forked from Indra4091/falcon_python
-
Notifications
You must be signed in to change notification settings - Fork 0
/
samplerz.py
153 lines (130 loc) · 3.96 KB
/
samplerz.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
# Importing dependencies
from math import floor
# Use high-quality randomness
# The "secrets" library could also work (Python >= 3.6)
from os import urandom
# Upper bound on all the values of sigma
MAX_SIGMA = 1.8205
INV_2SIGMA2 = 1 / (2 * (MAX_SIGMA ** 2))
# Precision of RCDT
RCDT_PREC = 72
# ln(2) and 1 / ln(2), with ln the natural logarithm
LN2 = 0.69314718056
ILN2 = 1.44269504089
# RCDT is the reverse cumulative distribution table of a distribution that
# is very close to a half-Gaussian of parameter MAX_SIGMA.
RCDT = [
3024686241123004913666,
1564742784480091954050,
636254429462080897535,
199560484645026482916,
47667343854657281903,
8595902006365044063,
1163297957344668388,
117656387352093658,
8867391802663976,
496969357462633,
20680885154299,
638331848991,
14602316184,
247426747,
3104126,
28824,
198,
1]
# C contains the coefficients of a polynomial that approximates exp(-x)
# More precisely, the value:
# (2 ** -63) * sum(C[12 - i] * (x ** i) for i in range(i))
# Should be very close to exp(-x).
# This polynomial is lifted from FACCT: https://doi.org/10.1109/TC.2019.2940949
C = [
0x00000004741183A3,
0x00000036548CFC06,
0x0000024FDCBF140A,
0x0000171D939DE045,
0x0000D00CF58F6F84,
0x000680681CF796E3,
0x002D82D8305B0FEA,
0x011111110E066FD0,
0x0555555555070F00,
0x155555555581FF00,
0x400000000002B400,
0x7FFFFFFFFFFF4800,
0x8000000000000000]
def basesampler(randombytes=urandom):
"""
Sample z0 in {0, 1, ..., 18} with a distribution
very close to the half-Gaussian D_{Z+, 0, MAX_SIGMA}.
Takes as (optional) input the randomness source (default: urandom).
"""
u = int.from_bytes(randombytes(RCDT_PREC >> 3), "little")
z0 = 0
for elt in RCDT:
z0 += int(u < elt)
return z0
def approxexp(x, ccs):
"""
Compute an approximation of 2^63 * ccs * exp(-x).
Input:
- a floating-point number x
- a scaling factor ccs
Both inputs x and ccs MUST be positive.
Output:
- an integral approximation of 2^63 * ccs * exp(-x).
"""
# y, z are always positive
y = C[0]
# Since z is positive, int is equivalent to floor
z = int(x * (1 << 63))
for elt in C[1:]:
y = elt - ((z * y) >> 63)
z = int(ccs * (1 << 63)) << 1
y = (z * y) >> 63
return y
def berexp(x, ccs, randombytes=urandom):
"""
Return a single bit, equal to 1 with probability ~ ccs * exp(-x).
Both inputs x and ccs MUST be positive.
Also takes as (optional) input the randomness source (default: urandom).
"""
s = int(x * ILN2)
r = x - s * LN2
s = min(s, 63)
z = (approxexp(r, ccs) - 1) >> s
for i in range(56, -8, -8):
p = int.from_bytes(randombytes(1), "little")
w = p - ((z >> i) & 0xFF)
if w:
break
return (w < 0)
def samplerz(mu, sigma, sigmin, randombytes=urandom):
"""
Given floating-point values mu, sigma (and sigmin),
output an integer z according to the discrete
Gaussian distribution D_{Z, mu, sigma}.
Input:
- the center mu
- the standard deviation sigma
- a scaling factor sigmin
- optional: the randomness source randombytes (default: urandom)
randombytes(k) should output k pseudorandom bytes
The inputs MUST verify 1 < sigmin < sigma < MAX_SIGMA.
Output:
- a sample z from the distribution D_{Z, mu, sigma}.
"""
s = int(floor(mu))
r = mu - s
dss = 1 / (2 * sigma * sigma)
ccs = sigmin / sigma
while(1):
# Sampler z0 from a Half-Gaussian
z0 = basesampler(randombytes=randombytes)
# Convert z0 into a pseudo-Gaussian sample z
b = int.from_bytes(randombytes(1), "little")
b &= 1
z = b + (2 * b - 1) * z0
# Rejection sampling to obtain a true Gaussian sample
x = ((z - r) ** 2) * dss
x -= (z0 ** 2) * INV_2SIGMA2
if berexp(x, ccs, randombytes=randombytes):
return z + s