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

Regular trajectories, pruning and quantum parity

Fix positive integers d,k,Cd, k, C and a real 0<ε≤1/20 < \varepsilon \le 1/2. There is an integer n0=n0(d,k,C,ε)n_0 = n_0(d, k, C, \varepsilon) such that, for every n≥n0n \ge n_0, every allowed QAC0\mathrm{QAC}^0 circuit of depth at most dd on at most C(n+1)kC(n+1)^k total qubits (with ancillary qubits initialized to ∣0⟩|0\rangle, one output qubit measured, and all other final registers discarded without restriction) has an input x∈{0,1}nx \in \{0,1\}^n on which its answer equals PARITYn(x)\mathrm{PARITY}_n(x) with probability strictly less than 1/2+ε1/2 + \varepsilon.

No constant-depth, polynomial-resource QAC0\mathrm{QAC}^0 circuit can implement the coherent fanout operation Fn∣y1,…,yn,b⟩=∣y1⊕b,…,yn⊕b,b⟩F_n|y_1, \dots, y_n, b\rangle = |y_1 \oplus b, \dots, y_n \oplus b, b\rangle, including Moore's original formulation in which ancillary qubits must be restored to zero.

At each fixed depth and fixed polynomial total-qubit bound, for every fixed A>0A > 0, eventually every allowed unitary state preparation ρ\rho from all-zero inputs (with arbitrary final ancillas discarded) has felinity 2∑ypypyˉ<n−A2 \sum_y p_y p_{\bar{y}} < n^{-A}. For every fixed 0<δ<10 < \delta < 1, eventually every such preparation and every integer kk with nδ≤k≤n/2n^\delta \le k \le n/2 also satisfy 12∥ρ−∣Dkn⟩⟨Dkn∣∥1>180k\frac{1}{2} \|\rho - |D_k^n\rangle\langle D_k^n|\|_1 > \frac{1}{80k}.

For symmetric functions fn:{0,1}n→{0,1}f_n : \{0,1\}^n \to \{0,1\} with either ρ(fn)≥nδ\rho(f_n) \ge n^\delta for some fixed δ>0\delta > 0 or pointwise-error-1/31/3 approximate degree Ω(n1/2+η)\Omega(n^{1/2+\eta}) for some fixed η>0\eta > 0 (including strict majority MAJn\mathrm{MAJ}_n), no fixed-depth, polynomial-total-qubit QAC0\mathrm{QAC}^0 family achieves success at least 1/2+1/(log⁡n)γ1/2 + 1/(\log n)^\gamma on every input at every sufficiently large length, for any fixed γ>0\gamma > 0. Under the same circuit conventions, adjoining unit-cost primitive reversible unbounded threshold gates strictly enlarges the class computed with worst-case success at least 2/32/3.

Along a regular trajectory from ζ\zeta (∥ζ∥≤1\|\zeta\| \le 1) for a circuit VV with a bounded number of analytic layers, for any fixed 0<k<K<10 < k < K < 1 and any operator EE with deg⁡E≤mk\deg E \le m^k and ∥E∥≤exp⁡(mk)\|E\| \le \exp(m^k), there is eventually an operator E′E' such that ∥VEζ−E′Vζ∥≤exp⁡(−mk)\|V E \zeta - E' V \zeta\| \le \exp(-m^k), deg⁡E′≤mK\deg E' \le m^K, and ∥E′∥≤exp⁡(mK)\|E'\| \le \exp(m^K).

If VV has a bounded number of analytic layers, A,BA, B are product tests at its initial and final endpoints, χ=V†PBVψ\chi = V^\dagger P_B V \psi with ∥ψ∥≤1\|\psi\| \le 1, and both trajectories from ψ\psi and χ\chi are regular for VV, then an exp⁡(−mp)\exp(-m^p) tail bound on ψ\psi for HA≥mhH_A \ge m^h (for all b<h<1,0<p<hb < h < 1, 0 < p < h) implies the same tail bound on χ\chi.

