Blueprint

🔍 Explore view: explanations, figures, other mathematical languages and labs are shown, marked commentary; they are not part of the paper. Switch to Paper for the statements, proofs and evidence only.

Robust qubit–memory transfer against quasi-static frequency noise costs time and exposure: exact bounds from Dubins curves

Carles Marín

Companion site: https://karlesmarin.github.io/blueprints/robust-offset/: every statement with its dependencies and its checks, and labs in the browser; each figure of this paper is a link to the lab that reproduces it with its data. This paper: doi:10.5281/zenodo.24250503 (Zenodo), with its supplementary material.

At a glance

🚀 Start here: Theorem 1: Robustness costs exposure and time: the disk inequality and the minimal duration T₀ + arccos(|ζ|/z₀)/G.

Main results:

  • Theorem 1: Robustness costs exposure and time: the disk inequality and the minimal duration T₀ + arccos(|ζ|/z₀)/G
  • Proposition 3: Fixed phase: the exact minimal exposure cost, from the shortest Dubins U-turn
  • Proposition 5: Second-order error under quasi-static noise, for every pulse; when the square pulse is optimal
  • Proposition 8: Slow noise together with memory loss: square-pulse optimality and the large-noise asymptotics
11 results ✅ 0 entirely in Lean ◐ 7 partly in Lean ✅ 3 in Isabelle 💻 0 checked by a program written proof only (a result can carry several tags)
🗺️ Map of the results (click a box to go to it)

Dependencies in the written proof: an arrow A → B means that this proof of B uses A. Every result rests on its written proof; the tags record what was checked besides, and each statement's “Evidence & scope” block says how far. The colour of a box is its strongest tag, in the order above; the tags in the box give the rest. Boxes with a thick border are the main results; click a box to read the statement.

cluster_0 Robustness costs time and exposure thm:robust Theorem 1 Robustness costs exposure and time: the disk inequality and the minimal duration T₀ + arccos(|ζ|/z₀)/G ✅ Isabelle cor:robustcost Corollary 2 A robust transfer costs at least max(√(γq γₘ), (π/2) min γ)/G in extra exposure ◐ part in Isabelle thm:robust->cor:robustcost auto:remark-4 Remark 4 Two representative pulses cancel the leakage in the least time π/G thm:robust->auto:remark-4 prop:slow2 Proposition 5 Second-order error under quasi-static noise, for every pulse; when the square pulse is optimal ◐ part in Isabelle thm:robust->prop:slow2 lem:planar Lemma 6 Planar exposure bound for pulses of fixed phase ◐ part in Isabelle thm:robust->lem:planar rem:planar Remark 7 The exact planar section: a Dubins car that turns back ◐ part in Isabelle thm:robust->rem:planar prop:noiseloss Proposition 8 Slow noise together with memory loss: square-pulse optimality and the large- noise asymptotics ◐ part in Isabelle thm:robust->prop:noiseloss rem:corr Remark 10 Correlated offsets: the same formula, and when correlation makes the square pulse optimal ◐ part in Isabelle thm:robust->rem:corr rem:unified Remark 11 One sufficient condition for square-pulse optimality, covering relaxation of both elements and correlated offsets ✅ Isabelle thm:robust->rem:unified prop:cost Proposition 3 Fixed phase: the exact minimal exposure cost, from the shortest Dubins U-turn ◐ part in Isabelle prop:cost->rem:corr prop:slow2->prop:noiseloss prop:slow2->rem:corr prop:slow2->rem:unified lem:planar->prop:noiseloss rem:relax Remark 9 Relaxation of the qubit too: the bound and the sufficient condition for the square pulse ✅ Isabelle lem:planar->rem:relax lem:planar->rem:corr rem:planar->prop:noiseloss rem:planar->rem:relax prop:noiseloss->rem:relax rem:relax->rem:unified rem:corr->rem:unified
📐 Setting (copied from the paper, §§1–2)

A complete transfer of one excitation from a qubit to a memory mode, with no losses, a complex coupling \(|g(t)|\le G\) and any detuning \(D(t)\), of duration \(T\) with \(U(T)|a\rangle \propto |b\rangle \). The exposures and the leakage are the entries of \(K=\int _0^T U^\dagger (t)P_aU(t)\, dt=\begin{pmatrix} A & \bar\zeta \\ \zeta & B \end{pmatrix}\), \(A+B=T\): \(A\) and \(B\) are the times spent in the qubit and in the memory, and \(\zeta \) is the first-order leakage under a static shift \(\Delta _0\) of the qubit level (\(U(T)\mapsto U(T)(I-i\Delta _0K)+O(\Delta _0^2)\)). The square (resonant) pulse has \(T_0=\pi /(2G)\), \(A_0=\pi /(4G)\) and \(z_0=1/(2G)\). With slow Gaussian offsets of variance \(\sigma _q^2\) on the qubit and memory decay at rate \(\gamma _m\), \(\Lambda =\sigma _q^2/(\gamma _mG)\).

Abstract.

A quasi-static offset of the qubit or memory frequency, the slowest part of the noise in a superconducting quantum memory, is described by no Markovian rate. For the exchange of one excitation between a qubit and a memory mode at a coupling bounded by \(G\), we characterise every complete transfer by the times \(A\) and \(B\) it spends in the qubit and in the memory and by its first-order leakage \(\zeta \). For a coupling of any phase and any detuning the duration obeys \(T\ge T_0+\arccos (|\zeta |/z_0)/G\) for \(|\zeta |\le z_0\) and the exposures obey \((|\zeta |-z_0)^2\le (A-A_0)(B-A_0)\), with \(T_0=\pi /(2G)\), \(A_0=\pi /(4G)\), \(z_0=1/(2G)\); both are sharp, and for a coupling of fixed phase and no detuning they follow from known results on paths of bounded curvature. Cancelling the leakage with a pulse of fixed phase and no detuning costs exactly \(({\gamma _q}+{\gamma _m})/(2G)\cdot \arccos (|{\gamma _q}-{\gamma _m}|/({\gamma _q}+{\gamma _m}))\) of first-order population loss above that of the square pulse (a third of it in average infidelity), through the shortest Dubins U-turn; for any pulse, with relaxation only, cancelling it costs at least \(\max (\sqrt{{\gamma _q}{\gamma _m}},\tfrac \pi 2\min ({\gamma _q},{\gamma _m}))/G\) in the same units; curves made of pieces in planes through the axis do no better in space. At second order in the offsets every pulse has error \([2\, \mathbb E(\xi _q-\xi _m)^2|\zeta |^2+\mathbb E(\xi _qA+\xi _mB)^2]/6\): with noise on the qubit no pulse goes below \(\pi ^2\sigma ^2/(96G^2)\), against \((8+\pi ^2)\sigma ^2/(96G^2)\) for the square pulse; with noise on both elements the square pulse is optimal for \(0.137\le \sigma _m^2/\sigma _q^2\le 7.30\) and explicit shaped pulses beat it outside \((0.1264,7.91)\) (certified by interval arithmetic); for pulses of fixed phase and no detuning, a planar exposure bound derived from the reachable set of the Dubins car (through our reading of Patsko and Fedotov’s description of it) replaces the relaxation, and the threshold ratio is exactly the crossing of the bang–bang family, \(0.12645352248541\ldots \); with memory loss the square pulse stops being optimal (for fixed phase and no detuning, exactly at \(\sigma ^2=7.9078851606875\ldots \, {\gamma _m}G\), the crossing of a short-wait turn–wait–turn pulse), and for large \(\sigma ^2/({\gamma _m}G)\) a pulse that waits in the memory attains the optimal error \(\pi ^2\sigma ^2/(96G^2)+\sigma \sqrt{\pi {\gamma _m}}/(6G^{3/2})\) to leading order; positively correlated offsets on the two elements favour the square pulse, with an elementary sufficient condition that also covers relaxation of both elements, and the band of optimality of the square pulse can only widen as the correlation grows; for fixed phase and no detuning we certify the exact band with correlated offsets and the exact threshold with qubit relaxation at several parameter values, prove, given the planar bound, that this threshold grows as \(2\sqrt r+\pi /2+4/\sqrt3+O(r^{-1/2})\) in the ratio \(r\) of qubit to memory relaxation, and above the lossy threshold the optimum is a turn–wait–turn pulse. The robustness inequalities for the space curve, the minimisation of the cost, the unitarity step of the second-order expansion and the algebra of the bounds are machine-checked in Isabelle/HOL; Section 4 says where each formal statement ends and the written proof begins.

tabular xltabular longtable


[enhanced,breakable,before skip=8pt,after skip=4pt,colback=accent!4,colframe=accent!60,boxrule=0.6pt,arc=3pt,left=4pt,right=8pt,top=4pt,bottom=4pt,title=Results at a glance,coltitle=white,colbacktitle=accent!85,fonttitle=]

  • Robustness costs time and exposure: the frontier between duration and leakage and an inequality between the times in the qubit and in the memory, for a coupling of any phase and any detuning (Theorem 1; for a coupling of fixed phase and no detuning they follow from Souères–Balluchi–Bicchi and Patsko–Fedotov, and the minimum time at zero leakage is due to Van Damme et al.). for the space curve

  • The exact cost of cancelling the leakage with a pulse of fixed phase and no detuning (Proposition 3); pieces in planes through the axis do not shorten the U-turn. , comparison

  • The second-order error of every pulse, a floor \(\pi ^2\sigma ^2/(96G^2)\) for noise on the qubit, the square pulse optimal or beaten in certified ranges, and the gain of waiting in a lossy memory (Propositions 5 and 8). , ranges

  • Exact thresholds for pulses of fixed phase and no detuning: the square pulse loses its optimality exactly at the crossing of a competing family, \(\varrho _c=0.12645352248541\ldots \) and \(\Lambda _c=7.9078851606875\ldots \), given a planar exposure bound from the reachable set of the Dubins car (Lemma 6); above \(\Lambda _c\) the optimum among these pulses is a member of the turn–wait–turn family (Proposition 8(iii)). ,

  • Relaxation and correlation: with qubit relaxation the exact threshold \(\Lambda _c({\gamma _q}/{\gamma _m})\), certified at nine ratios, grows from \(7.908\) to \(67.245\) (Remark 9); a common offset leaves the square pulse optimal, opposite offsets are cancelled by a closed curve, and in between the exact band \([\varrho _c(\kappa ),1/\varrho _c(\kappa )]\) widens with the correlation but never covers every ratio (Remark 10); one elementary condition covers both cases (Remark 11). ,

  • Scope: two states, one excitation, a static offset, first and second order; the limits of the exchange under Markovian relaxation and dephasing are in the companion paper [ 10 ] .

machine-checked in Isabelle/HOL   numbers enclosed by interval arithmetic
proof in the text, audited step by step and tested numerically

1 Introduction

A quantum memory attached to a superconducting qubit is written by exchanging one excitation between the qubit and a long-lived mode at a coupling rate bounded by \(G\) [ 10 ] :

\begin{equation} \tag {1} i\dot a=\bigl(D(t)-\tfrac {i}{2}{\gamma _q}\bigr)a+g(t)\, b,\qquad i\dot b=-\tfrac {i}{2}{\gamma _m}\, b+\overline{g(t)}\, a,\qquad |g(t)|\le G , \end{equation}
1

with a complex coupling \(g\) and a detuning \(D\) as controls. The companion paper [ 10 ] gives the exact limits of this exchange under Markovian relaxation and dephasing and shows that, there, no pulse transfers more population than the resonant swap. Slow noise is different: a frequency offset that stays constant during one transfer is described by no rate, and pulse shaping can then help. The error of the write is measured, as in , by the average infidelity \(r=(1-p)/6+(1-c)/3\) of the channel with transferred population \(p\) and transferred coherence \(c\).

This paper treats a static offset at first and second order. Section 2 sets up the exposures \(A\), \(B\) and the leakage \(\zeta \) of a lossless transfer, proves the frontier between duration and leakage and the inequality between exposures (Theorem 1), computes the exact cost of cancelling the leakage with a pulse of fixed phase and no detuning (Proposition 3), and gives the second-order error of every pulse, alone and with memory loss (Propositions 5 and 8). Section 3 states what is not covered, Section 4 what backs each statement, and Appendix A collects the formulas.

Earlier work. Van Damme, Ansel, Glaser and Sugny found the minimum time \(\pi /G\) of an inversion robust to first order [ 18 ] . We use the space-curve picture of quantum control of Barnes and coworkers [ 20 , 3 , 19 ] , the leakage whose vanishing Daems et al. [ 5 ] and Lu et al. [ 9 ] impose for robust inversion, and the geometry of curves of bounded curvature: the shortest paths of Dubins [ 6 ] and of Souères, Balluchi and Bicchi [ 15 ] , the reachable sets of Patsko and Fedotov [ 13 , 14 ] , and the estimate on angles in the proof of Schur’s comparison theorem [ 16 ] . Composite pulses with a detuning that alternates in sign [ 8 ] are robust to errors in the pulse area, a different requirement: their two-pulse inversion transfers completely but keeps a first-order leakage \(|\zeta |\approx 1.11\, z_0\), so it is not robust to a static offset in our sense (our check, consistent with Theorem 1). What is new and what is not is said in detail in Section 2.

2 Robustness costs time and exposure

Bloch curve of the Heisenberg-evolved projector; projection on an equatorial crossing; Cauchy–Schwarz on exposures

2.1 Exposures, leakage and the two robustness bounds

Under slow noise the relevant quantities of a complete transfer are time integrals of the projector on the qubit in the Heisenberg picture. Take no losses, a control with \(|g|\le G\) and any detuning \(D(t)\), and a duration \(T\) with \(U(T)|a\rangle \propto |b\rangle \). Put

\[ K=\int _0^T U^\dagger (t)P_aU(t)\, dt=\begin{pmatrix} A & \bar\zeta \\ \zeta & B \end{pmatrix},\qquad A+B=T . \]

\(A\) and \(B\) are the times spent in the qubit and in the memory, and \(\zeta \) is the first-order leakage: a static shift \(\Delta _0\) of the qubit level changes \(U(T)\) into \(U(T)(I-i\Delta _0K)+O(\Delta _0^2)\), so the amplitude left in the qubit is \(\Delta _0|\zeta |\) to first order and the phase of the transferred amplitude moves by \(\Delta _0A\). Write \(U^\dagger P_aU=\tfrac 12(I+\mathbf n\cdot \boldsymbol \sigma )\); then \(\mathbf n(0)=\mathbf e_z\), \(\mathbf n(T)=-\mathbf e_z\), \(|\dot{\mathbf n}|=2|g|\le \lambda :=2G\) (the detuning commutes with \(P_a\) and does not enter), \(A=\tfrac 12\int (1+n_z)\), \(B=\tfrac 12\int (1-n_z)\) and \(|\zeta |=\tfrac 12|\int \mathbf n_\perp |\). Let \(T_0=\pi /(2G)\), \(A_0=\pi /(4G)\) and \(z_0=1/(2G)\), the values of the square pulse. The curve \(\mathbf r(t)=\int _0^t\mathbf n\) is the one of the space-curve picture of quantum control [ 20 , 3 ] : its curvature is \(2|g(t)|\), its length is the duration, and minimum times become shortest curves under a curvature bound [ 19 , 12 ] . For gates the first-order robustness asks the curve to close; for the population of a transfer only its lateral displacement \(2|\zeta |\) counts (writing an arbitrary qubit state also keeps the longitudinal phase exposure \(A\)), and \(\zeta =0\) is the condition for robust inversion against a static detuning of Daems et al. [ 5 ] and Lu et al. [ 9 ] .

Theorem 1✅ Isabelle for the space curve
📕 p. 4📍 map❝ cite⚠ report

MarAthm:popsame-setting Robust_Disk.robust_disk, Robust_Disk.robust_disk_G, Robust_Disk.exposure_bound_general, Robust_Core.angle_bound, Robust_Core.equatorial_crossing, Robust_Core.frontier_general, Memory_Robust.concavity_step, Memory_Robust.tent_integral, Memory_Robust.frontier_step, Memory_Robust.disk_from_two_bounds, Memory_Robust.bang_bang_equality, Memory_Robust.turn_wait_turn_equality, Memory_Robust.robust_cost_amgm, Memory_Robust.min_robust_A For every complete transfer without losses, with \(|g|\le G\) and any detuning:

  1. \(A\ge A_0\), \(B\ge A_0\) and

    \[ \bigl(|\zeta |-z_0\bigr)^2\le (A-A_0)(B-A_0); \]
  2. if \(|\zeta |\le z_0\), then

    \[ T\ge T_0+\frac1G\arccos \frac{|\zeta |}{z_0}. \]

(b) is sharp for every \(|\zeta |\le z_0\): equality holds for \(g=+G\) during \((\pi +d)/(2G)\) followed by \(g=-G\) during \(d/(2G)\), \(d=\arccos (|\zeta |/z_0)\), and for its time reverse. (a) is an equality on the branch \(|\zeta |\ge z_0\): for \(+G\) during \(\alpha /(2G)\), no coupling and no detuning during \(w\), and \(+G\) again during \((\pi -\alpha )/(2G)\), for every \(\alpha \in [0,\pi ]\) and \(w\ge 0\) (the square pulse is \(w=0\)).

Isabelle · oursshow

robust_disk · Isabelle/MemoryAccessFloors/Robust_Disk.thy, line 677

theorem robust_disk:
  assumes "sphere_curve n n' T nu"
  shows "pi / (2 * nu) \<le> integral {0..T} (\<lambda>t. (1 + n t $ 3) / 2)"
    and "pi / (2 * nu) \<le> integral {0..T} (\<lambda>t. (1 - n t $ 3) / 2)"
    and "(sqrt ((integral {0..T} (\<lambda>t. n t $ 1))\<^sup>2 + (integral {0..T} (\<lambda>t. n t $ 2))\<^sup>2) / 2 - 1 / nu)\<^sup>2
      \<le> (integral {0..T} (\<lambda>t. (1 + n t $ 3) / 2) - pi / (2 * nu)) *
         (integral {0..T} (\<lambda>t. (1 - n t $ 3) / 2) - pi / (2 * nu))"

robust_disk_G · Isabelle/MemoryAccessFloors/Robust_Disk.thy, line 694

theorem robust_disk_G:
  assumes "0 < G" and "sphere_curve n n' T (2 * G)"
  shows "pi / (4 * G) \<le> integral {0..T} (\<lambda>t. (1 + n t $ 3) / 2)"
    and "pi / (4 * G) \<le> integral {0..T} (\<lambda>t. (1 - n t $ 3) / 2)"
    and "(sqrt ((integral {0..T} (\<lambda>t. n t $ 1))\<^sup>2 + (integral {0..T} (\<lambda>t. n t $ 2))\<^sup>2) / 2 - 1 / (2 * G))\<^sup>2
      \<le> (integral {0..T} (\<lambda>t. (1 + n t $ 3) / 2) - pi / (4 * G)) *
         (integral {0..T} (\<lambda>t. (1 - n t $ 3) / 2) - pi / (4 * G))"

exposure_bound_general · Isabelle/MemoryAccessFloors/Robust_Disk.thy, line 667

theorem exposure_bound_general:
  assumes "sphere_curve n n' T nu"
  shows "integral {0..T} (\<lambda>t. sqrt ((n t $ 1)\<^sup>2 + (n t $ 2)\<^sup>2)) - 2 / nu \<le>
    2 * sqrt ((integral {0..T} (\<lambda>t. (1 + n t $ 3) / 2) - pi / (2 * nu)) *
              (integral {0..T} (\<lambda>t. (1 - n t $ 3) / 2) - pi / (2 * nu)))"

angle_bound · Isabelle/MemoryAccessFloors/Robust_Core.thy, line 925

theorem angle_bound:
  assumes "bounded_speed_curve n n' T nu"
    and "0 \<le> s" "s \<le> T" "0 \<le> t" "t \<le> T" and "nu * \<bar>s - t\<bar> \<le> pi"
  shows "cos (nu * \<bar>s - t\<bar>) \<le> inner (n s) (n t)"

equatorial_crossing · Isabelle/MemoryAccessFloors/Robust_Core.thy, line 934

theorem equatorial_crossing:
  assumes "sphere_curve n n' T nu"
  shows "\<exists>ts. 0 \<le> ts \<and> ts \<le> T \<and> n ts $ 3 = 0"

frontier_general · Isabelle/MemoryAccessFloors/Robust_Core.thy, line 942

theorem frontier_general:
  assumes "sphere_curve n n' T nu" and "nu * T - pi \<le> pi"
  shows "cos ((nu * T - pi) / 2) \<le>
    nu * (sqrt ((integral {0..T} (\<lambda>t. n t $ 1))\<^sup>2 + (integral {0..T} (\<lambda>t. n t $ 2))\<^sup>2) / 2)"

concavity_step · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 20

lemma concavity_step:
  fixes s a :: real
  assumes "0 \<le> s" and "s \<le> pi" and "0 \<le> a" and "a \<le> s"
  shows "1 + cos (s / 2) \<le> cos (a / 2) + cos ((s - a) / 2)"

tent_integral · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 51

lemma tent_integral:
  fixes nu a ts :: real
  assumes "0 < nu" and "0 \<le> a" and "ts = (pi / 2 + a) / nu"
  shows "((\<lambda>t. cos (min (nu * (ts - t)) (pi / 2 + nu * t))) has_integral
           (2 * cos (a / 2) - 1) / nu) {0..ts}"

frontier_step · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 143

lemma frontier_step:
  fixes G nu s zeta z0 T T0 :: real
  assumes "0 < G" and "nu = 2 * G" and "T0 = pi / (2 * G)" and "z0 = 1 / (2 * G)"
    and "0 \<le> s" and "s \<le> pi" and "0 \<le> zeta" and "zeta \<le> z0"
    and "T = (pi + s) / nu" and "cos (s / 2) \<le> zeta / z0"
  shows "T0 + arccos (zeta / z0) / G \<le> T"

disk_from_two_bounds · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 170

lemma disk_from_two_bounds:
  fixes G nu zeta z0 Q X Y :: real
  assumes "0 < G" and "nu = 2 * G" and "z0 = 1 / (2 * G)"
    and "\<bar>2 * zeta - 2 / nu\<bar> \<le> Q - 2 / nu" and "Q - 2 / nu \<le> 2 * sqrt (X * Y)"
    and "0 \<le> X" and "0 \<le> Y"
  shows "(zeta - z0)\<^sup>2 \<le> X * Y"

bang_bang_equality · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 187

lemma bang_bang_equality:
  fixes G d T T0 A0 z0 zeta A B :: real
  assumes "0 < G" and "0 \<le> d" and "d \<le> pi / 2"
    and "T0 = pi / (2 * G)" and "A0 = pi / (4 * G)" and "z0 = 1 / (2 * G)"
    and "T = (pi + 2 * d) / (2 * G)" and "zeta = cos d / (2 * G)"
    and "A = A0 + (d - sin d) / (2 * G)" and "B = A0 + (d + sin d) / (2 * G)"
  shows "A + B = T"
    and "T = T0 + arccos (zeta / z0) / G"
    and "(zeta - z0)\<^sup>2 \<le> (A - A0) * (B - A0)"

turn_wait_turn_equality · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 220

lemma turn_wait_turn_equality:
  fixes G L alpha z0 A0 zeta A B :: real
  assumes "0 \<le> L" and "z0 = 1 / (2 * G)" and "A0 = pi / (4 * G)"
    and "zeta = z0 + L * sin alpha / 2"
    and "A = A0 + L * (1 + cos alpha) / 2" and "B = A0 + L * (1 - cos alpha) / 2"
  shows "(zeta - z0)\<^sup>2 = (A - A0) * (B - A0)" and "A + B = 2 * A0 + L"

robust_cost_amgm · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 237

lemma robust_cost_amgm:
  fixes G gq gm x y :: real
  assumes "0 < G" and "0 \<le> gq" and "0 \<le> gm" and "0 \<le> x" and "0 \<le> y"
    and "1 / (4 * G\<^sup>2) \<le> x * y"
  shows "sqrt (gq * gm) / G \<le> gq * x + gm * y"

min_robust_A · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 263

lemma min_robust_A:
  fixes G d A A0 zeta :: real
  assumes "0 < G" and "d = pi / 2" and "A0 = pi / (4 * G)"
    and "A = A0 + (d - sin d) / (2 * G)" and "zeta = cos d / (2 * G)"
  shows "A = (pi - 1) / (2 * G)" and "zeta = 0"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: the whole statement·Cite: Theorem 1, p. 4, doi:10.5281/zenodo.24250503

↳ Used in: Corollary 2, 4, Proposition 5, Lemma 6, 7, Proposition 8, 10, 11

Proof ▼

Choose a time \(t_e\) with \(n_z(t_e)=0\) and put \(\mathbf e=\mathbf n(t_e)\), \(x(t)=\mathbf n(t)\cdot \mathbf e\). Since \(\mathbf e\) is equatorial, \(\int x\, dt=\mathbf e\cdot \int \mathbf n_\perp \le 2|\zeta |\). The angle \(\psi (t)\) between \(\mathbf n(t)\) and \(\mathbf e\) is at most \(\lambda |t-t_e|\), and, because the end points are the poles, at most \(\pi /2+\lambda t\) and \(\pi /2+\lambda (T-t)\); in particular \(\lambda t_e\ge \pi /2\) and \(\lambda (T-t_e)\ge \pi /2\).

