Quantum minimum description of density matrices

1 Introduction

• For fixed d and rank r with distinct positive eigenvalues, the achievable memory cost is L_d,r(n, x) + o(1), with a matching converse through the additive constant.The construction has error O(log n/√n).Source: §1 · p. 2
• The theorem’s scope fixes dimension and spectrum; it makes no uniform claim as positive eigenvalues collide or approach zero.It asserts a sequence of codes with vanishing error, not the same expansion for every prescribed error schedule.Source: §1 · p. 2
• The construction uses covariant channels between GL(d, C) irreducible representations, selecting the PRV component to extend Werner’s symmetric-power cloner.Its finite trace-distance bound depends on highest-weight differences and adjacent-row gaps; for the padded target, the error is O(log n/√n).Source: §1 · p. 2
• The converse bounds memory for irreducible compact-group orbits with a simple largest eigenvalue separated by γ, and applies to arbitrary encoding and decoding channels.A uniform lower bound on the representation-state gap retains the additive constant when transferring the bound to the n-copy family.Source: §1 · p. 2
• For universal lossless compression, the smallest worst-case excess above nS(ρ) is L_d,r(n, x) − s_blk(x) + o(1).Here s_blk(x) is the limiting entropy of the irreducible representation state in a Schur block.Source: §1 · p. 2
• The spectral terms in the memory cost have a geometric interpretation: half the logarithm of the unitary orbit’s Hilbert–Schmidt volume reproduces them.The companion letter relates this comparison to physical and Voiculescu microstate free entropy.Source: §1 · p. 3
• The main theorems are formalized in Lean and Comparator checked.The paper also points to a Lean Certificate and an accompanying Proof Wiki guide.Source: §1 · p. 3

2.1 Related work

• Koashi–Imoto work supplies the algebraic basis for exact compression of mixed states, while purely classical memory does not suffice for unknown-basis families.The cited qudit scheme combines quantum and classical registers.Source: §2.1 · p. 3
• Earlier work established leading costs for mixed qubits and qudits, while this paper determines the additive constant for fixed spectra and extends the YCH construction to qudits.Prior work includes optimal leading-cost results for qudits and the YCH scheme.Source: §2.1 · p. 3

2.2 The YCH qubit protocol

• For qubits, YCH compresses n copies of a state with fixed spectrum diag(p, 1 − p), where the eigenbasis is rotated by unknown U(2).At p = 1/2, the orbit is a single known state and needs no memory.Source: §2.2 · p. 3
• YCH measures the Schur irrep, clones its state into a target irrep, discards the multiplicity register, then reconstructs using a regenerated register and inverse Schur transform.Only the irrep state carries the unknown eigenbasis; concentration near the target controls the error.Source: §2.2 · p. 4
• The YCH scheme achieves vanishing error with memory cost log n + log(2p − 1) + o(1).Its target spin is J⋆ = (p − 1/2)(n + 1) + O(1).Source: §2.2 · p. 5

3 Generalized cloning channels

• Schur–Weyl decomposition measures a block label λ, then a covariant channel maps each resulting irrep state to a fixed target irrep.The target is chosen to dominate the typical set.Source: §3 · p. 5

3.1 Generalized cloner from the PRV component

• The generalized cloning map is the covariant channel built from the PRV component of µ*⊗ν, which is minimal in dominance order and occurs with multiplicity one.The construction also supplies channels for atypical pairs; compression uses the Cartan case on typical pairs.Source: §3.1 · pp. 6–7

3.2.1 Cartan intertwiner formula

• When ν=µ+ω, an isometric Cartan embedding yields a channel whose Choi operator is the normalized projector onto the multiplicity-one PRV component.For symmetric powers, the construction replaces the auxiliary representation state by the maximally mixed state; for pure inputs it gives literal cloning.Source: §3.2.1 · pp. 7–8

3.2.2 Identity and qubit cases

• For GL(2), the generalized map reduces to Werner’s cloner when m≥n and to partial trace when m<n.For symmetric powers, its Kraus rank is |m−n|+1.Source: §3.2.2 · pp. 8–9

