst comes from counting ranks. The target reduced state is a multiple of the identity plus a rank-one piece, so each unitary conjugate adds one direction, while the source minus that multiple of the identity has full rank.
With \(\beta=1/(4(d-1))\), \(Q_d=\beta I+(\tfrac34-\beta)\ket1\bra1\). Any decomposition \(P_d=\sum_j\lambda_jU_jQ_dU_j^\dagger\) gives \[P_d-\beta I=\bigl(\tfrac34-\beta\bigr)\sum_{j=1}^M\lambda_j\ket{v_j}\bra{v_j},\qquad\ket{v_j}=U_j\ket1.\] Every \(p_i\) exceeds \(\beta\), so the left side has rank \(d\) while the right has rank at most \(M\); hence \(M\ge d\). Weights \(w_i=(p_i-\beta)/(\tfrac34-\beta)\) with permutations sending \(\ket1\) to \(\ket i\) achieve \(M=d\).
With a catalyst, up to permutations \(q\otimes u_2=(\tfrac38,\tfrac38)\oplus\tfrac14u_{2d-2}\) and \(p\oplus0_d=(\tfrac12,\tfrac14)\oplus\tfrac14(u_{d-2}\oplus0_d)\); each block on the left is majorized by the corresponding block on the right, so \(q\otimes u_2\prec p\oplus0_d\). Strict Schur concavity and additivity give \(H_\alpha(q)+1>H_\alpha(p)\), the endpoint gap is \(\log_2\tfrac32<1\), so \(\Delta_R<1\) and one bit suffices. Editorial check: the computed gaps rise from 0.727 at \(d=3\) to 0.883 at \(d=10^4\), staying below 1.
How to reuse it
Rank arguments give clean lower bounds on the number of terms needed to write a state as a mixture of unitary conjugates of an identity-plus-low-rank state. The same count bounds Kraus numbers of random-unitary channels and the number of outcomes in measurement-based protocols whose branches must all reach one target.
Context & limitations
Why it matters. It shows that catalysis changes not only what can be done but how cheaply, with an exact formula for the optimal one-shot catalytic cost, while the corresponding cost without a catalyst has no such closed form. The saving can be unbounded for a single conversion yet vanishes in the leading many-copy rate, a clean separation between one-shot and asymptotic behaviour.
Limitations. The result assumes exact conversion, full-support Schmidt vectors of equal length, and a catalyst returned exactly and uncorrelated with the target; relaxing return conditions changes the answer, up to embezzlement. The catalyst's Schmidt rank is unbounded and can be large, and how cost trades against catalyst size is not determined. Communication is measured by the number of distinct messages, not by message entropy. The authors state that ChatGPT and Claude assisted with language editing, literature searches and discussions of mathematical arguments.