(b) Put \(\alpha =\lambda t_e-\pi /2\) and \(\beta =\lambda (T-t_e)-\pi /2\), so \(\alpha +\beta =\lambda T-\pi =:\Theta \), and assume \(\Theta \le \pi \) (otherwise there is nothing to prove). The bound \(\bar\psi (t)=\min (\lambda |t-t_e|,\ \pi /2+\lambda t,\ \pi /2+\lambda (T-t))\) is at most \(\pi /2+\max (\alpha ,\beta )/2\le \pi \), where the cosine decreases, so \(x=\cos \psi \ge \cos \bar\psi \); integrating this piecewise linear angle on each side of \(t_e\) gives

\[ 2|\zeta |\ \ge \ \int _0^Tx\, dt\ \ge \ \frac2\lambda \Bigl(\cos \frac\alpha 2+\cos \frac\beta 2-1\Bigr)\ \ge \ \frac2\lambda \cos \frac\Theta 2, \]

the last step because \(\alpha \mapsto \cos (\alpha /2)+\cos ((\Theta -\alpha )/2)\) is concave on \([0,\Theta ]\) and so is smallest at an end point. With \(\lambda =2G\) this reads \(\cos (G(T-T_0))\le |\zeta |/z_0\), which is (b).

(a) On \(|t-t_e|\le \pi /(2\lambda )\) the speed bound gives \(x(t)\ge \cos (\lambda |t-t_e|)\ge 0\), so \(\int x_+\ge 2/\lambda \). With \(q=|\mathbf n_\perp |\ge |x|\) and \(Q=\int q\), this gives \(Q+2|\zeta |\ge \int (q+x)\ge 2\int x_+\ge 4/\lambda \), and with \(2|\zeta |\le Q\) we get \(\bigl|2|\zeta |-2/\lambda \bigr|\le Q-2/\lambda \). Let \(\theta =\arccos n_z\), the spherical distance of \(n\) from the north pole, which is \(\lambda \)-Lipschitz and hence absolutely continuous (the formal proof avoids \(\theta \) near the poles with a comparison argument, so that curves through a pole are covered), and \(h=1-\dot\theta /\lambda \ge 0\); the fundamental theorem of calculus gives \(\int (1\pm \cos \theta )\dot\theta =\pi \) and \(\int \sin \theta \, \dot\theta =2\) whether or not \(\theta \) is monotone, so \(A-\pi /(2\lambda )=\tfrac 12\int (1+\cos \theta )h\), \(B-\pi /(2\lambda )=\tfrac 12\int (1-\cos \theta )h\) and \(Q-2/\lambda =\int \sin \theta \, h\). Cauchy–Schwarz gives \(Q-2/\lambda \le 2\sqrt{(A-\pi /(2\lambda ))(B-\pi /(2\lambda ))}\), and the two inequalities give (a).

The equality cases are direct computations: for the bang–bang pulse of (b), \(T=(\pi +2d)/(2G)\), \(|\zeta |=\cos d/(2G)\), \(A=A_0+(d-\sin d)/(2G)\) and \(B=A_0+(d+\sin d)/(2G)\); its time reverse exchanges the two values of \(A-A_0\) and \(B-A_0\).

An unknown static frequency offset leaves a first-order error measured by the leakage \(\zeta \). Reducing it has a price that no pulse escapes: the exposures \(A,B\) must exceed \(A_0\) enough to satisfy \((|\zeta |-z_0)^2\le (A-A_0)(B-A_0)\), and the duration must exceed \(T_0\) by \(\arccos (|\zeta |/z_0)/G\). A fully robust transfer (\(\zeta =0\)) therefore takes at least \(\pi /G\), twice the resonant swap.

🛡️ Evidence & scope
  • Literature: Companion paper: this statement shares the setting of paper A (Exact limits on writing a superconducting quantum memory), Theorem 4 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: Figure 1 live: design a pulse and see where it falls against the disk and the frontier of the theorem

Integrates the lossless two-level dynamics exactly, piece by piece, for a pulse you choose (units \(G=1\), so \(z_0=1/2\), \(A_0=\pi /4\), \(T_0=\pi /2\)): the bang–bang pulse of parameter \(d\), the turn–wait–turn pulse \((\alpha ,w)\) of equality (a), the pulse \(C_{\pi +\phi }SC_\phi \) (a turn of \(\pi +\phi \), a wait, a turn of \(\phi \) back), or a random piecewise pulse with complex coupling \(|g|\le u\, G\) and detuning whose last piece is built to finish the transfer. It reads \(A\), \(B\), \(|\zeta |\) and \(T\) from \(K=\int U^\dagger P_aU\) and checks the two statements of the theorem: \((|\zeta |-z_0)^2\le (A-A_0)(B-A_0)\) and \(T\ge T_0+\arccos (|\zeta |/z_0)/G\). In 2D, duration against \(|\zeta |/z_0\) with the frontier, the turn–wait–turn line and 150 random pulses; in 3D, the point in \((A-A_0,B-A_0,|\zeta |)\) between the two sheets \(|\zeta |=z_0\pm \sqrt{(A-A_0)(B-A_0)}\), the bang–bang curves (orange, equality in (b)) and the turn–wait–turn lines (blue, equality in (a)). With \(u{\gt}1\) the hypothesis \(|g|\le G\) is broken and the theorem fails: that is the control.

Together, (a) and (b) give the whole frontier between duration and leakage: the smallest \(|\zeta |\) at duration \(T\in [T_0,\pi /G]\) is \(z_0\cos (G(T-T_0))\), and \(\zeta =0\) is reachable for every \(T\ge \pi /G\); at prescribed \(|\zeta |\ge z_0\), (a) and \(x+y\ge 2\sqrt{xy}\) give \(T\ge T_0+2(|\zeta |-z_0)\), attained by the pulse of the theorem with \(\alpha =\pi /2\). On the branch \(|\zeta |{\lt}z_0\), (a) is not attained by these pulses: at \(\zeta =0\), among pulses of fixed phase and no detuning and at fixed \(A-B\), the smallest product \((A-A_0)(B-A_0)\) is \((2\pi /3)^2z_0^2\approx 4.39\, z_0^2\) for \(A=B\) and \((\pi ^2/4-1)z_0^2\approx 1.47\, z_0^2\) for the pulses of least duration, and tends to \(z_0^2\) only as \(|A-B|\) grows without bound.

What is known. With \(|g|\le G\) the curve \(\mathbf r\) has unit speed and curvature at most \(\lambda \): it is a Dubins curve of minimum radius \(z_0=1/\lambda \), and a complete transfer is a U-turn, with lateral displacement \(2\zeta \) and axial displacement \(A-B\). Zeng and Barnes found the fastest robust phase gates as shortest closed plane curves of bounded curvature [ 19 ] ; the two ends of (b) are known values, \(T_0\) for \(|\zeta |=z_0\) (the \(\pi \) pulse, the elementary polar-angle speed limit for population inversion under \(|g|\le G\); related gate speed limits under a bounded spectral width of the Hamiltonian are derived in [ 12 ] ) and \(\pi /G\) for \(\zeta =0\) (below), and (b) is the exact curve between them. For \(g=e^{i\varphi }u(t)\) with \(u\) real, \(\varphi \) constant and no detuning (or a detuning compensated by the phase of \(g\)) the curve is planar; in general it leaves the plane. In the planar case (b) is the case of reversed heading of the shortest path of a bounded-curvature vehicle to a directed line with free arrival point, synthesised by Souères, Balluchi and Bicchi [ 15 ] (their paths \(C_aC_b\) at the edge \(a=b+\pi R\), which give the length in closed form), and (a) is, in substance, a consequence of the description of the reachable set of the Dubins car by Patsko and Fedotov [ 13 , 14 ] : at heading \(\pi \) the circle bounding (a) is their arc \(A_1\), which lies on the boundary of the section for every duration, the other boundary arcs lie inside its disk [ 14 , Property 8 and Theorem 11.1 ] , and for headings \(\pi +2\pi k\), \(k\ge 1\), the section is a smaller disk with the same centre; they do not write the inclusion as an inequality. The case \(\zeta =0\) of (b), \(T\ge \pi /G\) with the bang–bang pulse \((3\pi /(4G),\pi /(4G))\), is the minimum time of a first-order robust inversion found by Van Damme et al. [ 18 ] , already with two controls, by the maximum principle and a numerical search, and rederived for a real control by Ansel, Glaser and Sugny [ 1 ] . What the proof above adds is a complete elementary proof of that minimum, both statements for a coupling of any phase and any detuning, where the curve leaves the plane, (a) written as an inequality between exposures, and the reading of both in terms of leakage: (a) says how the extra time must be shared between the two subsystems. The bound on the angle to \(\mathbf n(t_e)\) is the step of the classical proof of Schur’s comparison theorem for curves of bounded curvature [ 16 ] , so (b) is an adaptation of that argument to an antipodal U-turn. In the sources on bounded-curvature curves in space that we read, neither statement appears: for the reachable set with prescribed final direction in \(\mathbb R^3\), Bae et al. [ 2 , Theorem 3 of the PDF v3, Theorem 8 of the HTML version ] give the possible boundary extremals (pieces \(CCC\), \(CSC\), their subsegments, or helicoidal arcs) but leave the explicit sections, the counterpart of [ 14 ] , for future work, so the antipodal-section inequalities used here are not there; and Sussmann’s classification of the shortest paths in space, a helicoidal arc or three pieces each a circle or a segment [ 17 , Theorem 1 ] , does not say when a helicoidal arc can be excluded, even for coplanar end data such as those of a U-turn. In space the planar paths need not be optimal: for displacements with equal end tangents a curve leaving the plane is shorter [ 11 ] .

🖼️ Figure 1
Theorem 1 as a picture. Left: in the coordinates \((A-A_0,B-A_0,|\zeta |)\) (units of \(1/G\)) every lossless complete transfer lies between the two sheets \(|\zeta |=z_0\pm \sqrt{(A-A_0)(B-A_0)}\) (translucent); dots: 400 random piecewise-constant pulses with complex coupling \(|g|\le G\) and detu…

Theorem 1 as a picture. Left: in the coordinates \((A-A_0,B-A_0,|\zeta |)\) (units of \(1/G\)) every lossless complete transfer lies between the two sheets \(|\zeta |=z_0\pm \sqrt{(A-A_0)(B-A_0)}\) (translucent); dots: 400 random piecewise-constant pulses with complex coupling \(|g|\le G\) and detuning (those inside the box shown), none of which violates (a) or (b) (smallest margins \(5\cdot 10^{-4}\) and \(2\cdot 10^{-2}\)). Orange: the bang–bang family, equality in (b); blue: turn–wait–turn pulses, equality in (a) for \(|\zeta |\ge z_0\). Right: duration against \(|\zeta |/z_0\), with the frontier \(T_0+\arccos (|\zeta |/z_0)/G\) of (b) and the turn–wait–turn line beyond \(z_0\).

Memory_Robust.robust_cost_amgm, Robust_Cost_Branches.robust_cost_min_branch, Robust_Cost_Branches.robust_cost_both_branches, Robust_Cost_Branches.robust_cost_equal_rates A complete transfer with \(\zeta =0\) satisfies

\[ {\gamma _q}(A-A_0)+{\gamma _m}(B-A_0)\ \ge \ \frac1G\max \Bigl(\sqrt{{\gamma _q}{\gamma _m}},\ \frac\pi 2\min ({\gamma _q},{\gamma _m})\Bigr). \]

With relaxation only (no dephasing), to first order in the rates the write channel loses \(({\gamma _q}A+{\gamma _m}B)/3\), and the square pulse, which does not cancel \(\zeta \), reaches the floor \(({\gamma _q}+{\gamma _m})A_0/3\); so cancelling the leakage costs at least a third of the right-hand side above that floor, whatever the pulse. For equal rates \(\gamma \) the bound is \(\pi \gamma /(2G)\) and is attained by the bang–bang pulse \((3\pi /(4G),\pi /(4G))\); the branch \(\sqrt{{\gamma _q}{\gamma _m}}/G\) is not attained by pulses of fixed phase and no detuning (see above).

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

robust_cost_amgm · Isabelle/MemoryAccessFloors/Memory_Robust.thy, line 237

lemma robust_cost_amgm:
  fixes G gq gm x y :: real
  assumes "0 < G" and "0 \<le> gq" and "0 \<le> gm" and "0 \<le> x" and "0 \<le> y"
    and "1 / (4 * G\<^sup>2) \<le> x * y"
  shows "sqrt (gq * gm) / G \<le> gq * x + gm * y"

robust_cost_min_branch · Isabelle/MemoryAccessFloors/Robust_Cost_Branches.thy, line 23

theorem robust_cost_min_branch:
  fixes G gq gm x y :: real
  assumes "0 < G" "0 \<le> gq" "0 \<le> gm" "0 \<le> x" "0 \<le> y" "pi / (2 * G) \<le> x + y"
  shows "pi / 2 * min gq gm / G \<le> gq * x + gm * y"

robust_cost_both_branches · Isabelle/MemoryAccessFloors/Robust_Cost_Branches.thy, line 39

theorem robust_cost_both_branches:
  fixes G gq gm x y :: real
  assumes "0 < G" "0 \<le> gq" "0 \<le> gm" "0 \<le> x" "0 \<le> y"
    and "1 / (4 * G\<^sup>2) \<le> x * y" and "pi / (2 * G) \<le> x + y"
  shows "max (sqrt (gq * gm)) (pi / 2 * min gq gm) / G \<le> gq * x + gm * y"

robust_cost_equal_rates · Isabelle/MemoryAccessFloors/Robust_Cost_Branches.thy, line 57

theorem robust_cost_equal_rates:
  fixes G g :: real
  assumes "0 < G" "0 \<le> g"
  shows "max (sqrt (g * g)) (pi / 2 * min g g) / G = pi * g / (2 * G)"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: part: the named pieces are machine-checked; the statement as a whole is a written proof·Cite: Corollary 2, p. 6, doi:10.5281/zenodo.24250503

↳ Not used by another result of the paper.

Proof ▼

By (a) with \(\zeta =0\), \((A-A_0)(B-A_0)\ge 1/(4G^2)\), and \({\gamma _q}u+{\gamma _m}v\ge 2\sqrt{{\gamma _q}{\gamma _m}\, uv}\). By (b) with \(\zeta =0\), \((A-A_0)+(B-A_0)=T-2A_0\ge \pi /(2G)\), and \({\gamma _q}u+{\gamma _m}v\ge \min ({\gamma _q},{\gamma _m})(u+v)\). The bang–bang pulse has \(T=\pi /G\). At first order \(1-p={\gamma _q}A+{\gamma _m}B\), since \(A=\int |a|^2\) and \(B=\int |b|^2\) along the lossless transfer, and \(1-c=\tfrac 12({\gamma _q}A+{\gamma _m}B)\), so \(r=(3-p-2c)/6=({\gamma _q}A+{\gamma _m}B)/3\).

2.2 The exact cost of cancelling the leakage

For pulses of fixed phase and no detuning the cost of cancelling the leakage can be computed exactly. With \(\zeta =0\) the curve \(\mathbf r\) is a planar U-turn back to its own axis, and by Dubins’ theorem [ 6 ] the shortest one with axial displacement \(s\) is a path of type \(CSC\) or \(CCC\). In units of \(z_0\), with \(q=\arcsin (\sqrt{4+s^2}/4)\) and \(s\ge 0\) by symmetry, the triangle of centres (sides \(2\), \(2\), \(\sqrt{4+s^2}\)) and the condition on the final heading give the eight candidates: \(LSL\) and \(RSR\) of length \(3\pi +\sqrt{4+s^2}\); \(LSR\) and \(RSL\), for \(s\ge 2\), of length \(2\pi -2\arctan (\sqrt{s^2-4}/2)+\sqrt{s^2-4}\); and, for \(s^2\le 12\), two pairs of \(CCC\) paths of lengths \(3\pi \mp 4q\) when \(s{\lt}2\), and \(5\pi -4q\) and \(\pi +4q\) when \(s{\gt}2\) (each checked against the integrated path for \(s\ne 2\)). At \(s=2\) the end arcs degenerate and the shortest \(CCC\) representatives have length \(2\pi \); this end point is treated separately from the open branches. For \(s{\lt}2\) the shortest is \(3\pi -4q{\lt}3\pi \). For \(s{\gt}2\), \(LSR\) has length at most \(2\pi +\sqrt{s^2-4}{\lt}3\pi \) when \(s^2\le 12\), which excludes \(LSL\) and \(5\pi -4q\); and \(f(s)=4q-\pi +2\arctan (\sqrt{s^2-4}/2)-\sqrt{s^2-4}\) tends to \(0\) as \(s\to 2\) with \(f'(s)=4s/\sqrt{(12-s^2)(4+s^2)}-\sqrt{s^2-4}/s{\gt}0\), since the first term is at least \(1\) exactly when \(s^2\ge 4\); so \(\pi +4q\ge LSR\). Beyond \(s^2=12\) only the \(CSC\) paths exist. Hence \(CCC\) for \(|s|\le 2z_0\) and \(CSC\) beyond, and the length is the least duration whose section at heading \(\pi \) in [ 14 ] contains the point \((s,0)\), which they do not compute. In units of \(z_0\) it is

\[ L(s)=\begin{cases} 3\pi -4\arcsin \dfrac {\sqrt{4+s^2}}4, & |s|\le 2,\\[6pt] 2\pi -2\arctan \dfrac {\sqrt{s^2-4}}2+\sqrt{s^2-4}, & |s|\ge 2. \end{cases} \]
🖼️ Figure 2
Proposition 3. Left: the curve \(\mathbf r(t)=\int \mathbf n\) of the shortest planar U-turns with \(\zeta =0\) for axial displacement \(s=0,1,\dots ,4\, z_0\) (\(CCC\) for \(s{\lt}2z_0\), \(CSC\) beyond); each ends on the axis at height \(s\). Right: the minimum cost \(C^*=\frac{{\gamma _q}+{\gamm…

Proposition 3. Left: the curve \(\mathbf r(t)=\int \mathbf n\) of the shortest planar U-turns with \(\zeta =0\) for axial displacement \(s=0,1,\dots ,4\, z_0\) (\(CCC\) for \(s{\lt}2z_0\), \(CSC\) beyond); each ends on the axis at height \(s\). Right: the minimum cost \(C^*=\frac{{\gamma _q}+{\gamma _m}}{2G}\arccos \frac{|{\gamma _q}-{\gamma _m}|}{{\gamma _q}+{\gamma _m}}\) for pulses of fixed phase and no detuning (colours, \(\log _{10}C^*\), with \(C^*\) the dimensionless cost of Proposition 3; dashed: equal rates). Its ratio to the bound of Corollary 2 for every pulse depends only on \({\gamma _m}/{\gamma _q}\): it is \(1\) for equal rates and has maximum \((\pi /2+2/\pi )\arctan (2/\pi )\approx 1.25140958\), attained at \({\gamma _m}/{\gamma _q}=4/\pi ^2\) and \(\pi ^2/4\), where the two branches of that bound meet.

MarAthm:popcompares Robust_Cost.csc_derivative, Robust_Cost.closed_form_value, Robust_Cost.csc_minimum, Robust_Cost.ccc_slope, Robust_Cost.ccc_branch, Robust_Cost.equal_rates, Robust_Cost.total_minimum, Dubins_U_Turn.lstar_le_out, Dubins_U_Turn.lstar_le_lin, Dubins_U_Turn.lstar_le_ccc_a, Dubins_U_Turn.lstar_le_ccc_b, Dubins_U_Turn.ccc_b_above_lin Among complete transfers with \(\zeta =0\) by pulses \(g=e^{i\varphi }u(t)\), \(u\) real, \(\varphi \) constant, without detuning,

\[ \min \bigl[{\gamma _q}(A-A_0)+{\gamma _m}(B-A_0)\bigr]=\frac{{\gamma _q}+{\gamma _m}}{2G}\, \arccos \frac{|{\gamma _q}-{\gamma _m}|}{{\gamma _q}+{\gamma _m}}, \]

attained for \({\gamma _q}\ne {\gamma _m}\), both positive, by a turn of \(\pi +\phi \), a wait without coupling, and a turn of \(\phi \) back, with \(|A-B|=({\gamma _q}+{\gamma _m})/(2G\sqrt{{\gamma _q}{\gamma _m}})\), and for equal positive rates by the bang–bang pulse. If exactly one rate vanishes the value \(0\) is an infimum, approached only by ever longer waits and not attained; if both vanish the cost is identically zero.

