Quantum information · Theory and numerics (AI-assisted)

The magical memory cost of entanglement spectrum compression

Matias Karjula, Vilma Nieminen, Teemu Ojanen, Tapio Ala-Nissila, Moein N. Ivaki

arXiv:2611.24710v1 ↗PDF ↗Submitted 18 November 2026

Question
The entanglement spectrum of a bipartite pure state, its list of Schmidt probabilities, sets how expensive the state is to represent as a matrix product state and how it can be transformed. Local unitaries cannot change the spectrum, and a Bell pair contributes a flat factor. Suppose each party may use any local unitary to separate out \(b\) Bell pairs and must keep everything else in a residual register of \(k\) qubits. How small can \(k\) be at a given fidelity, and which properties of the spectrum control it?
Result
The best fidelity for an allocation \((b,k)\) is \(F_{b,k}=2^{-b}\sum_j\bigl(\sum_s\sqrt{\lambda_{j2^b+s}}\bigr)^2\), a sum over consecutive blocks of the sorted Schmidt amplitudes. The least residual memory \(k_\epsilon\) at error \(\epsilon\) is bounded below by the smoothed entanglement spread, and by \((1-\epsilon)2^D-1\), where \(D\) is the nonlocal magic. Nonlocal magic that grows as \(\beta\log_2n\) therefore forces memory of order \(n^\beta\) at any fixed fidelity, so universally embezzling families need unbounded memory. Haar-typical states need only \(O(\log1/\epsilon)\) residual qubits, since \(1-F_k\le\tfrac12 4^{-k}\) in the limit, with \(F_{\rm NL}\approx0.844\) and \(D\approx0.24\) bits. In chaotic circuits the memory cost peaks and then falls while the entropy keeps growing. Stationary free-fermion states keep a cost that grows with size; their raw spread is \((4G/\pi\ln2-1)\,n\approx0.682\,n\). At the chaotic peak, the bond dimension needed at fixed fidelity exceeds the entropy estimate \(2^{S_1}\) by a factor that grows with system size.

The paper in detail

Write a pure state of \(n+n\) qubits as \(\sum_i\sqrt{\lambda_i}\ket{a_i}\ket{b_i}\) with \(\lambda_0\ge\lambda_1\ge\dots\). Each party splits its qubits into a Bell register of \(b\) qubits, a residual register of \(k\) qubits, and \(n-b-k\) qubits reset to \(\ket0\). The target is \(b\) Bell pairs times an arbitrary state \(\ket{\gamma_k}\) on the residual registers. Bell pairs are the standard currency of entanglement and carry a perfectly flat spectrum. The residual register must hold whatever non-flatness cannot be discarded within the error budget. The figure of merit is \(k_\epsilon\), the least residual size for which some \(b\) reaches fidelity \(1-\epsilon\), with the largest such \(b\) as a secondary output. Stabilizer states have \(k_0=0\), because local Cliffords turn them into Bell pairs and product states.

The optimisation over encoders has a closed form. By von Neumann's trace inequality, the best local unitaries align Schmidt bases, so only the spectra matter. Cauchy–Schwarz then fixes the best residual spectrum. The resulting fidelity separates into two losses: the probability beyond the \(2^{b+k}\) largest Schmidt weights, and a flattening error from replacing each block of \(2^b\) amplitudes by their mean. With no residual register, the optimum is the known nonlocal stabilizer fidelity \(F_{\rm NL}\), and \(D=-\log_2F_{\rm NL}\) is the nonlocal magic. Grouping the residual terms in dyadic blocks gives \(F_{\rm NL}\le F_k\le(k+1)F_{\rm NL}\), and with it a memory bound that diverges whenever \(D\) does.

A second bound uses entanglement spread, \(\Delta_0=\log_2(\mathrm{rank}\,\rho_A\,\|\rho_A\|_\infty)=S_0-S_\infty\), which measures how far the reduced spectrum is from flat. It is additive and unchanged by adding Bell pairs. Every admissible target has spread at most \(k\), so the smoothed spread \(\Delta_\epsilon\), the least spread among spectra within fidelity \(1-\epsilon\) of the true one, gives \(\lceil\Delta_\epsilon\rceil\le k_\epsilon\). The Supplemental Material computes \(\Delta_\epsilon\) exactly, with capped water-filling and a search over ranks. In the chaotic-circuit data, the spread bound is tighter than the magic bound, and both follow the rise and fall of the true cost.

