Counterexamples to additivity of minimum output $p$-Rényi entropy of quantum channels for $p>3/4$ and $0\leq p<1/4$

Abstract: Additivity of minimum output entropies is a central problem in quantum information theory. Nonadditivity is known for every Rényi order $p&gt;1$, at the von Neumann point $p=1$, and near $p=0$, while most of the interval $0&lt;p\&lt;1$ has remained open. We prove that for every Rényi order $p$ satisfying either $$p\&gt;3/4$$ or $0\leq p&lt;1/4$, there exist finite-dimensional projection-induced quantum channels such that additivity of the minimum output $p$-Rényi entropy fails. The proof combines two correlated random-projection constructions: a product-conjugate Bell-state witness for $p&gt;3/4$, and a transpose-complement rank-defect witness for $p&lt;1/4$. Thus the unresolved part of $0<p<1$ is reduced to $[1/4,3/4]$. Our estimates also improve the output dimension threshold for additivity violation of minimum output von Neumann entropy, first established in Belinschi, Collins and Nechida.

• The paper proves additivity failure for minimum output p-Rényi entropy for projection-induced quantum channels across 0≤p<1/4 and p>3/4, leaving only [1/4,3/4] unresolved.
• The authors use correlated random-projection constructions, combining product–conjugate Bell-state witnesses above p=3/4 with transpose-complement rank defects below p=1/4.
• The analysis establishes rigorous asymptotic output-body limits and improves the ensemble-specific von Neumann violation threshold from output dimension 183 to 182, while deterministic finite-dimensional examples remain open.

The additivity of minimum output entropies is a central open problem in quantum information theory, tied through Shor's equivalence to additivity of the Holevo capacity and entanglement of formation. Prior to this work, nonadditivity of the minimum output $p$-Rényi entropy was established for every $p>1$ (Hayden–Winter, with constructive versions by Grudka et al. and Derksen), at the von Neumann point $p=1$ (Hastings), and near $p=0$ (Cubitt–Harrow–Leung–Montanaro–Winter), while the interior of the interval $0$p>3/4$ and all $0\leq p<1/4$, thereby reducing the unresolved regime to $[1/4,3/4]$ and providing the first explicit uniform endpoints for the nonadditivity intervals adjacent to both ends of $(0,1)$.

Main result

The central statement is that for every $p\in[0,1/4)\cup(3/4,\infty]$, there exist finite-dimensional projection-induced channels $p>1$0 such that

$p>1$1

where $p>1$2 with the endpoint conventions $p>1$3 and $p>1$4 the von Neumann entropy. Two features of this result deserve emphasis. First, the new range $p>1$5 extends nonadditivity below the von Neumann point by a quantitative margin, and the asymptotic gap function used in the proof extends continuously to $p>1$6 and remains positive there, showing that the von Neumann point is not singular for this mechanism. Second, at the low-$p>1$7 end, the paper replaces the previously unspecified "sufficiently small $p>1$8" of Cubitt et al. by the explicit interval $p>1$9, improving on the numerically suggested endpoint of roughly $p=1$0 for the original explicit example.

Projection-induced channels and the limiting output body

The counterexamples are drawn from a random-channel ensemble built from Haar-distributed bipartite projections. For a projection $p=1$1 on $p=1$2 with $p=1$3, the locally normalized operator $p=1$4 is the Choi matrix of a trace-preserving channel, since $p=1$5. Fixing $p=1$6 and letting $p=1$7 with $p=1$8, the condition $p=1$9 guarantees invertibility of $p=0$0 almost surely.

The technical core is a deterministic identification of the large-$p=0$1 output set. Using strong asymptotic freeness and the strong block-modification theorem, together with a Legendre-type convex duality for the Bernoulli spectral edge, the authors prove that the output sets $p=0$2 converge in trace-norm Hausdorff distance, almost surely, to the compact convex body

$p=0$3

The proof proceeds through support functions: the support functional $p=0$4 of the random output set equals $p=0$5, and the random compression estimate identifies its limit as the support function of $p=0$6, with a $p=0$7-net argument promoting pointwise to uniform convergence. Since $p=0$8 is uniformly continuous on $p=0$9 for $0

The product–conjugate witness for $0

For the high-$0$p>3/4$0. The proof uses entrywise convergence of the Choi blocks (via a free-probability computation of the asymptotic normalized Choi purity $p>3/4$1) and a $p>3/4$2-twirling argument to pin down the two-eigenvalue form.

Comparing large-$p>3/4$3 expansions reduces the violation test to the scalar inequality

$p>3/4$4

with $p>3/4$5 by continuous extension. Since $p>3/4$6, the inequality holds for sufficiently small $p>3/4$7 exactly when $p>3/4$8; for $p>3/4$9 the ratio diverges. This covers $0\leq p<1/4$0 within the same mechanism, without a separate Hastings-type argument. The implication is that for every $0\leq p<1/4$1, including the von Neumann case, additivity fails for projection-induced channels of sufficiently large input dimension.