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

csc_derivative · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 139

theorem csc_derivative:
  fixes u :: real
  assumes "2 < u"
  shows "(Lcsc has_real_derivative sqrt (u\<^sup>2 - 4) / u) (at u)"

closed_form_value · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 225

theorem closed_form_value:
  fixes a b :: real
  assumes "b < a" and "0 < b"
  shows "sqrt (((a + b) / sqrt (a * b))\<^sup>2 - 4) = (a - b) / sqrt (a * b)"
    and "cost a b ((a + b) / sqrt (a * b)) = (a + b) * arccos ((a - b) / (a + b))"

csc_minimum · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 435

theorem csc_minimum:
  fixes a b u :: real
  assumes "b < a" and "0 < b" and "2 \<le> u"
  shows "(a + b) * arccos ((a - b) / (a + b)) \<le> cost a b u"

ccc_slope · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 495

theorem ccc_slope:
  fixes u :: real
  assumes "0 \<le> u" and "u \<le> 2"
  shows "(Lccc has_real_derivative - 4 * u / sqrt ((12 - u\<^sup>2) * (4 + u\<^sup>2))) (at u)"
    and "\<bar>- 4 * u / sqrt ((12 - u\<^sup>2) * (4 + u\<^sup>2))\<bar> \<le> 1"
    and "- 4 * u / sqrt ((12 - u\<^sup>2) * (4 + u\<^sup>2)) \<le> 0"

ccc_branch · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 595

theorem ccc_branch:
  fixes a b u :: real
  assumes "0 \<le> b" and "b \<le> a" and "0 \<le> u" and "u \<le> 2"
  shows "cost a b 2 \<le> cost a b u" and "cost a b 2 = (a + b) * pi / 2 - (a - b)"

equal_rates · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 618

theorem equal_rates:
  fixes a u :: real
  assumes "0 < a" and "0 \<le> u"
  shows "pi * a \<le> cost a a u"
    and "cost a a 2 = pi * a"
    and "(a + a) * arccos ((a - a) / (a + a)) = pi * a"

total_minimum · Isabelle/MemoryAccessFloors/Robust_Cost.thy, line 652

theorem total_minimum:
  fixes a b u :: real
  assumes "b < a" and "0 < b" and "0 \<le> u"
  shows "(a + b) * arccos ((a - b) / (a + b)) \<le> cost a b u"
    and "cost a b ((a + b) / sqrt (a * b)) = (a + b) * arccos ((a - b) / (a + b))"

lstar_le_out · Isabelle/MemoryAccessFloors/Dubins_U_Turn.thy, line 514

theorem lstar_le_out:
  fixes s :: real
  assumes "0 \<le> s"
  shows "Lstar s \<le> Lout s"

lstar_le_lin · Isabelle/MemoryAccessFloors/Dubins_U_Turn.thy, line 527

theorem lstar_le_lin:
  fixes s :: real
  assumes "2 \<le> s"
  shows "Lstar s \<le> Lin s"

lstar_le_ccc_a · Isabelle/MemoryAccessFloors/Dubins_U_Turn.thy, line 540

theorem lstar_le_ccc_a:
  fixes s :: real
  assumes "0 \<le> s" and "s \<le> sqrt 12"
  shows "Lstar s \<le> Lccc_a s"

lstar_le_ccc_b · Isabelle/MemoryAccessFloors/Dubins_U_Turn.thy, line 553

theorem lstar_le_ccc_b:
  fixes s :: real
  assumes "0 \<le> s" and "s \<le> sqrt 12"
  shows "Lstar s \<le> Lccc_b s"

ccc_b_above_lin · Isabelle/MemoryAccessFloors/Dubins_U_Turn.thy, line 478

theorem ccc_b_above_lin:
  fixes s :: real
  assumes "2 \<le> s" and "s \<le> sqrt 12"
  shows "Lin s \<le> Lccc_b s"

folded_curve · Isabelle/MemoryAccessFloors/Meridian_Unfolding.thy, line 615

theorem folded_curve: "sphere_curve m m' T nu"

folded_planar · Isabelle/MemoryAccessFloors/Meridian_Unfolding.thy, line 628

theorem folded_planar: "0 \<le> t \<Longrightarrow> t \<le> T \<Longrightarrow> m t $ 2 = 0"

folded_displacement · Isabelle/MemoryAccessFloors/Meridian_Unfolding.thy, line 631

theorem folded_displacement: "integral {0..T} m = integral {0..T} n"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: part: the named pieces are machine-checked; the statement as a whole is a written proof·Cite: Proposition 3, p. 8, doi:10.5281/zenodo.24250503

↳ Used in: 10

Proof ▼

With \(A-B=s\) and \(A+B\ge L(s)\), the cost is at least \(f(s)=\tfrac 12[({\gamma _q}+{\gamma _m})L(s)+({\gamma _q}-{\gamma _m})s]-({\gamma _q}+{\gamma _m})A_0\) (in units \(z_0=1\)). \(L\) is continuous, and differentiable except at \(|s|=2\): at \(s=2\) its slope is \(-1\) from the left and \(0\) from the right, and symmetrically at \(s=-2\). On \(|s|{\gt}2\), \(L'(s)=\operatorname {sgn}(s)\sqrt{s^2-4}/|s|\), so \(f\) is convex there with its minimum where \(\sqrt{s^2-4}/|s|=|{\gamma _q}-{\gamma _m}|/({\gamma _q}+{\gamma _m})\), i.e. \(|s|=({\gamma _q}+{\gamma _m})/\sqrt{{\gamma _q}{\gamma _m}}\); at that point the algebraic terms cancel and \(f=({\gamma _q}+{\gamma _m})\arctan (2/\sqrt{s^2-4})\), which is the stated arccosine. On \(|s|{\lt}2\), \(|L'(s)|=4|s|/\sqrt{(12-s^2)(4+s^2)}\le 1\) (equivalent to \(s^2\le 4\)), so on the favourable side, where \(({\gamma _q}-{\gamma _m})s\le 0\), \(f\) decreases towards \(|s|=2\); and since \(L\) is even, \(f(s)-f(-s)=({\gamma _q}-{\gamma _m})s\), so the other side is no better.

Among pulses of fixed phase and no detuning, the extra loss that full robustness forces, \(\gamma _q(A-A_0)+\gamma _m(B-A_0)\) at \(\zeta =0\), is computed exactly: it is the length of the shortest Dubins U-turn of the space curve, minimised over one parameter.

🛡️ Evidence & scope
  • Literature: Companion paper: this statement is compared with paper A (Exact limits on writing a superconducting quantum memory), Theorem 4 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: Figure 2 live: the cost of cancelling the leakage with fixed phase, and the curve r(t) of the optimal turn

For the rates \(\gamma _q,\gamma _m\) (in units of \(G\)) you choose, the cost \(\gamma _q(A-A_0)+\gamma _m(B-A_0)\) of a transfer with \(\zeta =0\) and fixed phase, as a function of the axial displacement \(s=(A-B)/z_0\) of the shortest Dubins U-turn (the least of the eight candidates \(LSL,RSR,LSR,RSL,CCC\) of the text), with its minimum found numerically, against the closed form \(C^*=\frac{\gamma _q+\gamma _m}{2G}\arccos \frac{|\gamma _q-\gamma _m|}{\gamma _q+\gamma _m}\) and the bound for every pulse of the corollary. The pulse that attains it (a turn of \(\pi +\phi \), a wait \(\cot \phi /G\), a turn of \(\phi \) back) is integrated exactly and its cost read from \(A,B\). In 3D, the curve \(\mathbf r(t)=\int \mathbf n\, dt\) drawn as time runs (its length is the duration), the open U-turn of the square pulse, and the same curve with the coupling phase turned: only the plane through the axis changes, nothing else.

For equal rates this is \(\pi \gamma /(2G)\), the value of Corollary 2, which holds for every pulse. For unequal rates we do not know whether a coupling of varying phase, or a detuning, can do better. The question is whether, with \(\zeta =0\), a non-planar path, for instance one with a helicoidal arc in Sussmann’s sense [ 17 ] , can make a shorter U-turn than \(L(s)\). Two numerical searches over curves that leave the plane (24 rotations of free axis with 30 starts, and a direct minimisation of the length at fixed \(s\) for \(0\le s\le 4\)) found no improvement beyond their tolerances: the first accepts end-state and leakage residuals up to \(10^{-4}\), compares costs to \(10^{-3}\), and for equal rates lands slightly below the formula within them; the second falls below \(L(s)\) by the same amount, \(3\cdot 10^{-3}\) or less, as at the control \(s=2\), where the planar path is proved optimal. This is evidence, not a proof; and in space planar paths are not always the shortest [ 11 ] . Mittenhuber’s mechanism, two planar pieces turned about the common tangent at their junction, does not apply here. Meridian_Unfolding.folded_curve, Meridian_Unfolding.folded_planar, Meridian_Unfolding.folded_displacement The data of the U-turn lie on the axis, so pieces in distinct planes through the axis can only meet on the axis with tangent \(\pm e_z\). Each piece can then be turned about the axis into one common plane without changing length, curvature or end data: a curve made of such pieces is no shorter than \(L(s)\). The unfolding is machine-checked () for a finite partition in which, on each piece, both the position \(\mathbf r\) and the tangent \(\mathbf n\) lie in one plane containing the axis, with the junction conditions, the final displacement on the axis and the pointwise differentiability hypotheses of the curve; that only \(\mathbf n\) be meridian on each piece is not enough, since turning the pieces could destroy a transverse cancellation that occurs only at the end. The comparison with \(L(s)\) then follows from the written planar Dubins argument. Pieces in planes that do not contain the axis, and helicoidal arcs, are not covered.

Two representative pulses cancel \(\zeta \) in the least time \(\pi /G\) (controls related to them by a constant phase of \(g\), or by a diagonal gauge transformation with a compensating detuning, have the same curve and exposures): the bang–bang pulse (\(+G\) during \(3\pi /(4G)\), \(-G\) during \(\pi /(4G)\)), with \(A=(\pi -1)/(2G)\) and \(B=(\pi +1)/(2G)\), and its time reverse, with the two values exchanged. Both exceed \(A_0\) in each exposure; the better one, when the qubit is the noisier element, spends \((\pi -1)/(2G)\) in the qubit, more than the square pulse, and so worsens the dephasing of the superposition with the vacuum. Robustness of the population is not the same as robustness of the write channel.

↳ Not used by another result of the paper.

2.3 Slow noise at second order

The exposures also give the effect of slow noise on the write channel at second order, for every pulse. Let the offsets \(\xi _q,\xi _m\) (static shifts of the qubit and memory levels, as in ) be centred, with \(\sigma _q^2,\sigma _m^2\) their second moments.

MarAprop:qsextendsMarAcor:channeluses Unitarity_Second_Order.unitary_coefficients, Unitarity_Second_Order.unitarity_aa, Unitarity_Second_Order.transfer_modulus, Slow_Noise_Second_Order.second_order_algebra,Slow_Noise_Second_Order.second_order_combination, Slow_Noise_Second_Order.square_value, Slow_Noise_Second_Order.qubit_only_floor, Slow_Noise_Second_Order.min_time_robust_value, Slow_Noise_Kernel.robust_worse_than_square_kernel, Slow_Noise_Kernel.square_vs_floor_kernel, Slow_Noise_Second_Order.csc_A_excess, Slow_Noise_Kernel.bang_bang_beats_square_kernel For every complete transfer without losses, with \(|g|\le G\) and any detuning, to second order in the offsets

\[ r=\frac{2\, \mathbb E(\xi _q-\xi _m)^2\, |\zeta |^2+\mathbb E(\xi _qA+\xi _mB)^2}{6}. \]
  1. For the square pulse this is the bound of , \((8+\pi ^2)(\sigma _q^2+\sigma _m^2)/(96G^2)\) for uncorrelated offsets: that bound is exact at this order.

  2. With noise on the qubit only, \(r\ge \pi ^2\sigma _q^2/(96G^2)\) for every pulse. The square pulse is \(1.81\) times this floor; the better of the two pulses of least duration that cancel \(\zeta \) gives \((\pi -1)^2\sigma _q^2/(24G^2)\), \(2.66\, \% \) above the square pulse. The floor is the infimum of the second-order coefficient: zero-leakage pulses that spend their extra time in the memory approach it only as \(T\to \infty \), outside the regime \(\sigma _qT\ll 1\) where the expansion holds (with memory loss, see Proposition 8).

  3. With independent offsets on both elements and \(\varrho =\sigma _m^2/\sigma _q^2\): for \(\varrho \le 0.1264\) the bang–bang pulse of Theorem 1(b) with \(d=0.78\) does better than the square pulse (by \(3.3\, \% \) at \(\varrho =0.1\) with \(d=0.848\)), and for \(\varrho \ge 7.91\) its time reverse does (certified by interval arithmetic, directly at \(7.91\)). The square pulse is optimal at this order among all pulses for \(0.137\le \varrho \le 7.30\) (the upper end certified through \(10/73\), whose reciprocal is \(7.3\)) (certified by interval arithmetic over the region allowed by Theorem 1; a floating-point minimisation over that relaxed region suggests that this certificate is lost near \(0.1357\), which is not established as a physical threshold). (The problem is symmetric: \(F/\varrho \) is the same problem with \(1/\varrho \) and \(A\), \(B\) exchanged.)

  4. For pulses of fixed phase and no detuning the set of ratios at which the square pulse is optimal at this order is a closed interval \([\varrho _c,1/\varrho _c]\), and \(\varrho _c\) is the crossing of the bang–bang family:

    \[ \varrho _c=\varrho _{\rm bb}:=\inf \{ \varrho :\ \min _d\, [F_{\rm bb}(\varrho ,d)-F_{\rm sq}(\varrho )]\ge 0\} =0.12645352248540975\ldots \]

    (enclosed in an interval of width \(10^{-12}\); certified by interval arithmetic, using the planar exposure bound of Lemma 6). Below \(\varrho _{\rm bb}\) the bang–bang pulse with \(d\) near \(d_*=0.7717755295532\ldots \) is strictly better, and its time reverse for the reciprocal ratios. For pulses of varying phase or with detuning, \(0.1264{\lt}\varrho {\lt}0.137\) and \(7.30{\lt}\varrho {\lt}7.91\) remain undecided.

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

unitary_coefficients · Isabelle/MemoryAccessFloors/Unitarity_Second_Order.thy, line 297

theorem unitary_coefficients:
  fixes X1 X2 :: cm and R :: "real \<Rightarrow> cm" and \<delta> K :: real
  assumes "\<delta> > 0" and "\<And>e. \<bar>e\<bar> \<le> \<delta> \<Longrightarrow> cm_bd K (R e)"
    and "\<And>e. \<bar>e\<bar> \<le> \<delta> \<Longrightarrow> cmul (cadj (cV X1 X2 (R e) e)) (cV X1 X2 (R e) e) = cone"
  shows "cadj X1 = X1"
    and "cadd X2 (cadj X2) = cmul X1 X1"

unitarity_aa · Isabelle/MemoryAccessFloors/Unitarity_Second_Order.thy, line 316

theorem unitarity_aa:
  fixes X1 X2 :: cm
  assumes "cadj X1 = X1" and "cadd X2 (cadj X2) = cmul X1 X1"
  shows "Im (a11 X1) = 0"
    and "2 * Re (a11 X2) = (norm (a21 X1))\<^sup>2 + (Re (a11 X1))\<^sup>2"

transfer_modulus · Isabelle/MemoryAccessFloors/Unitarity_Second_Order.thy, line 352

theorem transfer_modulus:
  fixes X1 X2 :: cm and e :: real
  assumes "cadj X1 = X1" and "cadd X2 (cadj X2) = cmul X1 X1"
  shows "(norm (1 - \<i> * of_real e * a11 X1 - (of_real e)\<^sup>2 * a11 X2))\<^sup>2
    = 1 - e\<^sup>2 * (norm (a21 X1))\<^sup>2
      + e^3 * (2 * Re (a11 X1) * Im (a11 X2) + e * (norm (a11 X2))\<^sup>2)"

second_order_algebra · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 25

lemma second_order_algebra:
  fixes e f z w :: real
  assumes "0 \<le> z" and "1 - 2 * w + f\<^sup>2 = 1 - e\<^sup>2 * z\<^sup>2"
  shows "2 * w = e\<^sup>2 * z\<^sup>2 + f\<^sup>2"

second_order_combination · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 31

lemma second_order_combination:
  fixes Ee2 Ef2 z P C p c :: real
  assumes "P = Ee2 * z\<^sup>2" and "C = Ee2 * z\<^sup>2 + Ef2" and "p = 1 - P" and "c = 1 - C / 2"
  shows "(3 - p - 2 * c) / 6 = (2 * (Ee2 * z\<^sup>2) + Ef2) / 6"

square_value · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 39

lemma square_value:
  fixes sq2 sm2 G z0 A0 :: real
  assumes "G > 0" and "sq2 \<ge> 0" and "sm2 \<ge> 0" and "z0 = 1 / (2 * G)" and "A0 = pi / (4 * G)"
  shows "(2 * (sq2 + sm2) * z0\<^sup>2 + sq2 * A0\<^sup>2 + sm2 * A0\<^sup>2) / 6
         = (8 + pi\<^sup>2) * (sq2 + sm2) / (96 * G\<^sup>2)"

qubit_only_floor · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 48

lemma qubit_only_floor:
  fixes sigma2 z A G :: real
  assumes "G > 0" and "sigma2 \<ge> 0" and "A \<ge> pi / (4 * G)"
  shows "(2 * sigma2 * z\<^sup>2 + sigma2 * A\<^sup>2) / 6 \<ge> pi\<^sup>2 * sigma2 / (96 * G\<^sup>2)"

min_time_robust_value · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 65

lemma min_time_robust_value:
  fixes sigma2 G A :: real
  assumes "G > 0" and "A = (pi - 1) / (2 * G)"
  shows "(2 * sigma2 * 0\<^sup>2 + sigma2 * A\<^sup>2) / 6 = (pi - 1)\<^sup>2 * sigma2 / (24 * G\<^sup>2)"

robust_worse_than_square_kernel · Isabelle/MemoryAccessFloors/Slow_Noise_Kernel.thy, line 25

lemma robust_worse_than_square_kernel: "(pi - 1)\<^sup>2 / 24 > (8 + pi\<^sup>2) / 96"

square_vs_floor_kernel · Isabelle/MemoryAccessFloors/Slow_Noise_Kernel.thy, line 39

lemma square_vs_floor_kernel: "(8 + pi\<^sup>2) / pi\<^sup>2 > 1.81" and "(8 + pi\<^sup>2) / pi\<^sup>2 < 1.82"

csc_A_excess · Isabelle/MemoryAccessFloors/Slow_Noise_Second_Order.thy, line 85

lemma csc_A_excess:
  fixes u r :: real
  assumes "u > 2" and "r = sqrt (u\<^sup>2 - 4)"
  shows "0 \<le> arctan (2 / r) + (r - u) / 2"
    and "arctan (2 / r) + (r - u) / 2 \<le> 2 / (u + r)"
    and "2 / (u + r) \<le> 2 / u"
    and "arctan (2 / r) + (r - u) / 2 \<le> 2 / u"

bang_bang_beats_square_kernel · Isabelle/MemoryAccessFloors/Slow_Noise_Kernel.thy, line 86

lemma bang_bang_beats_square_kernel:
  "F_cost_k (1 / 10) (cos (106 / 125)) (pi / 2 + 106 / 125 - sin (106 / 125))
            (pi / 2 + 106 / 125 + sin (106 / 125))
   < F_cost_k (1 / 10) 1 (pi / 2) (pi / 2)"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: part: the named pieces are machine-checked; the statement as a whole is a written proof·Cite: Proposition 5, p. 9, doi:10.5281/zenodo.24250503

↳ Used in: Proposition 8, 10, 11

Proof ▼

A static shift changes \(U(T)\) into \(U(T)(I-i\Xi _1-\Xi _2+\dots )\), with \(\Xi _1=\xi _qK+\xi _m(T-K)\), \(\Xi _2\) quadratic in the offsets and a remainder of third order, bounded uniformly for small offsets. With \(U(T)|a\rangle =e^{i\phi }|b\rangle \) the transferred amplitude is \(t=e^{i\phi }\bigl(1-i(\xi _qA+\xi _mB)-(\Xi _2)_{aa}+\dots \bigr)\) and the leaked one is \((\xi _q-\xi _m)\zeta \) to first order; unitarity, \(|t|^2=1-(\xi _q-\xi _m)^2|\zeta |^2\), fixes \(2\operatorname {Re}(\Xi _2)_{aa}=(\xi _q-\xi _m)^2|\zeta |^2+(\xi _qA+\xi _mB)^2\). For each fixed pulse and duration, write \(\xi =\epsilon X\) with \(X\) of fixed law and finite second moments: the averaged formula below then holds with a remainder \(o(\epsilon ^2)\); a cubic averaged remainder needs a further moment assumption. Averaging, the first-order terms vanish, \(p=1-\mathbb E(\xi _q-\xi _m)^2|\zeta |^2\) and \(c=|\mathbb Et|=1-\tfrac 12\mathbb E\bigl[(\xi _q-\xi _m)^2|\zeta |^2+(\xi _qA+\xi _mB)^2\bigr]\), and \(r=(3-p-2c)/6\) gives the formula. (i) is \(|\zeta |=z_0\), \(A=B=A_0\). (ii): \(A\ge A_0\) by Theorem 1(a); the minimum-time pulses have \(A=(\pi \mp 1)/(2G)\); and the zero-leakage pulses \(C_{\pi +\phi }SC_\phi \) (a turn of \(\pi +\phi \), a wait, a turn of \(\phi \) back) have \(A-A_0=O(1/(G^2T))\). (iii): the lower bound minimises \(2(\sigma _q^2+\sigma _m^2)|\zeta |^2+\sigma _q^2A^2+\sigma _m^2B^2\) over the region allowed by Theorem 1, \(A,B\ge A_0\), (a), (b) and \(T\ge T_0+2(|\zeta |-z_0)\). The difference \(g\) between a point of this region and the square pulse is linear in \(\varrho \), so it suffices to show \(g\ge 0\) at \(\varrho =0.137\) and at \(\varrho =1\); the region is symmetric in \(A\) and \(B\), and dividing by \(\varrho \) gives \(1\le \varrho \le 1/0.137\). For \(|\zeta |\ge z_0\) every term of \(g\) is non-negative; \(g\) increases with \(A-A_0\) and \(B-A_0\), so the latter can be taken at its least admissible value; near \(|\zeta |=z_0\), \(\arccos (|\zeta |/z_0)\ge \sqrt{2(1-|\zeta |/z_0)}\) gives \(g\ge 0\) by a one-line estimate; the rest of the region, a rectangle in \((|\zeta |,A-A_0)\), is covered by 551 accepted leaf boxes (1101 nodes of the subdivision evaluated; 554 leaves at \(\varrho =10/73\)) on each of which an interval lower bound of \(g\) is non-negative (at \(\varrho =0.137\) computed twice, with mpmath and with Arb in SageMath; at the exact rational \(\varrho =10/73\), which gives the decimal endpoint \(7.30\), with the Python certificate). The other side evaluates one bang–bang pulse, \(d=0.78\): its difference with the square pulse is linear in \(\varrho \) and negative at \(\varrho =0\) and \(\varrho =0.1264\) (enclosed by intervals), and its time reverse covers \(\varrho \ge 1/0.1264\approx 7.9114\); at \(\varrho =7.91\) the reversed pulse is checked directly (difference \(-7.6\cdot 10^{-4}\), slope \(-0.74\), both enclosed). The analytic patch and the subdivided region share the same exactly represented endpoint, rational parameters are enclosed directly, and every coefficient used as a lower bound is rounded downwards.

