Braverman's real gate approximators #
Over ℝ, the conjunction of bits b₁, …, b_K is approximated by
∏_c (1 - ∑_{j ∈ S c} (1 - bⱼ)) for a family S of sets of arguments (andApprox), and the
disjunction by its De Morgan dual 1 - ∏_c (1 - ∑_{j ∈ S c} bⱼ) (orApprox). At bits, a factor
is 1 minus the number of zeros in its set, so the approximator is exact unless some argument is
0 and no set contains exactly one zero (AndMiss, andApprox_bitVal). Negation is exactly
1 - b. Each approximator multiplies degrees by the number of sets (gateApprox_mem_lowDegree)
and its values stay bounded on bounded arguments (abs_andApprox_le, abs_orApprox_le,
abs_gateApprox_le).
Following Braverman, Lemma 8, take s sets at each of the rates
2⁻¹, …, 2⁻ℓ. When the arguments have z ≥ 1 zeros and 2 ^ r ≤ z < 2 ^ (r + 1), a random set of
rate p = 2 ^ -(r + 1) contains exactly one zero with probability
z p (1 - p) ^ (z - 1) ≥ 1 / (2e) > 9 / 50. So the s sets of that rate all miss with probability
at most (41 / 50) ^ s, and averaging over the sets against any weights ν on the inputs fixes
sets whose misses have weight at most (41 / 50) ^ s times the total (exists_sum_gateMiss_le).
Braverman pads the fan-in to k = 2 ^ ℓ and also needs the rate 1 / (2k) when all k
arguments are 0; here the fan-in is assumed to be less than 2 ^ ℓ, which avoids that case. We
choose the sets greedily, one at a time, each by an average over a single random set; the
constant 41 / 50 is Braverman's 0.82.
The sup-norm bounds are our own; they replace Braverman, Proposition 9 in the
circuit-level approximation (Braverman.exists_lowDegree_approx).
Main definitions #
Cslib.Circuits.Boolean.Braverman.andApprox,Cslib.Circuits.Boolean.Braverman.orApprox,Cslib.Circuits.Boolean.Braverman.gateApprox: the real approximators for a family of sets.Cslib.Circuits.Boolean.Braverman.AndMiss,Cslib.Circuits.Boolean.Braverman.GateMiss: the bits at which an approximator may err.
Main results #
Cslib.Circuits.Boolean.Braverman.gateApprox_bitVal: exactness away from misses.Cslib.Circuits.Boolean.Braverman.gateApprox_mem_lowDegree: the degree bound.Cslib.Circuits.Boolean.Braverman.abs_gateApprox_le: the sup-norm bound.Cslib.Circuits.Boolean.Braverman.exists_sum_gateMiss_le: Braverman's choice of sets.
References #
The AND approximator multiplies degrees by the number of sets.
The OR approximator multiplies degrees by the number of sets.
The approximator of a gate of the unbounded fan-in basis for a family S of sets of its
arguments: 1 - v for negation, andApprox and orApprox for conjunction and disjunction.
Instances For
The bits at which the approximator of a gate may err: never for negation, and AndMiss for
conjunction and, at the negated bits, for disjunction.
Instances For
Exactness of the gate approximators at bits where they do not miss.
Degrees of the gate approximators: a gate with a nonempty family of sets multiplies degrees by the number of sets.
Sup-norm bound for the gate approximators (our bound, replacing
Braverman, Proposition 9): with a nonempty family of t sets, a gate of fan-in
less than 2 ^ ℓ on arguments of absolute value at most 2 ^ E has absolute value at most
2 ^ (t (E + ℓ + 1)).
Choosing the sets #
Braverman's choice of sets for an AND gate Braverman, Lemma 8: for every
nonnegative weights ν on a finite set of points, at which the K < 2 ^ ℓ arguments of the gate
take the values u x, there are s sets at each of the rates 2⁻¹, …, 2⁻ℓ whose misses have
weight at most (41 / 50) ^ s times the total.
Braverman's choice of sets for a gate Braverman, Lemma 8: for every
nonnegative weights ν on a finite set of points, at which the arguments of a gate of fan-in less
than 2 ^ ℓ take the values u x, there are s sets at each of the rates 2⁻¹, …, 2⁻ℓ whose
misses have weight at most (41 / 50) ^ s times the total.