On the finite-dimensional side, numerical optimization of the asymptotic criterion gives $0\leq p<1/4$2: the smallest output dimension at which the Bell-state witness detects a von Neumann violation in this ensemble. This improves by one the threshold $0\leq p<1/4$3 established by Belinschi, Collins, and Nechita for the Haar–Stinespring ensemble, an improvement attributable to the local normalization changing the finite-$0\leq p<1/4$4 geometry of the limiting output body. The authors are careful to note that $0\leq p<1/4$5 is a witness-specific threshold, not a universal one.

The transpose-complement witness for $0\leq p<1/4$6

The low-$0\leq p<1/4$7 construction pairs a half-rank projection-induced channel $0\leq p<1/4$8 (with $0\leq p<1/4$9) with its transpose-orthogonal partner $[1/4,3/4]$0 built from $[1/4,3/4]$1. The marginal law of $[1/4,3/4]$2 is again Haar with rank ratio $[1/4,3/4]$3, so both one-channel output sets converge to $[1/4,3/4]$4. The essential additional structure is the exact orthogonality $[1/4,3/4]$5, which, via the rank-deficit lemma of Cubitt et al., produces a joint pure input whose output has rank at most $[1/4,3/4]$6 and hence $[1/4,3/4]$7-entropy at most $[1/4,3/4]$8 for every $[1/4,3/4]$9.

The comparison rests on the sharp expansions

$(0,1)$0

whose difference is $(0,1)$1, positive precisely for $(0,1)$2. At $(0,1)$3 the argument is purely rank-theoretic: one-channel outputs are almost surely full rank while the joint output has rank at most $(0,1)$4, recovering the nonmultiplicativity of minimum output rank. The paper is explicit that the projection construction itself is due to Cubitt et al.; the contribution here is the quantitative asymptotic entropy analysis that converts their qualitative result into a rigorous endpoint $(0,1)$5.

Relation to prior work and the status of the remaining interval

The paper situates itself carefully within the literature. The product–conjugate Bell-state analysis builds on Collins and Nechita, and the deterministic limiting output set refines work of Belinschi, Collins, and Nechita for the Haar-random-subspace ensemble; the ensemble here differs through the inverse-marginal normalization of the Choi matrix, which alters both the one-channel output body and the limiting isotropic spectrum. The authors also flag that a 2010 preprint by Yu and Ying announced nonadditivity for $(0,1)$6 with $(0,1)$7, but was withdrawn due to a crucial error, so those ranges are not regarded as established. The result leaves open precisely the interval $(0,1)$8, and the paper does not claim otherwise.

Limitations and open questions

Several limitations are acknowledged or evident. The counterexamples are probabilistic: for each $(0,1)$9 in the covered ranges, violation holds almost surely for sufficiently large input dimension, but no deterministic finite-dimensional instance is constructed, and the long-standing problem of an explicit Hastings-type realization remains open. The finite-$p\in[0,1/4)\cup(3/4,\infty]$0 thresholds $p\in[0,1/4)\cup(3/4,\infty]$1 and $p\in[0,1/4)\cup(3/4,\infty]$2 are properties of the specific ensemble and witnesses; the low-$p\in[0,1/4)\cup(3/4,\infty]$3 threshold grows as $p\in[0,1/4)\cup(3/4,\infty]$4 near the endpoint, so the construction becomes increasingly demanding as $p\in[0,1/4)\cup(3/4,\infty]$5, and nothing in the paper indicates whether additivity actually holds at $p\in[0,1/4)\cup(3/4,\infty]$6 or $p\in[0,1/4)\cup(3/4,\infty]$7 or fails there with a different witness. The paper also notes that the construction of Belinschi et al. appears difficult to adapt to the low-$p\in[0,1/4)\cup(3/4,\infty]$8 regime. Whether the mechanism extends through the unresolved central interval, and whether the Hausdorff convergence technique can be pushed to give uniform (rather than almost-sure) statements, are questions the paper leaves open.

Conclusion

This paper establishes additivity violation of minimum output $p\in[0,1/4)\cup(3/4,\infty]$9-Rényi entropy for projection-induced channels across all $p>1$00 and $p>1$01, via two correlated random-projection witnesses: a product–conjugate Bell-state test above the von Neumann point and a transpose-complement rank-defect test below it. The analysis combines strong asymptotic freeness, a Legendre-duality characterization of the Bernoulli spectral edge, and sharp large-$p>1$02 entropy asymptotics, and it yields a modest but concrete improvement of the von Neumann output-dimension threshold from 183 to 182. With these results, the additivity question for $p>1$03 is reduced to the interval $p>1$04.