(iv): for every pulse \(P\) the difference \(g_P(\varrho )\) with the square pulse is affine in \(\varrho \), so the set \(\bigcap _P\{ \varrho {\gt}0:g_P(\varrho )\ge 0\} \) is a closed interval; it contains \(1\) and is invariant under \(\varrho \mapsto 1/\varrho \), hence it is \([\varrho _c,1/\varrho _c]\). For pulses of fixed phase and no detuning the curve \(\mathbf r\) is planar, and the region of Theorem 1 can be replaced by the bound of Lemma 6: for \(|\zeta |{\lt}z_0\) and \(A-A_0\) below its bang–bang value, \(B-A_0\) is at least that of the family; symmetrically with \(A\) and \(B\) exchanged. Since \(F\) grows with \(A\) and with \(B\), the minimum over that branch lies on the family, a two-parameter set covered by intervals (605 accepted leaves, in the variables \((\phi ,w)\), which avoid the singularity at \(\phi \to 0\)). The rest is covered by the relaxed region (22 322 leaves in all at \(\varrho =0.12645395\)). By linearity in \(\varrho \) it suffices to add \(\varrho =1\), and the symmetry gives the upper end: \(\varrho _c\le 0.12645395\). At \(0.126453\) the bang–bang witness \(d=0.7717755\) is strictly better (difference \(-3.0\cdot 10^{-6}\), enclosed), and its difference increases with \(\varrho \), so it is better on all of \([0,0.126453]\) and \(\varrho _c{\gt}0.126453\); certificar_umbral_plano.py. The identification \(\varrho _c=\varrho _{\rm bb}\) is a second certificate (certificar_rho_c_igual_bb.py). A Krawczyk operator in \((\varrho ,d)\) proves that \(\{ h=0,\ \partial _dh=0\} \), \(h=F_{\rm bb}-F_{\rm sq}\), has a unique zero in a box, with enclosure widths below \(6\cdot 10^{-41}\) for \(\varrho _{\rm bb}\) and \(4\cdot 10^{-40}\) for \(d_*\) (regular Jacobian: \(\partial _\varrho h=5.80\), \(\partial _d^2h=2.54\)). Since \(h(\varrho ,0)=0\) always, at the crossing the global minimum of \(h\) is zero and is attained both at \(d=0\) and at the unique nonzero minimiser \(d=d_*\); throughout the enclosing interval \(E\) of \(\varrho \), the critical branch \(d(\varrho )\) is the unique critical point in \(U=[d_*\pm 0.01]\), and above the crossing it need not be the global minimiser of \(h\). On \(N=U\times [0,0.02]\) in the family variables \((\phi ,w)\) the Hessian of \(g\) is positive definite (by intervals), uniformly for \(\varrho \in E\). At \(w=0\) the exact identities \(g_\phi =h_d\) and \(g_w=h_d/2\) make the full gradient vanish at \((d(\varrho ),0)\); since \(N\) is convex, Taylor’s formula gives \(g\ge h(\varrho ,d(\varrho ))\) on \(N\), including its edge \(w=0\). Outside \(N\), the analytic near-square patch and the large-wait tail are treated separately; on the remaining retained branch boxes the interval lower bound is at least \(5\cdot 10^{-7}\) at both endpoints of \(E\), hence throughout \(E\) by affinity (equality at the square pulse is allowed in the analytic patch). The reflected branch is certified independently. For the middle region put \(x_0=d-\sin d\) and \(y_0=d+\sin d\) (in the units of Lemma 6): if \(x_0\le a\le y_0\), monotonicity in \(b\) and Theorem 1(b) reduce the cost to \(g(a,2d-a)\), which increases in \(a\) for \(\varrho \le \tfrac 14\); if \(a\ge y_0\), then \(g(a,b)\ge g(y_0,x_0)\ge g(x_0,y_0)\) because \(\varrho \le 1\); so the middle region is bounded below by \(h(\varrho ,d)\) without subdivision. Hence the square pulse is optimal at \(\varrho _{\rm bb}\); since \(h(\varrho _{\rm bb},d_*)=0\) and \(\partial _\varrho h{\gt}0\) at every positive \(d\), the same nonzero bang–bang pulse is strictly better for every \(\varrho {\lt}\varrho _{\rm bb}\). Controls (a neighbourhood too small, a ratio \(10^{-4}\) lower, an off-centre Krawczyk box) fail.

With slow random offsets \(\xi _q,\xi _m\) on both elements, the average error of a transfer at second order is \(r=[2\, \mathbb E(\xi _q-\xi _m)^2|\zeta |^2+\mathbb E(\xi _qA+\xi _mB)^2]/6\) for every pulse: the pulse enters only through \(\zeta \), \(A\) and \(B\), and the bounds of the theorem decide when the square pulse is optimal.

🛡️ Evidence & scope
  • Literature: Companion paper: this statement extends paper A (Exact limits on writing a superconducting quantum memory), Proposition 14 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

  • Literature: Companion paper: this statement uses paper A (Exact limits on writing a superconducting quantum memory), Corollary 9 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: Figure 3 live: when the bang–bang pulse beats the square pulse, and J against Λ with memory loss

Left (the second-order proposition): the relative change \((F-F_{\rm sq})/F_{\rm sq}\) of the second-order coefficient \(F=2(1+\varrho )|\zeta |^2+A^2+\varrho B^2\) when the square pulse is replaced by the bang–bang pulse of parameter \(d\), against \(d\) for the ratio \(\varrho =\sigma _m^2/\sigma _q^2\) you choose, and over \((\varrho ,d)\) in 3D (the slider sweeps \(\varrho \) and prints the best \(d\)): where the surface is below the orange plane the bang–bang pulse wins. Right (the proposition with memory loss, memory loss only, \(\Lambda =\sigma _q^2/(\gamma _mG)\)): \(J/J_{\rm square}\) for the best zero-leakage pulse \(C_{\pi +\phi }SC_\phi \) (the shortest U-turn of every displacement) and the lower bound for every pulse (minimum of \(J\) over the region allowed by Theorem (a), (b)), with the exact fixed-phase threshold \(\Lambda _c\) computed in the lab. All of it is calculated here in units \(G=1/2\); only ratios matter.

What was known. That a quasi-static shift of a level which the protocol must populate cannot be cancelled to first order is known. For the controlled-phase gate of Rydberg atoms with a bounded Rabi frequency, Jandura, Thompson and Pupillo [ 7 ] show that the first-order error is a phase equal to the detuning times the time spent in the Rydberg state, which is positive for every pulse, and that at second order the infidelity splits into a term fixed by those times and a term from the population left behind, which a longer composite pulse removes while the first stays; they minimise the first term numerically. Bergonzoni, Jandura and Pupillo [ 4 ] state the obstruction as general criteria, without bounds on the controls and without values. In the write of a memory the same two terms are those of the formula above, in \(A^2\) and in \(|\zeta |^2\); what (ii) adds is the closed floor \(\pi ^2\sigma _q^2/(96G^2)\) for every pulse with \(|g|\le G\), through \(A\ge A_0\).

2.4 Fixed phase and no detuning: the planar exposure bound

Lemma 6 planar exposure bound
📕 p. 11📍 map❝ cite⚠ report

Planar_Exposure_Bound.planar_exposure_bound, Planar_Exposure_Bound.planar_exposure_bound_inner, Planar_Exposure_Bound.arc_bounds, Planar_Exposure_Bound.arc_attains, Planar_Exposure_Bound.a_z_strict_mono, Planar_Exposure_Bound.theta_z_antimono, Planar_Exposure_Bound.multi_turn_dominated, Memory_Floors_Pieces.a_z_deriv, Memory_Floors_Pieces.b_z_deriv, Memory_Floors_Pieces.a_z_range, Memory_Floors_Pieces.odd_turns_integral For a pulse of fixed phase and no detuning, put \(a=2G(A-A_0)\), \(b=2G(B-A_0)\), \(z=|\zeta |/z_0\in [0,1)\) and \(d=\arccos z\), and for \(0{\lt}\phi \le d\)

\[ a_z(\phi )=\phi -(1+z)\tan \tfrac \phi 2,\qquad b_z(\phi )=\phi +(1-z)\cot \tfrac \phi 2 . \]

Then \(a_z'=(\cos \phi -z)/(1+\cos \phi )\ge 0\), \(b_z'=(z-\cos \phi )/(1-\cos \phi )\le 0\), \(a_z\) maps \((0,d]\) onto \((0,d-\sin d]\), and every such pulse with \(0{\lt}a\le d-\sin d\) has

\[ b\ \ge \ b_z\bigl(a_z^{-1}(a)\bigr), \]

and symmetrically with \(a\) and \(b\) exchanged. This is the type-\((1,0,-1)\) family of the remark below, written at fixed \(|\zeta |\).

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

planar_exposure_bound · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 433

theorem planar_exposure_bound:
  fixes z a b :: real
  assumes z: "0 \<le> z" "z < 1" and a: "0 < a" "a \<le> arccos z - sin (arccos z)"
    and th: "arccos z \<le> (a + b) / 2"
    and arc: "a < arccos z - sin (arccos z) \<Longrightarrow>
              \<exists>phi. 0 < phi \<and> phi < arccos z \<and> a - b = x3 ((a + b) / 2) phi
                      \<and> y3 ((a + b) / 2) phi \<le> 2 * z"
  shows "\<exists>beta. 0 < beta \<and> beta \<le> arccos z \<and> a_z z beta = a
                \<and> (\<forall>beta'. 0 < beta' \<longrightarrow> beta' \<le> arccos z \<longrightarrow> a_z z beta' = a \<longrightarrow> beta' = beta)
                \<and> b_z z beta \<le> b"

planar_exposure_bound_inner · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 391

theorem planar_exposure_bound_inner:
  fixes z a b :: real
  assumes z: "0 \<le> z" "z < 1" and a: "0 < a" "a < arccos z - sin (arccos z)"
    and arc: "\<exists>phi. 0 < phi \<and> phi < arccos z \<and> a - b = x3 ((a + b) / 2) phi
                      \<and> y3 ((a + b) / 2) phi \<le> 2 * z"
  shows "\<exists>beta. 0 < beta \<and> beta \<le> arccos z \<and> a_z z beta = a
                \<and> (\<forall>beta'. 0 < beta' \<longrightarrow> beta' \<le> arccos z \<longrightarrow> a_z z beta' = a \<longrightarrow> beta' = beta)
                \<and> b_z z beta \<le> b"

arc_bounds · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 358

theorem arc_bounds:
  fixes z phi theta :: real
  assumes "0 < phi" and "phi < pi" and "y3 theta phi \<le> 2 * z"
  shows "theta_z z phi \<le> theta" and "a_z z phi \<le> theta + x3 theta phi / 2"

arc_attains · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 335

theorem arc_attains:
  fixes z phi :: real
  assumes "0 \<le> z" and "z < 1" and "0 < phi" and "phi \<le> arccos z"
  defines "th \<equiv> theta_z z phi"
  shows "th + x3 th phi / 2 = a_z z phi" and "th - x3 th phi / 2 = b_z z phi"
    and "y3 th phi = 2 * z" and "0 \<le> th - phi"

a_z_strict_mono · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 181

theorem a_z_strict_mono:
  fixes z p1 p2 :: real
  assumes "0 \<le> z" and "z < 1" and "0 < p1" and "p1 < p2" and "p2 \<le> arccos z"
  shows "a_z z p1 < a_z z p2"

theta_z_antimono · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 219

theorem theta_z_antimono:
  fixes z p1 p2 :: real
  assumes "0 \<le> z" and "z < 1" and "0 < p1" and "p1 \<le> p2" and "p2 \<le> arccos z"
  shows "theta_z z p2 \<le> theta_z z p1"

multi_turn_dominated · Isabelle/MemoryAccessFloors/Planar_Exposure_Bound.thy, line 463

theorem multi_turn_dominated:
  fixes z :: real and k :: nat
  assumes "0 \<le> z" and "z < 1" and "1 \<le> k"
  shows "arccos z - sin (arccos z) < real k * pi" and "arccos z + sin (arccos z) < real k * pi"

a_z_deriv · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 423

theorem a_z_deriv:
  fixes z phi :: real
  assumes "- pi < phi" and "phi < pi"
  shows "(a_z z has_real_derivative (cos phi - z) / (1 + cos phi)) (at phi)"

b_z_deriv · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 443

theorem b_z_deriv:
  fixes z phi :: real
  assumes "0 < phi" and "phi < pi"
  shows "(b_z z has_real_derivative (z - cos phi) / (1 - cos phi)) (at phi)"

a_z_range · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 609

theorem a_z_range:
  fixes z y :: real
  assumes "0 \<le> z" and "z < 1" and "0 < y" and "y \<le> arccos z - sin (arccos z)"
  shows "\<exists>phi. 0 < phi \<and> phi \<le> arccos z \<and> a_z z phi = y"

odd_turns_integral · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 658

theorem odd_turns_integral:
  fixes k :: nat
  shows "((\<lambda>alpha. (cos (alpha / 2))\<^sup>2) has_integral ((2 * real k + 1) * pi / 2))
         {0..(2 * real k + 1) * pi}"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: part: the named pieces are machine-checked; the statement as a whole is a written proof·Cite: Lemma 6, p. 11, doi:10.5281/zenodo.24250503

↳ Used in: Proposition 8, 9, 10

Proof ▼

With fixed phase and no detuning the Bloch direction moves on one meridian, \(\mathbf n=(\sin \alpha ,0,\cos \alpha )\) with \(|\dot\alpha |\le 2G\), and \(\mathbf r\) is the path of a planar Dubins car of heading \(\alpha \); \(A\) and \(B\) are the times weighted by \(\cos ^2(\alpha /2)\) and \(\sin ^2(\alpha /2)\); the axial displacement is \(A-B\), and the magnitude of the lateral displacement is \(2|\zeta |\). After normalisation, \(t_f=\pi +a+b\), \(x=a-b\) and \(y=\pm 2z\). The write ends at the south pole, so the final heading is an odd multiple of \(\pi \). (1) Final headings \(\pm (2k+1)\pi \) with \(k\ge 1\): for \(f\ge 0\) and an absolutely continuous angle, \(2G\int _0^Tf(\alpha )\, dt\ge \int _0^Tf(\alpha )|\dot\alpha |\, dt\ge |\int _0^Tf(\alpha )\dot\alpha \, dt|\), so with the primitives of \(\cos ^2(\alpha /2)\) and \(\sin ^2(\alpha /2)\), \(A,B\ge A_0+k\pi /(2G)\), i.e. \(a,b\ge k\pi \), which exceeds the bang–bang values \(d\mp \sin d\) at the same \(|\zeta |\) in both exposures; such pulses are dominated by the bang–bang pulse (\(F\) grows with \(A\) and \(B\)). Equality \(a=b=k\pi \) forces a full-speed monotone turn and gives \(|\zeta |=z_0\), outside the sector \(z{\lt}1\); the square pulse has the same leakage and smaller exposures. (2) Heading \(\pi \) (by reflecting the control if necessary). Set \(\theta =(t_f-\pi )/2=(a+b)/2\); Theorem 1(b) gives \(\theta \ge d\). At heading \(\pi \) the classification of [ 14 , Lemma 3.1 and Theorem 11.1 ] gives their Case 1 for \(\pi {\lt}t_f{\lt}3\pi \) and Case 3 for \(t_f\ge 3\pi \); the non-simply connected Case 2 does not occur. The lower boundary on the negative-\(x\) side is the \(U_3=(1,0,-1)\) arc

\[ x_3=-2\sin \phi -2(\theta -\phi )\cos \phi ,\qquad y_3=2\cos \phi -2(\theta -\phi )\sin \phi . \]

In Case 1 the \(U_6\) branch satisfies \(a\ge \theta -\sin \theta \ge d-\sin d\) and does not enter the sector under consideration; in Case 3 it contributes no additional boundary arc. \(U_1\) has \(y\ge 2\), and \(U_2\) lies on the opposite axial side. The retained \(U_3\) arc includes \(0\le \phi \le d\): in Case 1 this follows from \(\theta \ge d\), and in Case 3 its first intersection with its reflected arc occurs after \(\phi =\pi /2\). Since \(x_3\) is increasing on the relevant arc, every reachable point in the strict sector \(a{\lt}d-\sin d\) lies above a unique \(U_3\) point at the same \(x\), with \(0{\lt}\phi {\lt}d\). (3) Both lateral signs are covered directly: \(y=\pm 2z\le 2z\) and \(y\ge y_3\) imply

\[ \theta \ \ge \ \theta _z(\phi ):=\phi +\frac{\cos \phi -z}{\sin \phi }, \]

and consequently \(a\ge a_z(\phi )\) (along the arc the wait is \(w=(\cos \phi -z)/(G\sin \phi )\) and the exposures are the closed forms \(a_z(\phi )\), \(b_z(\phi )\) of the remark, simplified). (4) Let \(\beta =a_z^{-1}(a)\). The strict monotonicity of \(a_z\) gives \(\phi \le \beta \), while \(\theta _z'(\phi )=\cos \phi \, (z-\cos \phi )/\sin ^2\phi \le 0\) for \(0{\lt}\phi \le d\); therefore \(b=2\theta -a\ge 2\theta _z(\beta )-a=b_z(\beta )\). At \(a=d-\sin d\) the same inequality follows from \(\theta \ge d\), and \(a=0\) is excluded by the disk bound for \(z{\lt}1\). Equality is attained by \(U_3\) with the stated non-negative wait. Finally, time reversal combined with \(\alpha \mapsto \pi -\alpha \) exchanges \(a\) and \(b\), giving the symmetric assertion. Steps (1), (3) and (4) are elementary; the passage from the sections of [ 14 ] , which are stated at fixed duration and unwound final heading, to exposures is our written derivation, not a lemma of that source. The closed forms were checked against piecewise pulses with exact integrals and against the parametrisation of the remark (to \(2\cdot 10^{-13}\), verificar_v289.py), and the inequality with both lateral signs on 7061 random turn–wait–turn pulses in the sector (verificar_v291.py).

Remark 7 the exact planar section
#
📕 p. 12📍 map❝ cite⚠ report

U_Turn_Exposure_Family.family_bang_bang, U_Turn_Exposure_Family.family_inside_disc, U_Turn_Exposure_Family.family_sum_above_boundary, U_Turn_Exposure_Family.family_x_increasing, U_Turn_Exposure_Family.family_y_decreasing, U_Turn_Exposure_Family.family_product_tends_to_hyperbola For a pulse of fixed phase and no detuning, the duration, \(\zeta \) and the axial displacement \(A-B\) are the time, the lateral position and the axial position of a planar Dubins car of radius \(z_0\) that turns back to heading \(\pi \). The set of such end points at a given duration is the section of the reachable set at heading \(\pi \), described exactly by Patsko and Fedotov [ 14 , Lemma 3.1 and Theorem 11.1 ] : its boundary is made of arcs of extremal controls of four types. In exposures, the part with \(|\zeta |{\lt}z_0\) and the smallest \(A-A_0\) is the arc of their type \((1,0,-1)\): a turn of \(\pi +\phi \) at full coupling, a wait \(w\), and a turn of \(\phi \) back,

\[ \begin{gathered} |\zeta |=z_0\bigl[\cos \phi -Gw\sin \phi \bigr],\\ A-A_0=\tfrac {1}{2G}\bigl[\phi -\sin \phi +Gw(1-\cos \phi )\bigr],\qquad B-A_0=\tfrac {1}{2G}\bigl[\phi +\sin \phi +Gw(1+\cos \phi )\bigr]. \end{gathered} \]

For prescribed \(z=|\zeta |/z_0\in [0,1)\) the branch is parametrised by \(0{\lt}\phi \le \arccos z\) and \(w=(\cos \phi -z)/(G\sin \phi )\ge 0\). At \(\phi =\arccos z\) (\(w=0\)) it is the bang–bang pulse of Theorem 1(b); at fixed \(z\), letting \(\phi \downarrow 0\) sends \(w\to \infty \) and gives \((A-A_0)(B-A_0)\to (z_0-|\zeta |)^2\), so the disk of Theorem 1(a) is approached along the branch only in that limit. This explains why the relaxation in the proof of (iii) stops at \(0.137\): its minimum lies where the disk and the frontier are both active, a point no pulse reaches. For \(0\le |\zeta |/z_0{\lt}1\) and \(0{\lt}\phi {\lt}\arccos (|\zeta |/z_0)\) the family satisfies both bounds of Theorem 1 strictly, \((A-A_0)(B-A_0){\gt}(z_0-|\zeta |)^2\) and \((A-A_0)+(B-A_0){\gt}2d/(2G)\) with \(d=\arccos (|\zeta |/z_0)\), \(A\) increases and \(B\) decreases with \(\phi \), and the product tends to \((z_0-|\zeta |)^2\); these facts are formalised from the closed forms above (the chain from the control of type \((1,0,-1)\) to these forms is the written computation).

We conjecture that the same holds for every pulse, of varying phase and with detuning: a numerical search over three-dimensional pulses with exact integrals, aimed at going below the family, lands on it in five values of \(|\zeta |\) and finds nothing below it beyond its constraint tolerance: over 40 starts per value the lowest margin is \(-8.7\cdot 10^{-8}\), at \(|\zeta |/z_0=0.664\) and \(2G(A-A_0)=0.019\), where the constraints are met to \(10^{-8}\) and the frontier is steep enough for that error to move the margin by up to about \(10^{-6}\) (conjetura_adversario.py). If it holds, the same planar exposure reduction, and hence the certified value of \(\varrho _c\) (and of \(\Lambda _c\) in Proposition 8), applies to all pulses.