3.2.3 Dominance order and dimension

• The PRV component need not have the smallest dimension: for the stated GL(3) example, its dimension is 24 while another summand has dimension 21.The comparison holds even in the dominant case ν=µ+ω.Source: §3.2.3 · p. 9

3.2.4 Reverse cloner

• Reversing the representation pair gives the Petz recovery map of the generalized cloning channel with respect to the maximally mixed input.In the pure-state case, the reverse cloner traces out the excess systems.Source: §3.2.4 · p. 9

3.2.5 Approximation of representation states

• For strictly decreasing positive spectra, the cloning error is controlled by the dominant weight difference relative to the input representation’s row gaps.The stated setting has rank r≤d and partitions with at most r rows.Source: §3.2.5 · p. 9

4 Achievability

• The encoder measures the Schur label without disturbing the representation state, discards the multiplicity register, and clones into a padded target representation.The decoder samples a Schur label, clones back to that representation, appends a maximally mixed multiplicity register, and applies the inverse Schur transform.Source: §4 · p. 10
• The protocol achieves Theorem 1’s memory cost using a sequence of compression codes whose memory is the target representation space HΛ.The memory dimension is evaluated using the asymptotic Weyl-dimension result.Source: §4 · pp. 10, 11, 13
• Padding makes Λ−λ a dominant, nonnegative weight difference for every typical source block, permitting the forward and reverse cloning bounds to apply.On typical blocks, the row-gap parameter bλ grows as Ω(n) under the stated distinct-positive-spectrum conditions.Source: §4 · pp. 12–13
• The typical-block cloning errors are bounded in both directions, while the atypical-block tail is superpolynomially small.The tail bound follows from the stated Sanov-theorem estimate.Source: §4 · pp. 12–13
• The asymptotic memory dimension is evaluated by applying the Weyl-dimension result to Λ, whose first r rows are nxi + O(√n log n).Rows beyond r are zero.Source: §4 · pp. 11, 13

5 Converse

• A spectral gap lets the irreducible-orbit memory bound constrain arbitrary encoding and decoding channels; at zero error it gives dim M ≥ dim H.For the representation states used here, the spectral gap is bounded below uniformly in the partition.Source: §5 · pp. 13–14
• For partitions with at most r rows, the representation state ρλ has a simple largest eigenvalue and a gap at least γx.This uniform gap supplies the condition needed for the irreducible-orbit memory bound.Source: §5 · p. 14
• Theorem 4 applies the converse to codes for the source orbit with Haar-average error and memory Mn, using Ld,r(n,x) as the memory-cost benchmark.The argument transfers an arbitrary code to a padded target representation; Haar-average error suffices for the worst-case error criterion.Source: §5 · p. 14

6 Universal quantum data compression and QMDL

• The excess ideal average length above nS(ρ) is Rn(x) = Ld,r(n,x) − sblk(x) + o(1).Here Ld,r(n,x) is the optimal QMDL cost, and sblk(x) is the finite nonnegative entropy in (51).Source: §6 · pp. 15–16
• The Haar average of ρ⊗n is an exact minimax code state for the unknown-eigenbasis coding problem.It attains the worst-case redundancy value for every eigenbasis.Source: §6 · p. 16
• The offset sblk(x) is the limiting entropy of a typical representation state, while QMDL measures the memory dimension needed to recover the orbit.The average block entropy converges to sblk(x), and the average log representation dimension is Ld,r(n,x) + o(1).Source: §6 · p. 16
• When the spectrum is also unknown, the exact cost split adds the classical penalty D(q∥Pn).The paper states that Appendix B evaluates this additional cost for Hayashi’s universal codes.Source: §6 · p. 16

7.1 Technical preliminaries

• Weight multiplicities are monotone under dominant highest-weight addition: mµ(δ) ≤ mµ+ω(δ) for every δ ∈ Q+.The proof injects Gelfand–Tsetlin patterns for µ into those for µ + ω while shifting their weights.Source: §7.1 · p. 17

