By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

The entropy photon-number inequality

For any finite number of modes n≥1n \ge 1, two independent nn-mode bosonic states ρA,ρB\rho_A, \rho_B with finite mean energy (allowing arbitrary entanglement among the modes within either input), and any transmissivity 0≤η≤10 \le \eta \le 1, the reduced state ρC\rho_C after passive beam-splitter mixing satisfies the entropy photon-number inequality g−1(S(ρC)/n)≥ηg−1(S(ρA)/n)+(1−η)g−1(S(ρB)/n)g^{-1}(S(\rho_C)/n) \ge \eta g^{-1}(S(\rho_A)/n) + (1 - \eta) g^{-1}(S(\rho_B)/n), where g(t)=(t+1)log⁡(t+1)−tlog⁡tg(t) = (t+1)\log(t+1) - t\log t.

For every finite n≥1n \ge 1, finite s,NB≥0s, N_B \ge 0, and 0≤η≤10 \le \eta \le 1, the exact minimum output entropy of the nn-fold tensor power of an identical thermal attenuator Eη,NB⊗n(ρ)\mathcal{E}_{\eta, N_B}^{\otimes n}(\rho) over all finite-energy nn-mode states ρ\rho (possibly entangled across modes) with S(ρ)=nsS(\rho) = ns is ng(ηg−1(s)+(1−η)NB)n g(\eta g^{-1}(s) + (1 - \eta) N_B), attained by the product thermal input τg−1(s)\tau_{g^{-1}(s)}.

For a degraded two-receiver pure-loss bosonic broadcast channel with 0≤ηC<ηB0 \le \eta_C < \eta_B, ηB+ηC≤1\eta_B + \eta_C \le 1, and mean photon constraint Nˉ<∞\bar{N} < \infty averaged over uses and the entire codebook (allowing codewords entangled across uses), the classical capacity region in nats per use is the closed convex hull of nonnegative pairs (RB,RC)(R_B, R_C) satisfying RB≤g(ηBβNˉ)R_B \le g(\eta_B \beta \bar{N}) and RC≤g(ηCNˉ)−g(ηCβNˉ)R_C \le g(\eta_C \bar{N}) - g(\eta_C \beta \bar{N}) for some 0≤β≤10 \le \beta \le 1.

If four families of positive diagonal quadratic forms with weights mf(x)exm f(x) e^x, mf(x)e−xm f(x) e^{-x}, mf(y)eym f(y) e^y, and mf(y)e−ym f(y) e^{-y} (where f(x)=x/sinh⁡xf(x) = x / \sinh x, f(0)=1f(0) = 1) satisfy comparison on a linear test space T\mathcal{T} supported on finitely many values of (m,x,y)(m, x, y) and invariant under scalar multiplication by functions of (m,x,y)(m, x, y), then the family with weight F=mf(x)f(y)/f(x+y)F = m f(x) f(y) / f(x + y) also satisfies comparison.

At each regularized minimum of JζJ_\zeta, the slice map LiL_i on the centered output space satisfies the component entropy-Hessian estimate ∥LiZ∥i2≤1+ε1−κ∥Z∥02\|L_i Z\|_i^2 \le \frac{1 + \varepsilon}{1 - \kappa} \|Z\|_0^2 for i=1,2i = 1, 2.

Methods

Variational contradiction with linked regularizations: assuming a strict violation of the inequality, the authors add an auxiliary thermal port, a small thermal replacement of the output, a relative-entropy penalty, and a fourth-moment penalty to obtain faithful Gibbs minimizers with finite seventh moments.

Source excerpts

“Assuming a strict violation, we add a thermal port, a small thermal replacement of the output, a relative-entropy penalty, and a fourth-moment penalty. The resulting functional has faithful Gibbs minimizers with enough moment control for all subsequent traces.” · Show in paper

“For \varepsilon, \zeta > 0 we minimize, over input states of finite fourth moment, the functional J_\zeta(\rho_1, \rho_2) = \frac{S(\rho_0) - n g(r_0)}{h_0} - \sum_i w_i \frac{S(\rho_i) - n g(r_i)}{h_i} + \varepsilon \sum_i \frac{w_i}{h_i} D(\rho_i \| \tau_{r_i}) + \zeta \sum_i Tr \rho_i W_i^4.” · Show in paper

Four-weight interpolation of positive diagonal quadratic forms using parallel-sum calculus, logarithmic means, and a Stieltjes representation derived via Herglotz representation and holomorphic extension.

Source excerpts

“The operations below belong to the parallel-sum and operator-mean calculus of Anderson-Duffin and Kubo-Ando [1, 25]. We give direct form proofs, including the limits needed for unbounded weights.” · Show in paper

“A Stieltjes representation of -B' turns the reciprocal into a limit of positive combinations of 1/m and reciprocals of \theta(ms) + (1-\theta)(mt) + rm, with r >= 0. The parallel-sum argument above will then construct E from the three known weights.” · Show in paper

Weighted trace-norm and form-domain generator pairing: evaluating the bosonic thermal-relaxation (quantum Ornstein-Uhlenbeck) generator as a trace-class expression at regularized minima without invoking log-Sobolev inequalities or entropy differentiation along an unbounded semigroup.

Source excerpts

“Here we use it as a trace-class expression at the regularized minima. The proof requires the weighted trace identities established below, rather than a log-Sobolev inequality or differentiation of entropy along an unbounded semigroup.” · Show in paper

Finite-alphabet Gaussian discretization and Markov-inequality deterministic code extraction for the degraded pure-loss bosonic broadcast channel under an averaged codebook photon constraint.

Source excerpts

“Partition the disk {|z| <= \ell} into finitely many sets of diameter at most 1/\ell, with its complement as one further cell, and let T_\ell, Z_\ell be the corresponding conditional expectations of T, Z.” · Show in paper

“Markov's inequality gives Pr{C_n > \bar{N}} + Pr{E_n > \sqrt{e_n}} <= N''/\bar{N} + \sqrt{e_n} < 1 for all sufficiently large n. Hence a deterministic code satisfies the entire-codebook photon constraint and has vanishing average error.” · Show in paper