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is not marked as formalized in the paper

family_bang_bang · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 117

lemma family_bang_bang:
  fixes z d :: real
  assumes "0 \<le> z" and "z < 1" and "d = arccos z"
  shows "fam_w z d = 0" and "fam_x z d = d - sin d" and "fam_y z d = d + sin d"
    and "fam_x z d + fam_y z d = 2 * d"

family_inside_disc · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 525

theorem family_inside_disc:
  fixes z \<phi> :: real
  assumes "0 \<le> z" and "z < 1" and "0 < \<phi>" and "\<phi> < arccos z"
  shows "(1 - z)\<^sup>2 < fam_x z \<phi> * fam_y z \<phi>"

family_sum_above_boundary · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 306

lemma family_sum_above_boundary:
  fixes z \<phi> :: real
  assumes "0 \<le> z" and "z < 1" and "0 < \<phi>" and "\<phi> < arccos z"
  shows "2 * arccos z < fam_x z \<phi> + fam_y z \<phi>"

family_x_increasing · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 260

lemma family_x_increasing:
  fixes z r s :: real
  assumes "0 \<le> z" and "z < 1" and "0 < r" and "r < s" and "s \<le> arccos z"
  shows "fam_x z r < fam_x z s"

family_y_decreasing · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 280

lemma family_y_decreasing:
  fixes z r s :: real
  assumes "0 \<le> z" and "z < 1" and "0 < r" and "r < s" and "s \<le> arccos z"
  shows "fam_y z s < fam_y z r"

family_product_tends_to_hyperbola · Isabelle/MemoryAccessFloors/U_Turn_Exposure_Family.thy, line 490

theorem family_product_tends_to_hyperbola:
  fixes z :: real
  shows "((\<lambda>t. fam_x z t * fam_y z t) \<longlongrightarrow> (1 - z)\<^sup>2) (at_right 0)"

Cite: Remark 7, p. 12, doi:10.5281/zenodo.24250503

↳ Used in: Proposition 8, 9

🖼️ Figure 3
Propositions 5 and 8. Left: relative excess \((F-F_{\rm sq})/F_{\rm sq}\) of the bang–bang pulse of Theorem 1(b) with parameter \(d\) over the square pulse, with \(F=2(1+\varrho )|\zeta |^2+A^2+\varrho B^2\) (negative: better than the square pulse; orange plane: zero; dot: \(-3.3\, \% \) at \(\varr…

Propositions 5 and 8. Left: relative excess \((F-F_{\rm sq})/F_{\rm sq}\) of the bang–bang pulse of Theorem 1(b) with parameter \(d\) over the square pulse, with \(F=2(1+\varrho )|\zeta |^2+A^2+\varrho B^2\) (negative: better than the square pulse; orange plane: zero; dot: \(-3.3\, \% \) at \(\varrho =0.1\), \(d=0.848\)). Right: leading-order error \(J\) with a lossy memory, relative to the square pulse, against \(\Lambda =\sigma _q^2/({\gamma _m}G)\), for the zero-leakage pulses \(C_{\pi +\phi }SC_\phi \) and for the lower bound of the proof of Proposition 8, valid for every pulse: the square pulse is optimal up to \(\Lambda =7.1\) and not beyond \(7.92\) (grey band; for pulses of fixed phase and no detuning the exact threshold is \(\Lambda _c=7.9079\), Proposition 8(iii)); the explicit family beats it from \(\Lambda \approx 15\).

2.5 Slow noise together with memory loss

When the memory also loses its excitation, waiting in it is no longer free. To leading order in both effects the error of a transfer with slow noise on the qubit and a memory decay rate \({\gamma _m}\) is \(J=\sigma _q^2(2|\zeta |^2+A^2)/6+{\gamma _m}B/3\).

MarAprop:qsanswersMarAcor:channeluses Noise_Loss_Bound.noiseloss_lower_core, Noise_Loss_Bound.noiseloss_minimise_z, Noise_Loss_Bound.noiseloss_combined, Noise_Loss_Bound.noiseloss_large_zeta Let the offset act on the qubit only, centred with variance \(\sigma _q^2\); let the memory decay at rate \({\gamma _m}\), with \({\gamma _q}=0\) and no dephasing; let \(A\), \(B\), \(\zeta \) be those of the lossless pulse, with \(|g|\le G\) and any detuning, and \(J\) the leading-order error above. Put \(\Lambda =\sigma _q^2/({\gamma _m}G)\).

  1. The square pulse minimises \(J\) over all pulses for \(\Lambda \le 7.1\), and for \(\Lambda \ge 7.92\) it does not (both certified by interval arithmetic; a floating-point minimisation over the relaxed region suggests that the certificate is lost near \(7.15\), which is not established as a physical threshold).

  2. As \(\Lambda \to \infty \),

    \[ \min J=\frac{\pi ^2\sigma _q^2}{96G^2}+\frac{\sigma _q\sqrt{\pi {\gamma _m}}}{6\, G^{3/2}}+O\Bigl(\frac{{\gamma _m}}G\Bigr), \]

    and the zero-leakage pulse \(C_{\pi +\phi }SC_\phi \) whose wait lasts about \(\sigma _q\sqrt\pi /(4G^{3/2}\sqrt{{\gamma _m}})\) attains the first two terms.

  3. For pulses of fixed phase and no detuning the square pulse minimises \(J\) exactly for \(\Lambda \le \Lambda _c\), and \(\Lambda _c\) is the crossing of the general turn–wait–turn family \(C_{\pi +\phi }SC_\phi \), with \(\phi \) and \(w\) free in the admissible domain:

    \[ \Lambda _c=\sup \{ \Lambda :\ \min _{\phi ,w}g\ge 0\} =7.9078851606875\ldots \]

    (enclosed in an interval of width \(10^{-12}\); certified by interval arithmetic, using Lemma 6); above it the member with \(\phi _*=0.742069\ldots \) and a short wait \(w_*=0.257473\ldots \) (units \(G=1/2\)) beats it; this critical member has \(|\zeta |/z_0\approx 0.6500698\), so it is not a zero-leakage pulse. For pulses of varying phase or with detuning, optimality remains undecided for \(7.1{\lt}\Lambda \le \Lambda _c\); for every \(\Lambda {\gt}\Lambda _c\) the explicit planar competitor already beats the square pulse. Moreover, for every \(\Lambda {\gt}\Lambda _c\) the minimum of \(J\) over pulses of fixed phase and no detuning is attained, in \((A,B,|\zeta |)\), by a member of the family \(C_{\pi +\phi }SC_\phi \).

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

noiseloss_lower_core · Isabelle/MemoryAccessFloors/Noise_Loss_Bound.thy, line 22

lemma noiseloss_lower_core:
  fixes sigma2 gm A0 z0 z x y :: real
  assumes "0 \<le> sigma2" and "0 < gm" and "0 < A0" and "0 < z0" and "0 \<le> z" and "z \<le> z0"
    and "0 \<le> x" and "0 \<le> y" and "(z0 - z)\<^sup>2 \<le> x * y"
  shows "(2 / 3) * (z0 - z) * sqrt (sigma2 * A0 * gm) \<le> sigma2 * A0 * x / 3 + gm * y / 3"

noiseloss_minimise_z · Isabelle/MemoryAccessFloors/Noise_Loss_Bound.thy, line 75

lemma noiseloss_minimise_z:
  fixes sigma2 A0 gm z0 z s :: real
  assumes "0 < sigma2" and "0 < A0" and "0 < gm" and "0 < z0" and "0 \<le> z" and "z \<le> z0"
    and "s = sqrt sigma2"
  shows "(2 / 3) * z0 * s * sqrt (A0 * gm) - A0 * gm / 3
         \<le> sigma2 * z\<^sup>2 / 3 + (2 / 3) * (z0 - z) * s * sqrt (A0 * gm)"

noiseloss_combined · Isabelle/MemoryAccessFloors/Noise_Loss_Bound.thy, line 103

lemma noiseloss_combined:
  fixes sigma2 gm A0 z0 z x y A B :: real
  assumes "0 < sigma2" and "0 < gm" and "0 < A0" and "0 < z0" and "0 \<le> z" and "z \<le> z0"
    and "0 \<le> x" and "0 \<le> y" and "(z0 - z)\<^sup>2 \<le> x * y" and "A = A0 + x" and "B = A0 + y"
  shows "sigma2 * A0\<^sup>2 / 6 + (2 / 3) * z0 * sqrt sigma2 * sqrt (A0 * gm)
         \<le> sigma2 * (2 * z\<^sup>2 + A\<^sup>2) / 6 + gm * B / 3"

noiseloss_large_zeta · Isabelle/MemoryAccessFloors/Noise_Loss_Bound.thy, line 146

lemma noiseloss_large_zeta:
  fixes sigma2 gm A0 z0 z A B :: real
  assumes "0 \<le> sigma2" and "0 \<le> gm" and "0 < A0" and "0 < z0" and "z0 \<le> z"
    and "A0 \<le> A" and "A0 \<le> B"
  shows "sigma2 * (A0\<^sup>2 + 2 * z0\<^sup>2) / 6 \<le> sigma2 * (2 * z\<^sup>2 + A\<^sup>2) / 6 + gm * B / 3"

symmetric_branch_bound · Isabelle/MemoryAccessFloors/Symmetric_Branch.thy, line 38

theorem symmetric_branch_bound:
  fixes p w :: real
  assumes p0: "0 < p" and p1: "p \<le> pi / 2" and w0: "0 \<le> w"
  shows "pi * p + sin p * (pi - 2 * sin p)
           \<le> 2 * ((cos p - w * sin p / 2)\<^sup>2 - 1)
              + (p + sin p + w * (1 + cos p) / 2) * (pi + (p + sin p + w * (1 + cos p) / 2))"

middle_zone_min · Isabelle/MemoryAccessFloors/Middle_Zone.thy, line 36

theorem middle_zone_min:
  fixes Lam A0 z x y x0 d :: real
  assumes "0 < A0" "2 / A0 < Lam" "0 \<le> x0" "x0 \<le> d" "x0 \<le> x" "x0 \<le> y" "2 * d \<le> x + y"
  shows "Lam * (2 * z^2 + (A0 + x0)^2) / 12 + (A0 + (2 * d - x0)) / 3
           \<le> Lam * (2 * z^2 + (A0 + x)^2) / 12 + (A0 + y) / 3"

Proof: in this note·Formalized: Isabelle/HOL, by the author (public formalizations repository)·Covers: part: the named pieces are machine-checked; the statement as a whole is a written proof·Cite: Proposition 8, p. 13, doi:10.5281/zenodo.24250503

↳ Used in: 9

Proof ▼

Lower bound: with \(x=A-A_0\), \(y=B-A_0\), (a) gives \(xy\ge (z_0-|\zeta |)^2\) when \(|\zeta |\le z_0\), so \(J\ge \sigma _q^2A_0^2/6+{\gamma _m}A_0/3+\sigma _q^2|\zeta |^2/3+\sigma _q^2A_0x/3+{\gamma _m}y/3\); minimising \(\sigma _q^2A_0x/3+{\gamma _m}y/3\) over \(xy\ge (z_0-|\zeta |)^2\) and then over \(|\zeta |\) (optimum \(|\zeta |=\sqrt{A_0{\gamma _m}}/\sigma _q\)) gives \(J\ge \pi ^2\sigma _q^2/(96G^2)+\tfrac 23z_0\sigma _q\sqrt{A_0{\gamma _m}}\), the term \({\gamma _m}A_0/3\) of the memory cancelling the one of the minimisation in \(|\zeta |\); this is (ii) from below; when \(|\zeta |{\gt}z_0\), \(J\ge \sigma _q^2(A_0^2+2z_0^2)/6\), which is larger when \(\Lambda \) is large. Upper bound: the pulse \(C_{\pi +\phi }SC_\phi \) with axial displacement \(|s|\) has \(A-A_0=z_0^2/|s|+O(|s|^{-2})\) and \(B=|s|+O(1)\), and the best \(|s|\) is \(z_0\sigma _q\sqrt{A_0/{\gamma _m}}\). (i): the difference between a pulse and the square pulse is linear in \(\Lambda \) and non-negative at \(\Lambda =0\), so it suffices to show it at \(\Lambda =7.1\), which is done as in Proposition 5(iii): an estimate near \(|\zeta |=z_0\) and 518 accepted leaf boxes elsewhere (1035 nodes evaluated). The bang–bang pulse with \(d=0.85\) has a difference negative at \(\Lambda =7.92\) and decreasing in \(\Lambda \).

(iii): as in Proposition 5(iv), the set of \(\Lambda \ge 0\) where the square pulse is optimal is an interval \([0,\Lambda _c]\) (affinity in \(\Lambda \) and optimality at \(\Lambda =0\)), and the relaxed region is cut by Lemma 6, with the same reading of [ 14 ] . It suffices to certify \(\Lambda _{\rm lo}=7.90788\): the family in its parameters \((\phi ,w)\) (2877 nodes; the minimum is interior, with lower bound \(6\cdot 10^{-9}\), since \(\Lambda _{\rm lo}\) is \(5\cdot 10^{-6}\) below the family crossing), the symmetric branch and the rest of the relaxed region (4663 nodes in all, certificar_umbral_plano_perdidas.py); the same certificate at \(\Lambda =7.90789\), \(7.91\) and \(7.92\) fails, as it should. Above the enclosure the competitor is not the bang–bang pulse but the member of the family with \(\phi =0.742069\) and wait \(w=0.257473\) (units \(G=1/2\)), whose difference with the square pulse is negative at \(\Lambda =7.90789\) (enclosed) and decreasing in \(\Lambda \); integrated as a piecewise pulse it closes at the south pole to \(3\cdot 10^{-16}\) and reproduces the closed form to \(5\cdot 10^{-16}\). The bang–bang pulse crosses only at \(7.9176\). The identification with the family crossing is a second certificate (certificar_lambda_c_identificacion.py): a Krawczyk operator in \((\Lambda ,\phi ,w)\) proves that \(\{ g=0,\ g_\phi =0,\ g_w=0\} \) has a unique zero in a box (the \(\Lambda \) coordinate of width below \(2.7\cdot 10^{-39}\), every coordinate below \(3.3\cdot 10^{-39}\)). On the fixed rational rectangle \(N=[\phi _*\pm 0.01]\times [w_*\pm 0.02]\), centred at rational decimal approximations of \((\phi _*,w_*)\), the Hessian satisfies \(H{\gt}0.3\, I\) uniformly for every \(\Lambda \) in the enclosing interval \(E\), as verified by interval arithmetic, and the Krawczyk-enclosed critical point lies strictly inside \(N\) and the physical family domain; at its parameter value, strict convexity makes this point the minimum over \(N\), with value zero. Outside \(N\), after the analytic near-square patch and the non-physical boxes are separated, the retained exterior family boxes have an interval lower bound at least \(1.88\cdot 10^{-7}\) at both endpoints of \(E\) (the square pulse remains a zero in the analytic patch); the symmetric branch and the rest of the region are certified as before, at both endpoints and extended by affinity; controls fail. \(\Lambda _c\) is close to, but distinct from, \(1/\varrho _c\): approximately \(7.90788516068753\) against \(7.90804384366106\); no relation between the two is asserted.

The minimiser above \(\Lambda _c\) (units \(G=1/2\), so \(z_0=1\) and \(A_0=\pi /2\); with \({\gamma _m}{\gt}0\) we compare the normalised objective \(\hat J=J/{\gamma _m}\), which has the same minimisers, and \(12(\hat J-\hat J_{\rm sq})=\Lambda Q+4y\) with \(Q=2(z^2-1)+x(2A_0+x)\), \(z=|\zeta |\)). At fixed \(z\), \(J\) is strictly increasing in \(x\) and in \(y\). Let \(d=\arccos z\) and \(x_0=d-\sin d\). If \(z\ge 1\), \(J\ge J_{\rm sq}\). If \(z{\lt}1\) and \(x\le x_0\), Lemma 6 gives \(y\ge Y(x;z)\), the \(y\) of the member of the family with the same \(x\) and \(z\), so \(J\) is at least its value there. If \(z{\lt}1\) and \(y\le x_0\), the symmetric statement bounds \(J\) from below by a point of the mirrored branch, with \((x,y)=(y_f,x_f)\) for a member \((x_f,y_f)\) of the family, where \(12(\hat J-\hat J_{\rm sq})=\Lambda Q_m+4x_f\) and \(Q_m=2(z^2-1)+y_f(\pi +y_f)\). With \(s=\sin \phi \) and \(c=\cos \phi \), the closed forms of Remark 7 give \(z\le c\) and \(1-z^2\le (1-z)(1+c)=s^2+(1+c)ws/2\), hence

\[ Q_m\ \ge \ \pi \phi +s(\pi -2s)+w(1+c)\bigl(\tfrac \pi 2-s\bigr)\ \ge \ \pi \phi +s(\pi -2s)\ {\gt}\ 0 , \]

so the mirrored branch never beats the square pulse, for any \(\Lambda \ge 0\). Symmetric_Branch.symmetric_branch_bound Middle_Zone.middle_zone_min If \(x,y{\gt}x_0\), Theorem 1(b) gives \(x+y\ge 2d\), and by monotonicity the minimum lies on \(x+y=2d\), where \(\tfrac {d}{dx}\hat J(x,2d-x)=\tfrac \Lambda 6(A_0+x)-\tfrac 13{\gt}0\) for \(\Lambda {\gt}2/A_0\); so it is the bang–bang point \(x=x_0\), a member of the family. Since \(\Lambda _c{\gt}2/A_0\), every case is bounded by the family or by the square pulse, which the family beats above \(\Lambda _c\); the family minimum exists because \(J\to \infty \) as \(w\to \infty \). This identifies the end point of the minimiser, not the control that reaches it. Numerically, a piecewise search with \(8\) to \(24\) segments and \(252\) starts per value lands on the family minimum to \(10^{-8}\) for \(\Lambda \) from \(\Lambda _c+0.1\) to \(1000\), on the frontier of Remark 7 (optimo_sobre_lambda_c/).

When the memory also decays, robustness must be paid in exposure, which the decay punishes. Measured by \(\Lambda =\sigma _q^2/(\gamma _mG)\), the square pulse stays optimal over all pulses for small \(\Lambda \) and stops being optimal beyond a threshold; for pulses of fixed phase and no detuning the threshold \(\Lambda _c\) is identified exactly, and for large noise the optimum is given to leading order.

🛡️ Evidence & scope
  • Literature: Companion paper: this statement answers a question left by paper A (Exact limits on writing a superconducting quantum memory), Proposition 14 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

  • Literature: Companion paper: this statement uses paper A (Exact limits on writing a superconducting quantum memory), Corollary 9 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

The expansion behind \(J\) needs \(\sigma _qT\ll 1\), and the optimal wait makes \(T\) grow like \(\sigma _q/\sqrt{{\gamma _m}}\): the regime of (ii) is \({\gamma _m}G\ll \sigma _q^2\ll \sqrt{{\gamma _m}}\, G^{3/2}\), a memory that keeps its excitation much longer than the slow noise of the qubit lets it keep its phase. Inside it the exact dynamics (losses and a Gaussian offset averaged by quadrature) confirms the gain of waiting: with \({\gamma _m}/G=2\cdot 10^{-6}\) and \(\Lambda =60\) the error falls from \(2.29\cdot 10^{-5}\) for the square pulse to \(1.76\cdot 10^{-5}\), and with \({\gamma _m}/G=2\cdot 10^{-8}\) and \(\Lambda =2000\) from \(7.45\cdot 10^{-6}\) to \(4.43\cdot 10^{-6}\), within \(1.1\, \% \) of \(J\); with \(\sigma _qT\approx 5\) the square pulse is better again.

Remark 9✅ Isabelle relaxation of the qubit; the bound and the sufficient condition
📕 p. 15📍 map❝ cite⚠ report

MarAprop:qsanswers Noise_Loss_Relax.relax_bound, Noise_Loss_Relax.relax_bound_G, Noise_Loss_Relax.relax_large_zeta_margin, Noise_Loss_Relax.relax_recovers_noiseloss, Square_Condition.square_pulse_optimal, Square_Condition.square_pulse_optimal_G, Dubins_Family_Identity.dubins_identity, Dubins_Family_Identity.dubins_quantity_nonneg, Dubins_Family_Identity.dubins_D_bound, Threshold_Localisation.threshold_lower_bound, Threshold_Localisation.threshold_lower_bound_k, Threshold_Localisation.complete_square_bound, Inf_Affine_Concave.inf_affine_concave, Inf_Affine_Concave.inf_affine_monotone With \({\gamma _q}{\gt}0\) the error is \(J=\sigma _q^2(2|\zeta |^2+A^2)/6+({\gamma _q}A+{\gamma _m}B)/3\), and the same argument, with \(a=\sigma _q^2A_0+{\gamma _q}\) in place of \(\sigma _q^2A_0\), gives

\[ J\ \ge \ \frac{\sigma _q^2A_0^2}{6}+\frac{({\gamma _q}+{\gamma _m})A_0}{3}+\frac23z_0\sqrt{a{\gamma _m}}-\frac{a{\gamma _m}}{3\sigma _q^2} \]

for every pulse (for \(|\zeta |{\gt}z_0\) the difference is at least \((\sigma _qz_0-\sqrt{a{\gamma _m}}/\sigma _q)^2/3\ge 0\); for \({\gamma _q}=0\) this is the bound of the proof above), while the zero-leakage pulses \(C_{\pi +\phi }SC_\phi \) exceed \(\tfrac 23z_0\sqrt{a{\gamma _m}}\) only by higher-order terms. So the price of waiting in the memory is \(\tfrac 23z_0\sqrt{(\sigma _q^2A_0+{\gamma _q}){\gamma _m}}\), to leading order, when \(\sqrt{a{\gamma _m}}\ll \sigma _q^2z_0\). Conversely, the same disk argument shows that the square pulse is optimal at this order whenever \(\sqrt{a{\gamma _m}}\ge \sigma _q^2z_0\): minimising \(\sigma _q^2u^2/3+\tfrac 23\sqrt{a{\gamma _m}}\, (z_0-u)\) over \(0\le u=|\zeta |\le z_0\) then gives \(u=z_0\), the value of the square pulse, and \(|\zeta |{\gt}z_0\) costs more. This is a sufficient condition, not a sharp threshold; for \({\gamma _q}=0\) it reads \(\Lambda \le \pi \), inside the certified range of Proposition 8(i). (Checked against the grid minimisation and the explicit family in four parameter sets: the family is within \(0.15\, \% \) of \(\tfrac 23z_0\sqrt{a{\gamma _m}}\) and the lower bound within the stated correction.)

For pulses of fixed phase and no detuning the sufficient condition is not sharp, and the exact threshold can be certified as in Proposition 8(iii): with \(r={\gamma _q}/{\gamma _m}\) the difference \(J-J_{\mathrm{sq}}\) stays linear in \(\Lambda \), so \(\Lambda _c(r)\) is again a minimum over the turn–wait–turn family, and interval arithmetic encloses it to \(10^{-5}\); a Krawczyk certificate identifies it, for each \(r\) below, with the zero, isolated uniquely in a certified box, of \(\{ g,g_\phi ,g_w\} \) on that family, to \(10^{-12}\) (as for \(r=0\); umbral_relajacion_cubit/):

\(r={\gamma _q}/{\gamma _m}\)

\(0\)

\(0.1\)

\(0.3\)

\(1\)

\(3\)

\(10\)

\(30\)

\(100\)

\(1000\)

\(\Lambda _c(r)\)

\(7.908\)

\(7.956\)

\(8.050\)

\(8.366\)

\(9.170\)

\(11.346\)

\(15.521\)

\(24.260\)

\(67.245\)

sufficient

\(3.141\)

\(3.264\)

\(3.485\)

\(4.113\)

\(5.374\)

\(8.087\)

\(12.637\)

\(21.632\)

\(64.835\)

(the last row is the condition above, \(\Lambda \le (\pi +\sqrt{\pi ^2+16r})/2\), rounded down). Given Lemma 6, \(\Lambda _c(r)\) is the infimum over the members of the family with \(Q{\lt}0\), a set that does not depend on \(r\), of \(4(y+rx)/(-Q)\), affine and non-decreasing in \(r\) since \(x\ge 0\); so \(\Lambda _c\) is non-decreasing and concave in \(r\). The threshold grows with the qubit relaxation like \(2\sqrt r\), with an exact constant: given Lemma 6, as \(r\to \infty \),

\[ \Lambda _c(r)=2\sqrt r+\frac\pi 2+\frac4{\sqrt3}+O(r^{-1/2}) . \]

For the threshold \(\Lambda _c^{\rm all}(r)\) over pulses of any phase and detuning two explicit bounds hold without Lemma 6: as \(r\to \infty \),

\[ 2\sqrt r+\frac\pi 2+O(r^{-1/2})\ \le \ \Lambda _c^{\rm all}(r)\ \le \ 2\sqrt r+\frac\pi 2+\frac4{\sqrt3}+O(r^{-1/2}), \]

the sharp constant of \(\Lambda _c^{\rm all}\) is not identified here. The lower bound is the sufficient condition above, valid for every pulse. The upper bound needs no reading of [ 14 ] : the member of the family with \(\phi =2/\sqrt r\) and \(w=2/\sqrt3\) (units \(G=1/2\)) has \(Q{\lt}0\) for large \(r\) (\(Q=-2w\phi +O(\phi ^2)\)), so it beats the square pulse for every \(\Lambda {\gt}4(y+rx)/(-Q)\), and that explicit ratio, with \(h=r^{-1/2}\), has a simple pole at \(h=0\) and a Laurent expansion \(2\sqrt r+\pi /2+4/\sqrt3+O(r^{-1/2})\) (\(\phi =2/\sqrt r\) minimises the leading term \(\sqrt r(u/2+2/u)\), \(u=\phi \sqrt r\), and \(w=2/\sqrt3\) the next one, \(w+\pi /2+4/(3w)\)). Both bounds therefore hold for pulses of any phase and detuning too. At finite \(r\) the two explicit quantities, not their expansions, are the bounds: the certified values lie between \((\pi +\sqrt{\pi ^2+16r})/2\) and \(4(y+rx)/(-Q)\) for \(10\le r\le 1000\), and approach the upper one (\(67.2454\) against \(67.2788\) at \(r=1000\), cotas_asintoticas.py).