7.2 Proof

• The forward-channel error reduces to the positive-part trace of the difference between the embedded target state and the retained highest-weight Kraus branch.The omitted Kraus branches contribute their missing trace, and the branch comparison can then be made separately in each total-weight sector.Source: §7.2 · p. 19
• For r ≥ 2, define gµ := min1≤i<r(µi − µi+1) and assume |δ| ≤ gµ; the projectors have equal rank, so ∥Pδ − Qδ∥ = ∥(I − Pδ)Qδ∥. For r = 1, they coincide.With D := ∥ω∥1, 0 ≤ (δ, ω) ≤ |δ|D.Source: §7.2 · pp. 20–21
• For strictly decreasing positive spectrum, the mean weight depth is uniformly bounded by K(x), and it vanishes when r = 1.This controls the averaged projector deficits and the tail mass used in estimating the eigenvalue ratio.Source: §7.2 · p. 21
• The eigenvalue-ratio deviation is controlled by sectors where multiplicities can differ, while the Weyl-dimension ratio is bounded using D and bµ.The multiplicities agree through depth gµ; the dimension estimate uses a numerator at most D and denominator at least bµ + 1.Source: §7.2 · pp. 21–22
• The proof bounds the full forward and reverse channel errors by averaging weight-block deficits, with scalar eigenvalue and Weyl-dimension ratios completing the estimates.The mean-depth bound is uniform over representations; the final estimates use bµ ≤ gµ and D ≥ 1.Source: §7 · p. 17, §7.2 · p. 24
• When r = 1, Z(x) = 1 and M_μ = M_ν = 0; the forward bound is 2 𝟙/(b_μ + 1), with zero reverse error.Source: §7.2 · p. 25
• For all cases, one may take C_d,x = 2 𝟙 + 3K(x), with K(x) = 0 when r = 1.The passage gives this choice after deriving the forward bound and zero reverse error for r = 1.Source: §7.2 · p. 25

• AI tools assisted with proof exploration, mathematical revisions, and language editing.Source: Generative AI usage · p. 25
• An initial strategy used GL(d) Clebsch–Gordan coefficients, followed by a projector-perturbation approach using the Davis–Kahan theorem.Source: Generative AI usage · p. 25
• The final proof of Lemma 7 used a direct Casimir argument that yielded Theorem 2’s finite bound.Source: Generative AI usage · p. 25
• The authors reviewed and revised AI-assisted material and take responsibility for the mathematical arguments and manuscript content.Source: Generative AI usage · p. 25

• The PRV component occurs with multiplicity one, making the intertwiner’s squared operator proportional to its projector.Source: A Proofs · p. 26
• The scaled adjoint of the cloning channel is the reverse cloning channel, as shown by identifying its Choi operator.Source: A Proofs · pp. 26–27
• The highest-weight line is the simple largest-eigenvalue space, with every other nonzero eigenvalue at most qx times its eigenvalue.Source: A Proofs · p. 27
• A uniform lower bound on the highest-weight eigenvalue follows by bounding weight multiplicities with positive-root partitions.Source: A Proofs · pp. 27–28
• When the weight depth is at most the row gap, multiplicities equal positive-root partition counts; this yields convergence of the associated entropy quantities as the gap grows.Source: A Proofs · p. 28

• When the spectrum is unknown, choosing code weights independently of it adds the classical penalty D(q∥Pn), while the eigenbasis contribution remains Rn(x).Source: B Additional cost of an unknown spectrum · p. 29
• The unknown-spectrum contribution has the same 2 log n leading term as classical coding with d −1 unknown spectral parameters.Source: B Additional cost of an unknown spectrum · p. 29
• An asymptotically optimizing code is a continuous mixture over ordered positive spectra, analogous to the quantum Jeffreys-prior construction.Source: B Additional cost of an unknown spectrum · p. 30
• 2 log n is the leading overhead for each of the unknown eigenbasis and unknown spectrum contributions.Source: B Additional cost of an unknown spectrum · p. 30