Inserting an arbitrary product projection PBP_B at the output of a regular trajectory and evolving backwards preserves regularity.

For any sequence of circuits U=Ut⋯U1U = U_t \cdots U_1 on mm qubits with a bounded number tt of analytic layers and any normalized product starting states ψ0\psi_0, retaining at each reflection layer precisely the gates with projection probability ∥Pgψj−1∥2≥m−6\|P_g \psi_{j-1}\|^2 \ge m^{-6} on the exact pruned-prefix state ψj−1\psi_{j-1} produces a circuit WW with a regular trajectory from ψ0\psi_0 satisfying ∥Uψ0−Wψ0∥≤2tm−2\|U\psi_0 - W\psi_0\| \le 2tm^{-2}.

Methods

Product-reflection representation of generalized Toffoli gates Rg=I−2PgR_g = I - 2P_g and decomposition of each physical circuit layer into two analytic layers: a product of disjoint reflections and a tensor product of one-qubit unitaries.

Source excerpts

“An analytic layer will be either a product of such reflections on pairwise disjoint supports, or a tensor product of one-qubit unitaries. Splitting a physical layer into its two gate types preserves its operation, because its gates have disjoint supports. Thus a physical circuit of depth d gives at most 2d analytic layers.” · Show in paper

Asymptotic regularity condition for trajectory sequences, requiring exponentially decaying tails ∥1HS≥mhξj−1∥≤exp⁡(−mp)\|\mathbf{1}_{H_S \ge m^h} \xi_{j-1}\| \le \exp(-m^p) for every set SS of at most mbm^b reflections in a reflection layer.

Source excerpts

“The sequence of these trajectories is regular if the following holds at every reflection layer Uj: for every fixed triple 0 ≤ b < h < 1, 0 < p < h, eventually, simultaneously for every set S of at most mb gates of that layer, ∥1HS≥mh ξj−1∥ ≤ exp(−mp). (2.1)” · Show in paper

Integer zero filter Q(x)Q(x) constructed from normalized Chebyshev polynomials Sℓ(x)S_\ell(x) multiplied by exact roots ∏i=1a(1−x/i)\prod_{i=1}^a (1 - x/i) to approximate 1j=0\mathbf{1}_{j=0} on integers 0≤j≤mL0 \le j \le m^L while controlling global growth on [0,m][0, m].

Source excerpts

“Define Q(x) = Sℓ(x) ∏a i=1 (1 − x/i).” · Show in paper

Nonnegative tail detector polynomial T(x)=(MJ(1−Sℓ(x)2))2T(x) = (M_J(1 - S_\ell(x)^2))^2 constructed by composing a squared Chebyshev-based band indicator with the odd-degree majority polynomial MJM_J.

Source excerpts

“Take T(x) = (MJ(B(x)))2.” · Show in paper

State-dependent pruning rule that retains reflections in each layer whose projection probability on the exact pruned-prefix state satisfies ∥Pgψj−1∥2≥m−6\|P_g \psi_{j-1}\|^2 \ge m^{-6}.

Source excerpts

“Retain precisely the gates satisfying qg := ∥Pgψ∥2 ≥ m−6. (6.3)” · Show in paper

Reference-register copying of the uniform coherent input ∣+⟩⊗n|+\rangle^{\otimes n} via CNOTs to convert worst-case parity success into a signed acceptance expectation ⟨Φ,ZRPBΦ⟩≤−ε\langle \Phi, Z_R P_B \Phi \rangle \le -\varepsilon.

Source excerpts

“Attach a register R of n reference qubits, initially zero, and let COPY be the layer of disjoint CNOTs from input qubit Ii to reference qubit Ri. Our normalized product starting state is ψ0 = |+⟩⊗n I ⊗ |0⟩⊗a ⊗ |0⟩⊗n R” · Show in paper