The exact constant needs a lower bound over every member of the family that could compete, including waits that tend to zero or to infinity. Work in units \(G=1/2\), with \(h=r^{-1/2}\), \(a=1-z\), \(D=-Q=2(1-z^2)-\pi x-x^2\) and \(R_r=4(y+rx)/D\) on the members with \(D{\gt}0\), so that, given Lemma 6, \(\Lambda _c(r)=\inf R_r\) (step (c)). (a) Localisation. With the closed forms of Remark 7, \(xy-a^2=E:=\phi ^2-2(1-\cos \phi )+w(\phi -\sin \phi )\ge 0\). Since \(D\le 2(1-z^2)=4a(1-a/2)\), AM–GM gives \(R_r\ge \frac2h\sqrt{1+E/a^2}/(1-a/2)\), and \(D\le 4a\le 4\sqrt{xy}\) gives \(R_r\ge (k+1/k)/h\) with \(k=h\sqrt{y/x}\). The explicit member above has \(R_r=2/h+\frac\pi 2+\frac4{\sqrt3}+O(h)\), so it is enough to bound the members with \(R_r\le 2/h+C\) for a fixed \(C\); for them \(a=O(h)\), \(E/a^2=O(h)\) and \(k=1+O(\sqrt h)\). As \(\phi \to 0\), \(a=\tfrac 12\phi (\phi +w)[1+O(\phi ^2)]\) and \(E=\tfrac 1{12}\phi ^3(\phi +2w)[1+O(\phi ^2)]\) uniformly in \(w\ge 0\) (each is a sum of two non-negative terms), so \(E/a^2\ge \phi /(3(\phi +w))\, [1+O(\phi ^2)]\); hence \(\phi /w=O(h)\), then \(x=\tfrac 14w\phi ^2[1+O(h)]\), \(y=w[1+O(h)]\), \(u:=\phi /h=2+O(\sqrt h)\), and \(m\le w\le M\) for constants independent of \(h\). (b) Expansion. On that compact set, uniformly, \(R_r=h^{-1}\bigl(\tfrac u2+\tfrac 2u\bigr)+c(u,w)+O(h)\) with \(c(u,w)=\frac{2-u^2/6}w+(u^2+4)\bigl(\frac w8+\frac\pi {16}\bigr)\). Since \(\tfrac u2+\tfrac 2u-2=\frac{(u-2)^2}{2u}\), \(c(2,w)=w+\frac4{3w}+\frac\pi 2\ge \frac\pi 2+\frac4{\sqrt3}\) and \(c\) is Lipschitz in \(u\) there, completing the square in \(u-2\) gives \(R_r\ge 2/h+\frac\pi 2+\frac4{\sqrt3}-O(h)\). (c) From the family to the threshold. The case analysis of Proposition 8(iii) carries over with the term \(4rx\) in \(12(\hat J-\hat J_{\rm sq})=\Lambda Q+4(y+rx)\): the cost still increases with \(y\), the mirrored branch has difference \(\Lambda Q_m+4x_f+4ry_f{\gt}0\), and on the middle zone \(\tfrac d{dx}\hat J(x,2d-x)=\tfrac \Lambda 6(A_0+x)+\tfrac {r-1}3{\gt}0\) for \(\Lambda {\gt}4/\pi \), while below \((\pi +\sqrt{\pi ^2+16r})/2\ge \pi {\gt}4/\pi \) the sufficient condition decides. So the sufficient condition misses the threshold by exactly \(4/\sqrt3\) asymptotically. (The identity for \(E\) and the coefficients \(u/2+2/u\) and \(c(u,w)\) checked symbolically; \(\inf R_r\) equal to the certified \(\Lambda _c(r)\) to \(4\cdot 10^{-15}\) at the six certified values \(r=1,3,10,30,100,1000\); \(\sqrt r\, (\Lambda _c-2\sqrt r-\frac\pi 2-\frac4{\sqrt3})\) equal to \(3.77\) at \(r=10^4\) and \(3.76\) at \(r=10^6\); with \(w=1\) fixed the constant is \(3.907\), larger, as it must be: constante_asintotica.py.)

Isabelle · oursshow

relax_bound · Isabelle/MemoryAccessFloors/Noise_Loss_Relax.thy, line 72

theorem relax_bound:
  fixes s g gm A0 z0 z A B :: real
  assumes "0 < s" and "0 \<le> g" and "0 < gm" and "0 < A0" and "0 < z0" and "0 \<le> z"
    and "A0 \<le> A" and "A0 \<le> B" and "(z - z0)\<^sup>2 \<le> (A - A0) * (B - A0)"
  shows "s * A0\<^sup>2 / 6 + (g + gm) * A0 / 3 + (2 / 3) * z0 * sqrt ((s * A0 + g) * gm)
           - (s * A0 + g) * gm / (3 * s)
         \<le> s * (2 * z\<^sup>2 + A\<^sup>2) / 6 + (g * A + gm * B) / 3"

relax_bound_G · Isabelle/MemoryAccessFloors/Noise_Loss_Relax.thy, line 126

corollary relax_bound_G:
  fixes s g gm G z A B :: real
  assumes "0 < G" and "0 < s" and "0 \<le> g" and "0 < gm" and "0 \<le> z"
    and "pi / (4 * G) \<le> A" and "pi / (4 * G) \<le> B"
    and "(z - 1 / (2 * G))\<^sup>2 \<le> (A - pi / (4 * G)) * (B - pi / (4 * G))"
  shows "s * (pi / (4 * G))\<^sup>2 / 6 + (g + gm) * (pi / (4 * G)) / 3
           + (2 / 3) * (1 / (2 * G)) * sqrt ((s * (pi / (4 * G)) + g) * gm)
           - (s * (pi / (4 * G)) + g) * gm / (3 * s)
         \<le> s * (2 * z\<^sup>2 + A\<^sup>2) / 6 + (g * A + gm * B) / 3"

relax_large_zeta_margin · Isabelle/MemoryAccessFloors/Noise_Loss_Relax.thy, line 39

lemma relax_large_zeta_margin:
  fixes s c z0 z :: real
  assumes "0 < s" and "0 \<le> z0" and "z0 \<le> z"
  shows "(sqrt s * z0 - c / sqrt s)\<^sup>2 / 3
         \<le> s * z\<^sup>2 / 3 - ((2 / 3) * z0 * c - c\<^sup>2 / (3 * s))"

relax_recovers_noiseloss · Isabelle/MemoryAccessFloors/Noise_Loss_Relax.thy, line 157

corollary relax_recovers_noiseloss:
  fixes sigma2 gm A0 z0 z A B :: real
  assumes "0 < sigma2" and "0 < gm" and "0 < A0" and "0 < z0" and "0 \<le> z"
    and "A0 \<le> A" and "A0 \<le> B" and "(z0 - z)\<^sup>2 \<le> (A - A0) * (B - A0)"
  shows "sigma2 * A0\<^sup>2 / 6 + (2 / 3) * z0 * sqrt sigma2 * sqrt (A0 * gm)
         \<le> sigma2 * (2 * z\<^sup>2 + A\<^sup>2) / 6 + gm * B / 3"

square_pulse_optimal · Isabelle/MemoryAccessFloors/Square_Condition.thy, line 39

theorem square_pulse_optimal:
  fixes s g gm A0 z0 z A B :: real
  assumes "0 < s" and "0 \<le> g" and "0 < gm" and "0 < A0" and "0 < z0" and "0 \<le> z"
    and "A0 \<le> A" and "A0 \<le> B" and "(z - z0)\<^sup>2 \<le> (A - A0) * (B - A0)"
    and "s * z0 \<le> sqrt ((s * A0 + g) * gm)"
  shows "s * (2 * z0\<^sup>2 + A0\<^sup>2) / 6 + (g + gm) * A0 / 3
         \<le> s * (2 * z\<^sup>2 + A\<^sup>2) / 6 + (g * A + gm * B) / 3"

square_pulse_optimal_G · Isabelle/MemoryAccessFloors/Square_Condition.thy, line 111

corollary square_pulse_optimal_G:
  fixes s gm G z A B :: real
  assumes "0 < G" and "0 < s" and "0 < gm" and "0 \<le> z"
    and "pi / (4 * G) \<le> A" and "pi / (4 * G) \<le> B"
    and "(z - 1 / (2 * G))\<^sup>2 \<le> (A - pi / (4 * G)) * (B - pi / (4 * G))"
    and "s / (gm * G) \<le> pi"
  shows "s * (2 * (1 / (2 * G))\<^sup>2 + (pi / (4 * G))\<^sup>2) / 6 + (0 + gm) * (pi / (4 * G)) / 3
         \<le> s * (2 * z\<^sup>2 + A\<^sup>2) / 6 + (0 * A + gm * B) / 3"

dubins_identity · Isabelle/MemoryAccessFloors/Dubins_Family_Identity.thy, line 29

lemma dubins_identity:
  "dub_x \<phi> w * dub_y \<phi> w - (1 - dub_z \<phi> w)\<^sup>2
     = \<phi>\<^sup>2 - 2 * (1 - cos \<phi>) + w * (\<phi> - sin \<phi>)"

dubins_quantity_nonneg · Isabelle/MemoryAccessFloors/Dubins_Family_Identity.thy, line 55

lemma dubins_quantity_nonneg:
  assumes "0 \<le> \<phi>" "0 \<le> w"
  shows "0 \<le> dub_x \<phi> w * dub_y \<phi> w - (1 - dub_z \<phi> w)\<^sup>2"

dubins_D_bound · Isabelle/MemoryAccessFloors/Dubins_Family_Identity.thy, line 67

lemma dubins_D_bound:
  fixes a D :: real
  assumes "a = 1 - dub_z \<phi> w" "D = 2 * (1 - (dub_z \<phi> w)\<^sup>2) - pi * dub_x \<phi> w - (dub_x \<phi> w)\<^sup>2"
    "0 \<le> dub_x \<phi> w"
  shows "D \<le> 4 * a * (1 - a / 2)"

threshold_lower_bound · Isabelle/MemoryAccessFloors/Threshold_Localisation.thy, line 34

lemma threshold_lower_bound:
  fixes x y a h r D :: real
  assumes "0 < x" "0 < y" "0 < a" "0 < h" "0 < r" "r = 1 / h\<^sup>2" "0 < D"
    "D \<le> 4 * a * (1 - a / 2)" "a < 2" "a\<^sup>2 \<le> x * y"
  shows "(2 / h) * sqrt (x * y) / (a * (1 - a / 2)) \<le> 4 * (y + r * x) / D"

threshold_lower_bound_k · Isabelle/MemoryAccessFloors/Threshold_Localisation.thy, line 65

lemma threshold_lower_bound_k:
  fixes x y h r D :: real
  assumes "0 < x" "0 < y" "0 < h" "r = 1 / h\<^sup>2" "0 < D" "D \<le> 4 * sqrt (x * y)"
  shows "((h * sqrt (y / x)) + 1 / (h * sqrt (y / x))) / h \<le> 4 * (y + r * x) / D"

complete_square_bound · Isabelle/MemoryAccessFloors/Threshold_Localisation.thy, line 91

lemma complete_square_bound:
  fixes h L t :: real
  assumes "0 < h" "0 \<le> L"
  shows "- (3 / 2) * L\<^sup>2 * h \<le> t\<^sup>2 / (6 * h) - L * \<bar>t\<bar>"

inf_affine_concave · Isabelle/MemoryAccessFloors/Inf_Affine_Concave.thy, line 40

lemma inf_affine_concave:
  assumes "I \<noteq> {}" "\<forall>r\<ge>0. bdd_below ((\<lambda>i. \<alpha> i + \<beta> i * r) ` I)"
  shows "concave_on {0..} (inf_affine I \<alpha> \<beta>)"

inf_affine_monotone · Isabelle/MemoryAccessFloors/Inf_Affine_Concave.thy, line 21

lemma inf_affine_monotone:
  assumes "I \<noteq> {}" "\<forall>i\<in>I. 0 \<le> \<beta> i"
    "\<forall>r\<ge>0. bdd_below ((\<lambda>i. \<alpha> i + \<beta> i * r) ` I)"
  shows "monotone_on {0..} (\<le>) (\<le>) (inf_affine I \<alpha> \<beta>)"

Cite: Remark 9, p. 15, doi:10.5281/zenodo.24250503

↳ Used in: 11

🛡️ Evidence & scope
  • Literature: Companion paper: this statement answers a question left by paper A (Exact limits on writing a superconducting quantum memory), Proposition 14 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: Qubit relaxation: the threshold Λc(r) beyond which the square pulse stops being optimal

With memory decay \(\gamma _m\), qubit relaxation \(\gamma _q=r\gamma _m\) and slow noise on the qubit (\(\Lambda =\sigma _q^2/(\gamma _mG)\)), \(g=12(J-J_{\rm sq})=\Lambda Q+4(y+rx)\), \(Q=2(z^2-1)+x(2A_0+x)\), over the family \(C_{\pi +\phi }SC_\phi \) and its mirror (fixed phase, no detuning; units \(G=1/2\)). \(g\) is linear in \(\Lambda \), so \(\Lambda _c(r)=\min 4(y+rx)/(-Q)\), found by Dinkelbach iterations on the minimum over \((\phi ,w)\), and the lab says whether your \(\Lambda \) lies below it, the member that wins above it, the sufficient condition \((\pi +\sqrt{\pi ^2+16r})/2\) of the remark and the large-\(r\) expansion. 2D: \(\Lambda _c(r)\) and the sufficient condition against \(r\), with your point. 3D: the minimum over the wait of the relative change of \(J\), over \((\Lambda /\Lambda _c,\phi )\), one surface per \(r\) (slider): the valley opens exactly at \(\Lambda /\Lambda _c=1\). Numerical; the Krawczyk and interval certificates are in the paper.

Statements (i)–(iii) of Proposition 5 were checked against the exact dynamics averaged by Gauss–Hermite quadrature: the formula agrees to \(4\cdot 10^{-5}\) in relative terms on 300 random complete transfers with complex couplings and real detunings, none of which goes below the floor of (ii). Moving the noise from the qubit to the memory replaces \(A^2\) by \(B^2\) in the coefficient, as the same check confirms (\(r/\sigma ^2=22.70\) against \((2|\zeta |^2+B^2)/6=22.70\) and \((2|\zeta |^2+A^2)/6=0.47\) for one zero-leakage pulse); since \(B\ge A_0\), the symmetric floor \(\pi ^2\sigma _m^2/(96G^2)\) holds as well.

2.6 Correlated offsets

MarAprop:qsextends Band_Quadratic.band_between, Memory_Floors_Pieces.square_corr, Memory_Floors_Pieces.square_corr_far, Memory_Floors_Pieces.square_corr_uncorrelated_fails, Memory_Floors_Pieces.square_corr_equal_variances, Correlation_Monotone.six_delta_difference, Correlation_Monotone.delta_monotone_in_C, Correlation_Monotone.delta_nonneg_transfer The formula of Proposition 5 holds for correlated offsets: with \(C=\mathbb E(\xi _q\xi _m)\) and \(V=\sigma _q^2+\sigma _m^2-2C\), \(6r=2V|\zeta |^2+\sigma _q^2A^2+2CAB+\sigma _m^2B^2\). At second order, the two limiting covariance structures simplify exactly. For a common offset, \(\xi _q=\xi _m\), \(6r=\sigma ^2T^2\), so the square pulse, of least duration, minimises this coefficient. For opposite offsets, \(\xi _m=-\xi _q\) with \(\sigma ^2{\gt}0\), \(6r=\sigma ^2(8|\zeta |^2+(A-B)^2)\) vanishes exactly when \(\zeta =0\) and \(A=B\), i.e. when the curve \(\mathbf r\) closes; among pulses of fixed phase and no detuning the shortest such transfer is the U-turn with \(s=0\), of duration \(L(0)=\tfrac 73\, T_0\) (Proposition 3), the closed-curve condition of [ 19 , 12 ] . In between, for \(C\ge 0\), the square pulse is optimal at second order among all pulses whenever

\[ \frac\pi 2\sqrt{(\sigma _q^2+C)(\sigma _m^2+C)}\ \ge \ \sigma _q^2+\sigma _m^2-2C . \]

Indeed, with \(x=A-A_0\), \(y=B-A_0\ge 0\) and \(u=|\zeta |\), the difference from the square pulse is \(6\Delta =2V(u^2-z_0^2)+2A_0[(\sigma _q^2+C)x+(\sigma _m^2+C)y]+\sigma _q^2x^2+2Cxy+\sigma _m^2y^2\); the quadratic part is non-negative for \(C\ge 0\), and for \(u\le z_0\) the disk of Theorem 1(a) and \(ax+by\ge 2\sqrt{abxy}\) give \(6\Delta \ge 2(z_0-u)\bigl[2A_0\sqrt{(\sigma _q^2+C)(\sigma _m^2+C)}-2Vz_0\bigr]\), with \(A_0/z_0=\pi /2\); for \(u{\gt}z_0\) every term is non-negative. For \(C=0\) and \(\sigma _q^2+\sigma _m^2{\gt}0\) this sufficient condition cannot hold, since \((\pi /2)\sigma _q\sigma _m\le (\pi /4)(\sigma _q^2+\sigma _m^2){\lt}\sigma _q^2+\sigma _m^2\) (if both variances vanish it holds trivially and every pulse has zero noise cost); the content is the correlated case: for equal positive variances the condition is \(\kappa \ge (4-\pi )/(4+\pi )\), approximately \(0.1201983\). For pulses of fixed phase and no detuning the band can be computed exactly. With \(\kappa =C/(\sigma _q\sigma _m)\) the square pulse is optimal on \([\varrho _c(\kappa ),1/\varrho _c(\kappa )]\) and on no larger interval, where \(\varrho _c\) decreases from \(0.12645\) at \(\kappa =0\) through \(0.0902\) (\(\kappa =0.1\)), \(0.0476\) (\(0.3\)), \(0.0272\) (\(0.5\)) to \(0.00948\) at \(\kappa =1\). At \(\kappa =0,\, 0.1,\, 0.1202\) (\(=601/5000\)), \(0.3,\, 0.5,\, 1\) the threshold is identified exactly, under the same reading of [ 14 ] : it is the zero of the stationarity system of the competing family isolated uniquely in a certified box, enclosed to \(10^{-12}\) by a Krawczyk certificate, the bang–bang family for the first three values (\(\varrho _c=0.12645352249\), \(0.09019042063\), \(0.0843166534\)) and the turn–wait–turn family with a positive wait for the last three (\(0.047565938906\), \(0.027186606021\), \(0.009479169051\); every digit printed is fixed by the enclosure); a strictly convex neighbourhood of the critical point, \(\partial g/\partial s{\gt}0\) there, and interval bounds on the rest of the region at both ends of the enclosure make the identification exact (banda_correlada/). Two facts reduce the certificate to three values of \(\varrho \): with \(s=\sqrt\varrho \) the difference from the square pulse is \(Q_x+bs+Q_ys^2\) with \(b\ge 0\) at each point of the region, so optimality at \(s_{\rm hi}\) and at \(s=1\) gives it in between; and \(\varrho \mapsto 1/\varrho \) exchanges the two elements. A rational turn–wait–turn witness beats the square pulse on \((0,\varrho _{\rm lo}]\). So correlation widens the band but never makes the square pulse optimal for every ratio: as \(\varrho \to 0\) the correlation term \(C=\kappa \sigma _q\sigma _m\) vanishes and the bang–bang pulse wins again. The competitor is the bang–bang pulse up to \(\kappa \approx 0.2072\) and a turn–wait–turn pulse with a positive wait beyond it (numerical). The sufficient condition above holds at equal variances exactly for \(\kappa \ge (4-\pi )/(4+\pi )\); at \(\kappa =0\) and \(0.1\) it holds for no ratio, and at \(\kappa =601/5000,\, 0.3,\, 0.5,\, 1\) its band lies inside the exact one, with lower end exceeding \(\varrho _c\) by factors \(11.80\), \(3.79\), \(3.01\) and \(2.53\) (interval arithmetic, with a failing control, banda_suficiente.py).

The band can only widen as the correlation grows, for pulses of any phase and detuning and without Lemma 6: fix positive variances and the relaxation rates; if the square pulse minimises the leading-order error among the pulses of an admissible class at \(\kappa _1\in [0,1]\), it does at every \(\kappa _2\in [\kappa _1,1]\). Indeed, in the difference \(6\Delta \) above, \(\partial (6\Delta )/\partial C=4(z_0^2-u^2)+2[A_0(x+y)+xy]\), non-negative for \(u\le z_0\) since \(x,y\ge 0\) (Theorem 1(a)), while for \(u{\gt}z_0\) every term of \(6\Delta \) is non-negative at every \(C\ge 0\); relaxation adds terms that do not depend on \(C\). So where the square pulse is optimal exactly on \([\varrho _c(\kappa ),1/\varrho _c(\kappa )]\) these bands are nested and \(\varrho _c\) is non-increasing (not shown to be strictly decreasing).