The numerics compare three ensembles. In brickwork circuits of random single-qubit gates and controlled-\(Z\) gates, the memory cost rises to a size-dependent peak and then falls to a common late-time value. Meanwhile the Bell-pair fraction keeps the entropy growing. Late-time states behave like Haar states, for which the Marchenko–Pastur limit gives a size-independent trade-off curve. Free-fermion (matchgate) circuits instead settle into a stationary ensemble whose memory cost keeps growing with size. A data collapse over the sizes studied gives an exponent near 0.59, consistent with the \(\sqrt n\) lower bound implied by their nonlocal magic. Open: how efficiently encoders, for example quantum autoencoders, can approach the optimal fidelity.

Transferable methods

  1. Von Neumann's trace inequality solves the encoder optimisation

    The encoders act on \(n\) qubits each and are unrestricted, so a brute-force optimisation is hopeless. But fidelity under local unitaries depends only on how the two states' Schmidt coefficients are paired. The trace inequality says the best pairing is largest with largest. What remains is a one-line Cauchy–Schwarz over the residual spectrum. The closed form exposes the two losses, truncation and flattening, and it allows bounds directly in terms of nonlocal magic.

    Write both states as coefficient matrices \(C_\psi\) and \(C_\Omega\), whose singular values are the Schmidt amplitudes. Local unitaries map \(C_\psi\) to \(U_AC_\psi U_B^T\), and \(|\operatorname{tr}(C_\Omega^\dagger U_AC_\psi U_B^T)|\le\sum_i\sigma_i(C_\Omega)\sigma_i(C_\psi)\), with equality when the singular vectors are aligned. The target's amplitudes are \(\sqrt{\mu_j/2^b}\), each repeated \(2^b\) times. With \(A_j=\sum_{s<2^b}\sqrt{\lambda_{j2^b+s}}\), Cauchy–Schwarz gives \[\Bigl(2^{-b/2}\sum_jA_j\sqrt{\mu_j}\Bigr)^2\le2^{-b}\sum_jA_j^2=F_{b,k},\] with equality for \(\mu_j\propto A_j^2\), which is already sorted. Subtracting from 1 splits the loss into the tail \(\sum_{r\ge2^{b+k}}\lambda_r\) plus \(\sum_{j,s}(\sqrt{\lambda_{j2^b+s}}-\bar a_j)^2\), where \(\bar a_j\) is the block mean.

    For the magic bound, put \(T_j=A_j/2^{b/2}\), a non-increasing sequence with \(F_{b,k}=\sum_jT_j^2\). The term \(T_0^2\) is a Bell-sector fidelity, so it is at most \(F_{\rm NL}\). For each dyadic group \(j=m,\dots,2m-1\), monotonicity gives \[\sum_{j=m}^{2m-1}T_j^2\le m\Bigl(\frac1m\sum_{j<m}T_j\Bigr)^2=\frac1{2^bm}\Bigl(\sum_{i<2^bm}\sqrt{\lambda_i}\Bigr)^2\le F_{\rm NL},\] since \(2^bm\) is again a power of two. There are \(k\) such groups, so \(F_k\le(k+1)F_{\rm NL}\), and \(F_k\ge1-\epsilon\) forces \(k\ge(1-\epsilon)2^D-1\). The editorial check found no violation on 2000 random spectra, and no random encoder exceeded the closed form.

    How to reuse it

    Many tasks fix a target with a structured spectrum and allow free local unitaries on one side: entanglement dilution with bounded memory, catalytic transformations and embezzlement, matrix product state truncation with flat symmetry sectors, and autoencoders that separate Bell pairs. Each reduces to sums over sorted Schmidt amplitudes of this kind, with the same split between truncation and flattening.

  2. Marchenko–Pastur quantiles turn Haar compression into block averaging

    To claim that typical states need only accuracy-dependent memory, one must control the flattening error for a spectrum of \(2^n\) random numbers. In the large-\(n\) limit the sorted spectrum becomes a deterministic quantile function. The compression fidelity becomes a statement about approximating that function by block averages, and a Poincaré inequality bounds the error with an explicit constant.