Isabelle · ours · partshow

Only a part is in Isabelle: the named pieces are machine-checked; the statement as a whole is a written proof

band_between · Isabelle/MemoryAccessFloors/Band_Quadratic.thy, line 56

theorem band_between:
  fixes Qx b Qy sa s :: real
  assumes "0 \<le> sa" "sa \<le> s" "s \<le> 1" "0 \<le> b" "0 \<le> Qx + b * sa + Qy * sa^2" "0 \<le> Qx + b + Qy"
  shows "0 \<le> Qx + b * s + Qy * s^2"

square_corr · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 93

theorem square_corr:
  fixes A0 z0 sq sm C V x y u :: real
  assumes "0 < z0" and "A0 = pi / 2 * z0" and "0 \<le> sq" and "0 \<le> sm" and "0 \<le> C"
    and "V = sq + sm - 2 * C" and "0 \<le> V" and "0 \<le> x" and "0 \<le> y" and "0 \<le> u" and "u \<le> z0"
    and "(z0 - u)\<^sup>2 \<le> x * y"
    and "V \<le> pi / 2 * sqrt ((sq + C) * (sm + C))"
  shows "0 \<le> 2 * V * (u\<^sup>2 - z0\<^sup>2) + 2 * A0 * ((sq + C) * x + (sm + C) * y)
              + sq * x\<^sup>2 + 2 * C * x * y + sm * y\<^sup>2"

square_corr_far · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 150

theorem square_corr_far:
  fixes A0 z0 sq sm C V x y u :: real
  assumes "0 < z0" and "0 \<le> A0" and "0 \<le> sq" and "0 \<le> sm" and "0 \<le> C"
    and "V = sq + sm - 2 * C" and "0 \<le> V" and "0 \<le> x" and "0 \<le> y" and "z0 < u"
  shows "0 \<le> 2 * V * (u\<^sup>2 - z0\<^sup>2) + 2 * A0 * ((sq + C) * x + (sm + C) * y)
              + sq * x\<^sup>2 + 2 * C * x * y + sm * y\<^sup>2"

square_corr_uncorrelated_fails · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 166

theorem square_corr_uncorrelated_fails:
  fixes sq sm :: real
  assumes "0 < sq" and "0 < sm"
  shows "pi / 2 * sqrt (sq * sm) < sq + sm"

square_corr_equal_variances · Isabelle/MemoryAccessFloors/Memory_Floors_Pieces.thy, line 202

theorem square_corr_equal_variances:
  fixes s0 kappa :: real
  assumes "0 < s0" and "0 \<le> kappa"
  shows "s0 + s0 - 2 * (kappa * s0) \<le> pi / 2 * sqrt ((s0 + kappa * s0) * (s0 + kappa * s0))
         \<longleftrightarrow> (4 - pi) / (4 + pi) \<le> kappa"

six_delta_difference · Isabelle/MemoryAccessFloors/Correlation_Monotone.thy, line 22

lemma six_delta_difference:
  "six_delta sq sm z0 A0 C2 u x y - six_delta sq sm z0 A0 C1 u x y
   = (C2 - C1) * (4 * (z0\<^sup>2 - u\<^sup>2) + 2 * (A0 * (x + y) + x * y))"

delta_monotone_in_C · Isabelle/MemoryAccessFloors/Correlation_Monotone.thy, line 28

lemma delta_monotone_in_C:
  assumes "0 \<le> u" "u \<le> z0" "0 \<le> x" "0 \<le> y" "0 \<le> A0" "C1 \<le> C2"
  shows "delta_fn sq sm z0 A0 C1 u x y \<le> delta_fn sq sm z0 A0 C2 u x y"

delta_nonneg_transfer · Isabelle/MemoryAccessFloors/Correlation_Monotone.thy, line 44

lemma delta_nonneg_transfer:
  assumes "0 \<le> u" "u \<le> z0" "0 \<le> x" "0 \<le> y" "0 \<le> A0" "C1 \<le> C2"
    "0 \<le> delta_fn sq sm z0 A0 C1 u x y"
  shows "0 \<le> delta_fn sq sm z0 A0 C2 u x y"

Cite: Remark 10, p. 17, doi:10.5281/zenodo.24250503

↳ Used in: 11

🛡️ Evidence & scope
  • Literature: Companion paper: this statement extends paper A (Exact limits on writing a superconducting quantum memory), Proposition 14 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: Correlated offsets: the band of ϱ where the square pulse is optimal, computed for the κ you choose

With \(\sigma _q^2=1\), \(\sigma _m^2=\varrho \) and correlation \(\kappa \) (\(C=\kappa \sigma _q\sigma _m\)) the lab minimises \(g=6(J-J_{\rm sq})=2V(z^2-1)+x(2A_0+x)+2C(A_0(x+y)+xy)+\varrho y(2A_0+y)\) over the turn–wait–turn family \(C_{\pi +\phi }SC_\phi \) and its mirror (pulses of fixed phase and no detuning; \(x=A-A_0\), \(y=B-A_0\); \(g\) is quadratic in the wait, so the minimum over the wait is closed form and \(\phi \) is searched by grid and golden section) and finds \(\varrho _c(\kappa )\) by bisection on the sign of the minimum: the square pulse is optimal exactly on \([\varrho _c,1/\varrho _c]\). It then says whether your \((\kappa ,\varrho )\) is inside, which competitor wins outside, and whether the sufficient condition \(\frac\pi 2\sqrt{(\sigma _q^2+C)(\sigma _m^2+C)}\ge \sigma _q^2+\sigma _m^2-2C\) holds. In 3D, the minimum of \(g\) over \((\kappa ,\log _{10}\varrho )\) with the zero plane: flat on the band, a valley outside it, and the edge \(\varrho _c(\kappa )\) drawn on the plane. This is a numerical computation of the observation in the remark; the interval-arithmetic certificate (at \(\kappa =0,\, 0.1,\, 0.1202,\, 0.3,\, 0.5,\, 1\), under the same reading of Patsko–Fedotov) is in the paper.

Remark 11✅ Isabelle one sufficient condition for both
📕 p. 18📍 map❝ cite⚠ report

MarAprop:qsanswers Square_Unified_Condition.square_unified, Square_Unified_Condition.square_unified_no_relax, Square_Unified_Condition.square_unified_qubit_only, Square_Unified_Condition.square_unified_uncorrelated_needs_relax The sufficient conditions of Remarks 9 and 10 are cases of one. With slow offsets on both elements (second moments \(\sigma _q^2\), \(\sigma _m^2\), covariance \(C\ge 0\), \(V=\sigma _q^2+\sigma _m^2-2C\ge 0\)) and Markovian relaxation of both (\({\gamma _q},{\gamma _m}\ge 0\), no pure dephasing), the leading-order error is \(6J=2V|\zeta |^2+\sigma _q^2A^2+2CAB+\sigma _m^2B^2+2({\gamma _q}A+{\gamma _m}B)\), and the square pulse minimises it over all pulses whenever

\[ V z_0\ \le \ \sqrt{\bigl((\sigma _q^2+C)A_0+{\gamma _q}\bigr)\bigl((\sigma _m^2+C)A_0+{\gamma _m}\bigr)} . \]

The proof is the disk of Theorem 1(a), \((z_0-|\zeta |)^2\le (A-A_0)(B-A_0)\), followed by the arithmetic–geometric mean inequality. Without relaxation it is the condition of Remark 10; with \(\sigma _m=C=0\) it is the condition of Remark 9; with \({\gamma _q}={\gamma _m}=C=0\) and \(\sigma _q^2+\sigma _m^2{\gt}0\) it cannot hold (\(\tfrac \pi 2\sqrt{\sigma _q^2\sigma _m^2}{\lt}\sigma _q^2+\sigma _m^2\); if both variances also vanish it holds trivially and every pulse has zero leading-order cost), in agreement with the bang–bang pulse beating the square pulse outside \([\varrho _c,1/\varrho _c]\). It is sufficient, not sharp: at \({\gamma _q}={\gamma _m}=C=0\) the exact band of Proposition 5 is wider. (Checked on 3620 random parameter sets meeting the condition, over the region the disk allows: minimum of \(6(J-J_{\mathrm{sq}})\) equal to \(3.6\cdot 10^{-4}\ge 0\); the control with \({\gamma _q}={\gamma _m}=C=0\), \(\varrho =0.05\) and the bang–bang pulse of \(d=0.8\) violates the condition and gives \(6(J-J_{\mathrm{sq}})=-0.115\); condicion_unificada.py.)

Isabelle · oursshow

square_unified · Isabelle/MemoryAccessFloors/Square_Unified_Condition.thy, line 52

theorem square_unified:
  fixes z0 A0 sq sm C gq gm x y u :: real
  assumes "0 < z0" "A0 = pi/2 * z0" "0 \<le> sq" "0 \<le> sm" "0 \<le> C" "0 \<le> gq" "0 \<le> gm"
      "0 \<le> sq + sm - 2*C" "0 \<le> x" "0 \<le> y" "0 \<le> u"
      "u \<le> z0 \<Longrightarrow> (z0 - u)^2 \<le> x*y"
      "(sq + sm - 2*C) * z0 \<le> sqrt (((sq + C)*A0 + gq) * ((sm + C)*A0 + gm))"
  shows "0 \<le> 2*(sq + sm - 2*C)*(u^2 - z0^2)
            + 2*(((sq + C)*A0 + gq)*x + ((sm + C)*A0 + gm)*y)
            + sq*x^2 + 2*C*x*y + sm*y^2"

square_unified_no_relax · Isabelle/MemoryAccessFloors/Square_Unified_Condition.thy, line 123

corollary square_unified_no_relax:
  fixes z0 A0 sq sm C x y u :: real
  assumes "0 < z0" "A0 = pi/2 * z0" "0 \<le> sq" "0 \<le> sm" "0 \<le> C"
      "0 \<le> sq + sm - 2*C" "0 \<le> x" "0 \<le> y" "0 \<le> u"
      "u \<le> z0 \<Longrightarrow> (z0 - u)^2 \<le> x*y"
      "(sq + sm - 2*C) * z0 \<le> A0 * sqrt ((sq + C) * (sm + C))"
  shows "0 \<le> 2*(sq + sm - 2*C)*(u^2 - z0^2)
            + 2*((sq + C)*A0*x + (sm + C)*A0*y) + sq*x^2 + 2*C*x*y + sm*y^2"

square_unified_qubit_only · Isabelle/MemoryAccessFloors/Square_Unified_Condition.thy, line 149

corollary square_unified_qubit_only:
  fixes z0 A0 sq gq gm x y u :: real
  assumes "0 < z0" "A0 = pi/2 * z0" "0 \<le> sq" "0 \<le> gq" "0 \<le> gm"
      "0 \<le> x" "0 \<le> y" "0 \<le> u"
      "u \<le> z0 \<Longrightarrow> (z0 - u)^2 \<le> x*y"
      "sq * z0 \<le> sqrt ((sq*A0 + gq) * gm)"
  shows "0 \<le> 2 * sq * (u^2 - z0^2) + 2*((sq*A0 + gq)*x + gm*y) + sq*x^2"

square_unified_uncorrelated_needs_relax · Isabelle/MemoryAccessFloors/Square_Unified_Condition.thy, line 163

corollary square_unified_uncorrelated_needs_relax:
  fixes z0 A0 sq sm :: real
  assumes "0 < z0" "A0 = pi/2 * z0" "0 \<le> sq" "0 \<le> sm" "0 < sq + sm"
  shows "\<not> ((sq + sm - 2*0) * z0 \<le> sqrt (((sq + 0)*A0 + 0) * ((sm + 0)*A0 + 0)))"

Cite: Remark 11, p. 18, doi:10.5281/zenodo.24250503

↳ Not used by another result of the paper.

🛡️ Evidence & scope
  • Literature: Companion paper: this statement answers a question left by paper A (Exact limits on writing a superconducting quantum memory), Proposition 14 (silent link, not printed in the paper). https://karlesmarin.github.io/blueprints/memory-floors/

🧪 Lab: One sufficient condition for both: correlated offsets and relaxation, with its margin and a numerical check

Give the offset variances \(\sigma _q^2,\sigma _m^2\), the correlation \(\kappa \) (\(C=\kappa \sigma _q\sigma _m\)), the relaxation rates \(\gamma _q,\gamma _m\) and the coupling bound \(G\) (default \(G=1/2\), so \(z_0=1\), \(A_0=\pi /2\)). The lab evaluates, in closed form, \(V=\sigma _q^2+\sigma _m^2-2C\), \(a=(\sigma _q^2+C)A_0+\gamma _q\), \(b=(\sigma _m^2+C)A_0+\gamma _m\) and the condition \(Vz_0\le \sqrt{ab}\) with its margin, says which special case it reduces to (no relaxation: the condition of the remark on correlated offsets; \(\sigma _m=C=0\): that of the remark on relaxation; \(\gamma _q=\gamma _m=C=0\) with a variance: it cannot hold; all zero: trivially), and the smallest \(\kappa \) that makes it hold. Then a numerical check, as in the paper: 4000 complete transfers sampled in the region that the disk of the disk of the main theorem, part (a), \((|\zeta |-z_0)^2\le (A-A_0)(B-A_0)\) with \(A,B\ge A_0\), allows (half of them on its edge), and the minimum of \(6(J-J_{\rm sq})\) with \(6J=2V|\zeta |^2+\sigma _q^2A^2+2CAB+\sigma _m^2B^2+2(\gamma _qA+\gamma _mB)\): it is \(\ge 0\) whenever the condition holds; when it fails the condition is only sufficient, and the minimum may still be \(\ge 0\). It also evaluates the exact \(6(J-J_{\rm sq})\) of the bang–bang pulse of the parameter \(d\) you choose (the paper’s control: \(\gamma =C=0\), \(\varrho =0.05\), \(d=0.8\), \(G=1\) gives \(-0.115\)). 2D: the normalised margin \((\sqrt{ab}-Vz_0)/(\sqrt{ab}+Vz_0)\) against \(\kappa \) with your point; 3D: the same margin over \((\kappa ,\log _{10}\varrho )\) with the zero plane. What is proved: that the square pulse is optimal over all pulses whenever the condition holds is a written proof (disk plus AM–GM) and is formalised in Isabelle (Square_Unified_Condition.square_unified); the sampling here is numerical evidence only.

3 Limits of this paper

Questions of the companion paper addressed here.

[ 10 ] gives the error of the square pulse under a quasi-static offset, \((8+\pi ^2)\sigma ^2/(96G^2)\), and leaves open which pulse is best under slow noise, alone or together with loss. The companion questions are addressed through exact second-order floors, certified square-pulse optimality ranges, and explicit improving protocols (Propositions 5 and 8): the floor for every pulse is \(\pi ^2\sigma ^2/(96G^2)\), and the square pulse is optimal in certified ranges of \(\sigma _m^2/\sigma _q^2\) and of \(\sigma _q^2/({\gamma _m}G)\). For fixed-phase, zero-detuning pulses, the two square-pulse thresholds are identified (\(\varrho _c\) and \(\Lambda _c\)). With memory loss, the general optimal pulse at every finite noise-to-loss ratio is not determined here; the waiting family attains the stated asymptotic optimum in the large-ratio regime, and for fixed phase and no detuning it contains the end point of the optimum for every \(\Lambda {\gt}\Lambda _c\) (given Lemma 6; the control that reaches it is not shown to be unique). The exact thresholds with qubit relaxation, \(\Lambda _c({\gamma _q}/{\gamma _m})\), and with correlated offsets, \(\varrho _c(\kappa )\), are certified at the listed parameter values only; for fixed phase and no detuning, \(\Lambda _c(r)=2\sqrt r+\pi /2+4/\sqrt3+O(r^{-1/2})\) is proved given Lemma 6, and \(\Lambda _c\) is non-decreasing and concave in \(r\); for arbitrary phase and detuning the two explicit bounds with constants \(\pi /2\) and \(\pi /2+4/\sqrt3\) hold, but the sharp constant is not identified; \(\varrho _c\) is proved non-increasing in \(\kappa \); the change of competitor at \(\kappa \approx 0.2072\) is numerical.

The results are for two states, one excitation and a static offset: exactly at first order in it (Theorem 1) and at second order (Propositions 5 and 8), with relaxation of either or both elements included at the leading order used in Propositions 3 and 8 and Remarks 9 and 11, and in the regime \(\sigma T\ll 1\). Offsets that change during a transfer, and Markovian relaxation and dephasing beyond that order, are outside; the latter are treated in [ 10 ] . The exact cost of Proposition 3 is proved for pulses of fixed phase and no detuning only, and whether a pulse of varying phase, or a detuning, can do better is open; curves made of pieces in planes through the axis are excluded. For pulses of varying phase or with detuning, undecided ranges remain in Proposition 5(iii) (\(0.1264\)–\(0.137\) and \(7.30\)–\(7.91\); for fixed phase and no detuning (iv) identifies the threshold with the bang–bang crossing, given our reading of [ 14 ] in Lemma 6) and in Proposition 8(i) (\(7.1{\lt}\Lambda \le \Lambda _c\); for every \(\Lambda {\gt}\Lambda _c\) the explicit planar competitor already beats the square pulse, and for fixed phase and no detuning (iii) identifies the threshold with the crossing of a short-wait turn–wait–turn pulse, under the same reading). Those bounds come from a relaxation of the reachable set; the conjectured spatial exposure boundary of Remark 7 would extend the planar constraints to all pulses and could sharpen them, but determining the resulting optima would still require the remaining minimisation. The reduction from the two-level dynamics to the space curve is a written proof, and the paper has not been refereed by a journal.

4 What backs each statement

A formal label refers to the Isabelle statement named in each entry, not to the physical claim as a whole. Two bridges are written proofs throughout. (i) Reduction: the passage from the two-level dynamics to the space curve \(\mathbf n\) and to the exposures \(A\), \(B\), \(\zeta \), and the Dyson expansion that gives the first-order term \(\Xi _1\). (ii) Inputs: where a lemma takes a formula as a hypothesis (an integrated exposure, the candidate lengths of Dubins’ classification), its derivation is written. The Isabelle theories are public, in https://github.com/karlesmarin/formalizations, folder Isabelle/MemoryAccessFloors; each certificate is listed with its script, and the commands and archived outputs are in the supplement.

Theorem 1 and Corollary 2 (robustness)

Backing: proof in the text; numerical minima of the duration at prescribed leakage fall on the frontier to five digits; the bang–bang family checked by integrating the two-level dynamics, the turn–wait–turn family on the Bloch curve (2000 cases, slack below \(5\cdot 10^{-13}\)); for a real coupling at \(\zeta =0\) the exact shortest U-turns (Dubins paths, checked by integrating each path and by an independent solver) lie above the bound. Status: the elementary steps formalised in Isabelle/HOL (concavity_step, tent_integral, frontier_step, disk_from_two_bounds, both equality families, robust_cost_amgm); (a) and (b) formalised for continuous unit curves \(n\) that are differentiable at every interior time, with the speed bound \(2G\) imposed pointwise, through the poles included (Robust_Disk.robust_disk, Robust_Core.frontier_general); the written proof establishes the same statements for Lipschitz curves, including those of switching controls; the equality checks are formalised conditional on the explicit exposure formulas, whose derivation from the named controls is written; the reduction from the two-level dynamics to the curve (\(A\), \(B\), \(\zeta \) as integrals of \(n\)) is the written proof.

Proposition 3 (exact cost for fixed-phase, zero-detuning pulses)

Backing: proof in the text from Dubins’ theorem, with the comparison of the eight candidates done analytically; the closed forms of the candidates agree to \(7\cdot 10^{-15}\) with their geometry for \(0\le s\le 4\), each path checked by integration and by an independent solver, and the optimal paths, run as pulses through the two-level dynamics, transfer completely with \(\zeta =0\); the minimum over \(s\) agrees with the formula to \(10^{-7}\) and falls at the predicted \(|s|\); curves out of the plane found nothing lower. Status: the minimisation over \(s\) formalised in Isabelle/HOL, with \(L(s)\) as a definition (csc_minimum, ccc_branch, closed_form_value, total_minimum); the comparison of the eight candidates formalised too (Dubins_U_Turn), with their closed forms as definitions; the derivation of the candidate lengths from the paths is written and checked numerically for \(s\ne 2\), with the degenerate end point \(s=2\) treated separately (at \(s=2\) the candidate Lccc_b keeps an extra-turn representative of length \(4\pi \); the shortest degenerate representative has length \(2\pi \), and the minimum at \(s=2\) is \(2\pi \)); that the shortest is \(CSC\) or \(CCC\) is Dubins’ theorem, cited; unequal rates with varying phase or detuning open (pieces in planes through the axis excluded, Meridian_Unfolding).

Proposition 5 (slow noise at second order)

Backing: proof in the text; exact dynamics averaged by Gauss–Hermite on 300 random transfers (relative error \({\lt}4\cdot 10^{-5}\)), control with the noise in the memory; thresholds by interval certificates (certificar_cuadrado_optimo, certificar_robustez). Status: the unitarity step of the expansion (Unitarity_Second_Order: a family \(I-i\epsilon X_1-\epsilon ^2X_2+O(\epsilon ^3)\) unitary for small \(\epsilon \) has \(X_1\) Hermitian and \(X_2+X_2^\dagger =X_1^2\), hence the formula for \(2\operatorname {Re}(\Xi _2)_{aa}\); the formal result proves the coefficient identities and the truncated-modulus algebra, while the dynamical construction, the identification of \(X_1\), the stochastic averaging and the averaged remainder are written proofs), the algebra, (i), (ii) given Theorem 1(a), and the bang–bang example at \(\varrho =0.1\) formalised in Isabelle/HOL, with the kernel only (Slow_Noise_Second_Order, Slow_Noise_Kernel); in (iii) the square pulse beaten for \(\varrho \le 0.1264\) and \(\varrho \ge 7.91\) certified by interval arithmetic (one pulse, the difference being linear in \(\varrho \)), its optimality on \([0.137,7.30]\) certified by branch and bound with intervals, with a control at \(\varrho =0.12\) that fails (the certificate at \(\varrho =0.137\) is implemented with both mpmath intervals and Arb; the extension to the decimal endpoint \(7.30\) uses the Python certificate at the exact rational \(\varrho =10/73\)); (iv) the enclosure \(0.126453{\lt}\varrho _c\le 0.12645395\) certified by intervals given Lemma 6 (certificar_umbral_plano: optimality over the cut region at \(0.12645395\) and \(1\), 55 567 nodes and 22 322 accepted leaves; strict witness \(d=0.7717755\) at \(0.126453\)); the identification \(\varrho _c=\varrho _{\rm bb}\) by a second certificate (certificar_rho_c_igual_bb: Krawczyk in \((\varrho ,d)\), enclosure widths below \(6\cdot 10^{-41}\) for \(\varrho _{\rm bb}\) and \(4\cdot 10^{-40}\) for \(d_*\); Hessian positive definite on \(N\) uniformly for \(\varrho \) in the enclosure; margin at least \(5\cdot 10^{-7}\) on the retained branch boxes outside \(N\), the near-square patch and the tail being treated separately; five controls that fail; 1.3 s). The displayed root values are rounded approximations; the existence and uniqueness tests use the full internal intervals. The gradient and Hessian formulas were checked against finite differences (agreement \(2\cdot 10^{-8}\); a sign error gives \(10^{-2}\)), which detects errors but is not part of the proof.