    For a balanced cut, the rescaled weights \(x=2^n\lambda\) have density \(\rho(x)=(2\pi)^{-1}\sqrt{(4-x)/x}\) on \((0,4)\). Substituting \(x=4\cos^2\theta\), the fraction of weights above \(x\) is \(u=(2\theta-\sin2\theta)/\pi\). The cumulative amplitude is \(X(u)=\int_0^u\sqrt{x}\,du'=\tfrac{8}{3\pi}\sin^3\theta\). With \(r=n-b-k\) reset qubits and blocks of width \(h=2^{-(k+r)}\), \[\mathcal F_{k,r}=\frac1h\sum_{j=0}^{2^k-1}\bigl[X((j+1)h)-X(jh)\bigr]^2.\] This gives \(\mathcal F_{0,0}=0.7205\) and \(\mathcal F_{0,1}=0.8444\), so with no residual register it is optimal to give up one qubit per side. With one residual qubit and \(r=0\), \(\mathcal F_{1,0}=0.9236\) already beats every allocation with \(r\ge1\), which is capped by \(P(1/2)=0.8937\).

    At \(r=0\) nothing is truncated, and with \(f=\sqrt{x(u)}\) and \(\int_0^1f^2=1\), \[1-F_k=\sum_j\int_{jh}^{(j+1)h}\bigl(f-\bar f_j\bigr)^2du\le\frac{h^2}{\pi^2}\int_0^1|f'(u)|^2du=\frac12\,4^{-k}.\] The last step uses \(f'=-\pi/(2\sin\theta)\) and \(du=4\sin^2\theta\,d\theta/\pi\), so the integral equals \(\pi\int_0^{\pi/2}d\theta=\pi^2/2\). The derivative diverges at the top of the spectrum, but its square is integrable. Each extra residual qubit halves the block width and quarters the bound. The editorial check found the exact limiting error at 61 to 76 per cent of the bound for \(k=1\) to 6.

    How to reuse it

    For typical states of large random ensembles, sums over sorted spectra can be replaced by integrals over the limiting quantile function, and block-averaging errors then follow from Poincaré or Wirtinger inequalities. This gives quick truncation estimates for random tensor networks, Page-curve states, and induced mixed-state ensembles with other density laws.

Source: Main text (Eqs. (1)–(8)); Supplemental Material (optimal compression fidelity, Eqs. (S1)–(S3); memory bound from fidelity, Eq. (S4); Haar compression, Eqs. (S5)–(S9)). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. The paper gives two measures discussed mostly as diagnostics, nonlocal magic and entanglement spread, an operational meaning. They bound the quantum memory needed to store a state's entanglement beyond Bell pairs. It also links a structural property, the multiscale spectra behind entanglement embezzlement and the transient "magic barrier" of chaotic dynamics, to a practical cost: matrix product state truncation estimated from entropy alone can underestimate the needed bond dimension by a growing factor.

Limitations. The setting is a single copy, one cut, and unrestricted local unitaries. The cost of finding or implementing the encoders is not addressed, and the paper names it as open (for example, for quantum autoencoders). The circuit results are numerical at finite sizes; the single-size chaotic panels use \(N=26\) qubits, and the range of sizes appears only in the figures. The exponents \(\beta\approx0.45\) and \(\beta\approx0.59\) come from collapsing data over limited sizes, not from proofs. The growth \(D\approx\tfrac12\log_2n\) for free-fermion states is cited from other work, not derived. The authors used the GPT-6 family for literature discovery, cross-checking reasoning, and assistance with mathematical proofs, code and automation.

Description based on the full arXiv HTML of v1, read 2026-11-22: the main text and the Supplemental Material (optimal fidelity, memory bound from fidelity, the Haar limit with Eqs. (S5)–(S9), and the smoothed-spread algorithm, Eqs. (S10)–(S17)). Editorial checks: the Haar limit computed from the Marchenko–Pastur quantile parametrisation reproduces \(F_{0,0}=0.7205\), \(F_{0,1}=0.8444\), \(F_{1,0}=0.9236\), \(P(1/2)=0.8937\), \(P(1/4)=0.6229\) and \(D=0.244\) bits. For \(k=1\) to 6, \(1-F_k\) is 0.61 to 0.76 of the bound \(\tfrac124^{-k}\). Twenty Haar states with nine qubits per side give mean \(F_0,\dots,F_3=0.844\), 0.924, 0.980 and 0.995, and \(2^n\lambda_{\max}=3.93\), against the limit 4. The bounds \(F_{\rm NL}\le F_k\le(k+1)F_{\rm NL}\) held for 2000 random heavy-tailed spectra. The Catalan expression and the arcsine integral both equal 0.68253. Random local unitaries and residual states never exceeded Eq. (2) for two qubits per side.