Lemma 6 (planar exposure bound)

Backing: written derivation from the sections of [ 14 , Lemma 3.1 and Theorem 11.1 ] (branch selection and the passage to exposures, both lateral signs covered directly); the inequality checked with both lateral signs on 7061 random turn–wait–turn pulses in the sector (verificar_v291: the only point below the bound is a member of the equality family itself at \(a=5.6\cdot 10^{-8}\), short by \(2.8\cdot 10^{-9}\), the rounding of inverting \(a_z\) near \(a=0\)); the closed forms \(a_z,b_z\) against the family of the remark to \(2\cdot 10^{-13}\) (verificar_v289). Status: the elementary parts formalised in Isabelle/HOL (Memory_Floors_Pieces: the derivatives and signs of \(a_z\), \(b_z\), their values at \(d\), the monotonicity and the range of \(a_z\), and the integral of step (1): a_z_deriv, b_z_deriv, a_z_mono, b_z_antimono, a_z_range, odd_turns_integral); the properties of the family are formalised too (Remark 7); the cut used in step (2), that the arc \(U_3=(1,0,-1)\) is the lower envelope of the section in the sector, is certified within the extremal families \(U_1,U_2,U_3,U_6\) by interval arithmetic for \(z\le 1-2^{-12}\) (envolvente_U3: 67 687 nodes, minimum margin \(4\cdot 10^{-4}\), six controls that fail), with closed-form identities checked to 60 digits rather than proved symbolically; the identification of these closed forms with the trajectories and the end region \(z{\gt}1-2^{-12}\) rest on the written argument, so this is not a complete certificate of the lemma; that these families form the boundary of the reachable set is taken from the classification of Dubins extremals [ 14 , Lemma 3.1 and Theorem 11.1 ] , and steps (2) and (3) are otherwise written.

Remark 7 (the exact planar section)

Backing: the identification with the type-\((1,0,-1)\) arc of [ 14 , Lemma 3.1 and Theorem 11.1 ] is a written argument; the closed forms checked against piecewise pulses with exact integrals (hueco_17iii_b); the conjecture by an adversarial search over three-dimensional pulses (conjetura_adversario). Status: the properties of the family (strict bounds of Theorem 1, monotonicity, asymptote, bang–bang end point) formalised in Isabelle/HOL from the closed forms (U_Turn_Exposure_Family); the conjecture open.

Remark 9 (relaxation of the qubit)

Backing: the bound and the sufficient condition are written proofs from Theorem 1(a), checked against grid minimisation and the explicit family in four parameter sets; the exact threshold \(\Lambda _c(r)\) enclosed by interval arithmetic at nine ratios \(r={\gamma _q}/{\gamma _m}\) from \(0\) to \(1000\) (cert_relajacion; the control \(r=0\) reproduces the enclosure of Proposition 8(iii)) and identified with the crossing of the turn–wait–turn family by Krawczyk certificates (certificar_identificacion_r: unique zero to \(10^{-12}\), controls that fail; at \(r=1000\) with a narrower neighbourhood \(N\)), given Lemma 6; the large-\(r\) bounds are written proofs (the sufficient condition below; an explicit member of the family above, whose ratio is expanded in \(r^{-1/2}\)), compared with the certified values at \(r=10,30,100,1000\) and evaluated at \(r=10^4,10^6,10^8\) as high-precision numerical checks, with no threshold certified there (cotas_asintoticas.py; with \(w=1\) instead of \(2/\sqrt3\) the constant is larger, as it must be). Status: the bound and the sufficient condition formalised in Isabelle/HOL (Noise_Loss_Relax, Square_Condition); the thresholds certified; the sharp constant and the \(O(r^{-1/2})\) remainder established by the written localisation and uniform-expansion argument, given Lemma 6, with its elementary steps formalised (Dubins_Family_Identity.dubins_identity, dubins_quantity_nonneg, dubins_D_bound; Threshold_Localisation.threshold_lower_bound, threshold_lower_bound_k, complete_square_bound) and the uniform expansion not formalised; non-decrease and concavity in \(r\) from the infimum over a fixed domain, the lemma formalised (Inf_Affine_Concave.inf_affine_concave, inf_affine_monotone); the symbolic checks confirm the identities and coefficients and do not replace the uniform argument.

Remark 10 (correlated offsets)

Backing: the two exact limits and the sufficient condition are short written proofs (AM–GM on the disk of Theorem 1(a)); the exact band certified by interval arithmetic at six values of \(\kappa \) given Lemma 6 (cert_banda: optimality at \(s_{\rm hi}\) and \(s=1\), a rational witness on \((0,\varrho _{\rm lo}]\), enclosures of relative width \(1.2\cdot 10^{-3}\); the control \(\kappa =0\) reproduces \(\varrho _c\); twelve controls that fail); the identification with the family crossing at the same six values by Krawczyk certificates (certificar_rho_c_bb_kappa for the bang–bang competitor at \(\kappa =0,\, 0.1,\, 601/5000\), with the middle zone bounded analytically; certificar_identificacion_kappa for the turn–wait–turn competitor at \(\kappa =0.3,\, 0.5,\, 1\); enclosures of \(\varrho _c\) of width \(10^{-12}\); the control \(\kappa =0\) reproduces the original certificate to 30 digits; Krawczyk off the zero, a tiny neighbourhood, a shifted end point and a dropped correlation term all fail); the change of competitor (\(\kappa _w\approx 0.2072\), where the Hessian determinant of the family vanishes) is numerical; that \(\varrho _c\) is non-increasing in \(\kappa \) is a short written proof valid for any phase and detuning (the derivative of \(6\Delta \) in \(C\)), its algebra formalised (Correlation_Monotone.six_delta_difference, delta_monotone_in_C, delta_nonneg_transfer); the comparison of the sufficient band with \(\varrho _c\) by banda_suficiente.py. Status: the interval lemma behind the band formalised (Band_Quadratic.band_between); the sufficient condition formalised in Isabelle/HOL with the disk of Theorem 1(a) as a hypothesis (Memory_Floors_Pieces: square_corr, square_corr_far, square_corr_uncorrelated_fails, square_corr_equal_variances); the two limits are written.

Remark 11 (one sufficient condition)

Backing: written proof (disk of Theorem 1(a) and AM–GM); 3620 random parameter sets meeting the condition, minimum of \(6(J-J_{\rm sq})\) equal to \(3.6\cdot 10^{-4}\), and a control that violates it and gives \(-0.115\) (condicion_unificada). Status: formalised in Isabelle/HOL with the disk as a hypothesis, kernel only (Square_Unified_Condition: square_unified and its three special cases).

Proposition 8 (slow noise and memory loss)

Backing: lower bound in closed form from Theorem 1(a); upper bound in the large-\(\Lambda \) regime by the explicit zero-leakage family; their difference relative to the excess \(J_{\rm lb}-\pi ^2\sigma _q^2/(96G^2)\) falls from \(7.9\, \% \) at \(\Lambda =200\) to \(0.24\, \% \) at \(\Lambda =2\cdot 10^5\) (brecha_cotas_noiseloss.py); the thresholds by interval certificates (certificar_ruido_perdidas: \(\Lambda \le 7.1\) and \(\Lambda \ge 7.92\), with a control at \(\Lambda =8.5\) that fails; for fixed phase and no detuning certificar_umbral_plano_perdidas: \(7.90788\le \Lambda _c{\lt}7.90789\), 4663 nodes, strict witness at \(7.90789\), controls at \(7.90789\), \(7.91\), \(7.92\) that fail, given Lemma 6; the identification with the crossing \(7.9078851606875\ldots \) of the turn–wait–turn family by certificar_lambda_c_identificacion: Krawczyk in \((\Lambda ,\phi ,w)\), \(\Lambda \) coordinate of width below \(2.7\cdot 10^{-39}\) and every coordinate below \(3.3\cdot 10^{-39}\); Hessian \({\gt}0.3\, I\) on \(N\) uniformly for every \(\Lambda \) in the enclosure (57 nodes, 29 leaves) and the isolated critical point strictly inside \(N\); margin at least \(1.88\cdot 10^{-7}\) on the retained exterior boxes at both endpoints, the near-square patch and the non-physical boxes being separated; five controls that fail); exact dynamics with losses and averaged offsets confirms the gain inside the regime \(\sigma _qT\ll 1\) and its loss outside. The optimum above \(\Lambda _c\): written case analysis given Lemma 6 and Theorem 1(b); each step checked on \(1.7\cdot 10^6\) points with a control that fails (prueba_Qm); a piecewise search (\(8\)–\(24\) segments, \(252\) starts per value) agrees with the family minimum to \(10^{-8}\) for \(\Lambda \) up to \(1000\) (optimo_sobre_lambda_c). Status: the lower bound formalised in Isabelle/HOL (Noise_Loss_Bound), given Theorem 1(a), and its extension to \({\gamma _q}{\gt}0\) of the remark that follows (Noise_Loss_Relax); the bound \(Q_m\ge \pi \phi +s(\pi -2s)\) on the mirrored branch formalised (Symmetric_Branch.symmetric_branch_bound) and the middle-zone step (Middle_Zone.middle_zone_min); the threshold enclosures and the crossing identification in (i) and (iii) certified by interval arithmetic, (iii) given Lemma 6, while the optimality of the turn–wait–turn family for every \(\Lambda {\gt}\Lambda _c\) is a written proof with the mirrored-branch and middle-zone inequalities formalised as above; for varying phase or detuning, \(7.1{\lt}\Lambda \le \Lambda _c\) open.

A Formula sheet

quantity

formula

robustness (complete lossless transfer, \(|g|\le G\), any detuning)

\(\bigl(|\zeta |-z_0\bigr)^2\le (A-A_0)(B-A_0)\), \(A,B\ge A_0\); \(T\ge T_0+\dfrac 1G\arccos \dfrac {|\zeta |}{z_0}\) for \(|\zeta |\le z_0\); \(T_0=\dfrac {\pi }{2G}\), \(A_0=\dfrac {\pi }{4G}\), \(z_0=\dfrac 1{2G}\)

cost of \(\zeta =0\), fixed phase, no detuning, \({\gamma _q},{\gamma _m}{\gt}0\)

\(\min \bigl[{\gamma _q}(A-A_0)+{\gamma _m}(B-A_0)\bigr]=\dfrac {{\gamma _q}+{\gamma _m}}{2G}\arccos \dfrac {|{\gamma _q}-{\gamma _m}|}{{\gamma _q}+{\gamma _m}}\)

slow noise, second order (static offsets, no losses)

\(r=\dfrac {2\, \mathbb E(\xi _q-\xi _m)^2|\zeta |^2+\mathbb E(\xi _qA+\xi _mB)^2}{6}\); qubit only: \(r\ge \dfrac {\pi ^2\sigma _q^2}{96G^2}\)

noise on both elements, fixed phase, no detuning

square pulse optimal exactly for \(\varrho _c\le \sigma _m^2/\sigma _q^2\le 1/\varrho _c\), \(\varrho _c=0.12645352248541\ldots \) (the bang–bang crossing)

correlated offsets, \(C=\mathbb E(\xi _q\xi _m)\ge 0\)

\(6r=2V|\zeta |^2+\sigma _q^2A^2+2CAB+\sigma _m^2B^2\), \(V=\sigma _q^2+\sigma _m^2-2C\); square pulse optimal if \(\dfrac \pi 2\sqrt{(\sigma _q^2+C)(\sigma _m^2+C)}\ge V\); fixed phase, no detuning: exactly on \([\varrho _c(\kappa ),1/\varrho _c(\kappa )]\), \(\varrho _c(0.5)=0.0272\), \(\varrho _c(1)=0.00948\)

offsets and relaxation of both elements

square pulse optimal if \(Vz_0\le \sqrt{\bigl((\sigma _q^2+C)A_0+{\gamma _q}\bigr)\bigl((\sigma _m^2+C)A_0+{\gamma _m}\bigr)}\)

slow noise on the qubit and memory decay

\(\min J=\dfrac {\pi ^2\sigma _q^2}{96G^2}+\dfrac {\sigma _q\sqrt{\pi {\gamma _m}}}{6G^{3/2}}+O({\gamma _m}/G)\) for \({\gamma _m}G\ll \sigma _q^2\ll \sqrt{{\gamma _m}}G^{3/2}\); square pulse optimal for \(\sigma _q^2\le 7.1\, {\gamma _m}G\), not for \(\sigma _q^2\ge 7.92\, {\gamma _m}G\); fixed phase, no detuning: optimal exactly for \(\sigma _q^2\le \Lambda _c{\gamma _m}G\), \(\Lambda _c=7.9078851606875\ldots \); with qubit relaxation \(r={\gamma _q}/{\gamma _m}\): \(\Lambda _c(1)=8.366\), \(\Lambda _c(1000)=67.245\), \(\Lambda _c(r)=2\sqrt r+\pi /2+4/\sqrt3+O(r^{-1/2})\) given Lemma 6, non-decreasing and concave in \(r\); any phase and detuning: \(2\sqrt r+\pi /2\lesssim \Lambda _c^{\rm all}(r)\lesssim 2\sqrt r+\pi /2+4/\sqrt3\) up to \(O(r^{-1/2})\)

B The companion site

This paper has a companion website, its blueprint, at https://karlesmarin.github.io/blueprints/robust-offset/. It shows what the paper proves, how its results depend on each other and what exactly has been checked, and lets the reader go from any statement to its proof, its dependencies and its checks:

  • The statements. Every numbered statement of the paper, copied verbatim with its number (1 theorem, 3 propositions, 1 lemma, 1 corollary and 5 remarks), each with its proof.

  • The dependencies. What each statement uses and where it is used, as a list and as a graph, including the links to the companion paper.

  • The map. The main results, the tools they rest on and the fields they feed.

  • What was checked, and how. For each statement, the backing and status of Section 4: 10 statements have named pieces machine-checked in Isabelle/HOL (94 theorems, linked by name to the public repository), and each numerical claim names the interval certificate of the supplementary material that backs it.

  • Labs. Next to Theorem 1, Propositions 3 and 5 and Remarks 9, 10 and 11, a form runs the paper’s own computations in the reader’s browser (Python, nothing to install) for values the reader chooses.

  • Figures in motion. Figures 1–3 are drawn live by the same computations, with the parameters of the paper; each figure of this PDF is a link to its lab.

  • The paper itself. Its PDF, opened at the right page from every statement.

The site is a reading aid, not a part of the proofs: the statements and proofs are those of this paper, whose text prevails, and a formal or computational check is reported for exactly the part it covers.

Acknowledgement

We worked with an AI assistant.

📊 Tables

Every table of the paper, in its order: click one to unfold it, with where the paper cites it, how its numbers were checked, and its data as CSV. The typeset table is one click away, on its page of the PDF.

Table without a number, in the text of §2.5 Slow noise together with memory loss
$r={\gamma_q}/{\gamma_m}$$0$$0.1$$0.3$$1$$3$$10$$30$$100$$1000$
$\Lambda_c(r)$$7.908$$7.956$$8.050$$8.366$$9.170$$11.346$$15.521$$24.260$$67.245$
sufficient$3.141$$3.264$$3.485$$4.113$$5.374$$8.087$$12.637$$21.632$$64.835$
Table without a number, in the text of §A Formula sheet
quantityformula
robustness (complete lossless transfer, $|g|\le G$, any detuning)$\bigl(|\zeta|-z_0\bigr)^2\le(A-A_0)(B-A_0)$, $A,B\ge A_0$; $T\ge T_0+\dfrac1G\arccos\dfrac{|\zeta|}{z_0}$ for $|\zeta|\le z_0$; $T_0=\dfrac{\pi}{2G}$, $A_0=\dfrac{\pi}{4G}$, $z_0=\dfrac1{2G}$
cost of $\zeta=0$, fixed phase, no detuning, ${\gamma_q},{\gamma_m}>0$$\min\bigl[{\gamma_q}(A-A_0)+{\gamma_m}(B-A_0)\bigr]=\dfrac{{\gamma_q}+{\gamma_m}}{2G}\arccos\dfrac{|{\gamma_q}-{\gamma_m}|}{{\gamma_q}+{\gamma_m}}$
slow noise, second order (static offsets, no losses)$r=\dfrac{2\,\mathbb E(\xi_q-\xi_m)^2|\zeta|^2+\mathbb E(\xi_qA+\xi_mB)^2}{6}$; qubit only: $r\ge\dfrac{\pi^2\sigma_q^2}{96G^2}$
noise on both elements, fixed phase, no detuningsquare pulse optimal exactly for $\varrho_c\le\sigma_m^2/\sigma_q^2\le1/\varrho_c$, $\varrho_c=0.12645352248541\ldots$ (the bang–bang crossing)
correlated offsets, $C=\mathbb E(\xi_q\xi_m)\ge0$$6r=2V|\zeta|^2+\sigma_q^2A^2+2CAB+\sigma_m^2B^2$, $V=\sigma_q^2+\sigma_m^2-2C$; square pulse optimal if $\dfrac\pi2\sqrt{(\sigma_q^2+C)(\sigma_m^2+C)}\ge V$; fixed phase, no detuning: exactly on $[\varrho_c(\kappa),1/\varrho_c(\kappa)]$, $\varrho_c(0.5)=0.0272$, $\varrho_c(1)=0.00948$
offsets and relaxation of both elementssquare pulse optimal if $Vz_0\le\sqrt{\bigl((\sigma_q^2+C)A_0+{\gamma_q}\bigr)\bigl((\sigma_m^2+C)A_0+{\gamma_m}\bigr)}$
slow noise on the qubit and memory decay$\min J=\dfrac{\pi^2\sigma_q^2}{96G^2}+\dfrac{\sigma_q\sqrt{\pi{\gamma_m}}}{6G^{3/2}}+O({\gamma_m}/G)$ for ${\gamma_m} G\ll\sigma_q^2\ll\sqrt{{\gamma_m}}G^{3/2}$; square pulse optimal for $\sigma_q^2\le7.1\,{\gamma_m} G$, not for $\sigma_q^2\ge7.92\,{\gamma_m} G$; fixed phase, no detuning: optimal exactly for $\sigma_q^2\le\Lambda_c{\gamma_m} G$, $\Lambda_c=7.9078851606875\ldots$; with qubit relaxation $r={\gamma_q}/{\gamma_m}$: $\Lambda_c(1)=8.366$, $\Lambda_c(1000)=67.245$, $\Lambda_c(r)=2\sqrt r+\pi/2+4/\sqrt3+O(r^{-1/2})$ given Lemma 6, non-decreasing and concave in $r$; any phase and detuning: $2\sqrt r+\pi/2\lesssim\Lambda_c^{\rm all}(r)\lesssim2\sqrt r+\pi/2+4/\sqrt3$ up to $O(r^{-1/2})$

References

1

Q. Ansel, S. J. Glaser and D. Sugny, Selective and robust time-optimal rotations of spin systems, J. Phys. A 54, 085204 (2021); we used arXiv:2010.12454v2, §4.

2

J. Bae, J. H. Bai, B.-Y. Lee, J.-Y. Lee and C.-H. Lee, On the reachability of 3-dimensional paths with a prescribed curvature bound, arXiv:2403.18707v3 (2025); Theorems 3 and 5, and p. 10.

3

E. Barnes, F. A. Calderon-Vargas, W. Dong, B. Li, J. Zeng and F. Zhuang, Dynamically corrected gates from geometric space curves, Quantum Sci. Technol. 7, 023001 (2022); we used arXiv:2103.16015v1, Eqs. (6)–(8) and §II.B.

4

M. Bergonzoni, S. Jandura and G. Pupillo, iSWAP gate with polar molecules: robustness criteria for entangling operations, Phys. Rev. A 112, 032621 (2025); we used arXiv:2502.21238v2.

5

D. Daems, A. Ruschhaupt, D. Sugny and S. Guérin, Robust quantum control by a single-shot shaped pulse, Phys. Rev. Lett. 111, 050404 (2013); we used arXiv:1304.4016, Eqs. (7)–(10).

6

L. E. Dubins, On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and tangents, Amer. J. Math. 79, 497–516 (1957). 🔎 find it

7

S. Jandura, J. D. Thompson and G. Pupillo, Optimizing Rydberg gates for logical-qubit performance, PRX Quantum 4, 020336 (2023); we used arXiv:2210.06879v3, Sections II–III and Appendix A.

8

E. Kyoseva, H. Greener and H. Suchowski, Detuning-modulated composite pulses for high-fidelity robust quantum control, Phys. Rev. A 100, 032333 (2019); we used arXiv:1811.02623v1.

9

X.-J. Lu, X. Chen, A. Ruschhaupt, D. Alonso, S. Guérin and J. G. Muga, Fast and robust population transfer in two-level quantum systems with dephasing noise and/or systematic frequency errors, Phys. Rev. A 88, 033406 (2013); we used arXiv:1305.6127v1.

10

C. Marín, Exact limits on writing a superconducting quantum memory: when can pulse shaping beat the resonant swap?, companion paper, Zenodo (2026), doi:10.5281/zenodo.24250501.

11

D. Mittenhuber, Dubins’ problem is intrinsically three-dimensional, ESAIM: Control Optim. Calc. Var. 3, 1–22 (1998); Theorem 1.1 and the planar and spatial shifts, pp. 2–3. 🔎 find it

12

H. Nelson and E. Barnes, Exact quantum speed limits for quantum gates from space curve geometry, Phys. Rev. A 114, 052426 (2026), doi:10.1103/28zn-ch8y; we used arXiv:2604.23031v1.

13

V. S. Patsko and A. A. Fedotov, Analytic description of a reachable set for the Dubins car, Trudy Inst. Mat. Mekh. UrO RAN 26(1), 182–197 (2020), in Russian; Eq. (2.6), §§3.3–3.4 and Theorem 1. 🔎 find it

14

V. S. Patsko and A. A. Fedotov, Three-dimensional reachable set for the Dubins car: foundation of analytical description, Commun. Optim. Theory 2022, Art. 23, 1–42, doi:10.23952/cot.2022.23; Eq. (3.5), §3.4, Property 8 (§3.7), Theorem 11.1.

15

P. Souères, A. Balluchi and A. Bicchi, Optimal feedback control for route tracking with a bounded-curvature vehicle, Proc. IEEE Int. Conf. Robotics and Automation (ICRA 2000), vol. 3, 2473–2478; Property 5 and the synthesis. 🔎 find it

16

J. M. Sullivan, Curves of finite total curvature, in Discrete Differential Geometry, Birkhäuser (2008); we used arXiv:math/0606007, §5 (Schur’s comparison theorem and its proof).

17

H. J. Sussmann, Shortest 3-dimensional paths with a prescribed curvature bound, Proc. 34th IEEE CDC (1995) 3306–3312; Theorem 1 and the proof on p. 3312. 🔎 find it

18

L. Van Damme, Q. Ansel, S. J. Glaser and D. Sugny, Robust optimal control of two-level quantum systems, Phys. Rev. A 95, 063403 (2017); we used arXiv:1704.07653.

19

J. Zeng and E. Barnes, Fastest pulses that implement dynamically corrected single-qubit phase gates, Phys. Rev. A 98, 012301 (2018); we used arXiv:1802.10241v2.

20

J. Zeng, C. H. Yang, A. S. Dzurak and E. Barnes, Geometric formalism for constructing arbitrary single-qubit dynamically corrected gates, Phys. Rev. A 99, 052321 (2019); we used arXiv:1811.04864v2, Eqs. (1)–(8).