Datasets:
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
mathematics
optimization
gradient-descent
convex-optimization
learning-theory
recursive-self-improvement
File size: 6,202 Bytes
868c741 | 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 | #!/usr/bin/env python3
"""Recreate mathematical illustrations; decimal plot data are not proofs."""
from __future__ import annotations
import csv
import math
from pathlib import Path
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from verify_witness import actual_trajectory
ROOT=Path(__file__).resolve().parents[1]
FIG=ROOT/'figures'
DATA=ROOT/'data'
FIG.mkdir(exist_ok=True);DATA.mkdir(exist_ok=True)
plt.rcParams.update({
'font.family':'DejaVu Sans','font.size':9,'axes.spines.top':False,
'axes.spines.right':False,'axes.labelcolor':'#193449','text.color':'#193449',
'axes.edgecolor':'#84949E','axes.titlesize':10,'figure.titlesize':12,
'pdf.fonttype':42})
TEAL='#006978';PLUM='#72476E';GREEN='#397D5A';GREY='#8899A6'
def save(fig,name):
fig.savefig(FIG/(name+'.pdf'),bbox_inches='tight')
fig.savefig(FIG/(name+'.png'),dpi=210,bbox_inches='tight')
plt.close(fig)
def witness():
m=32;M=m+1
a,mu,h,states,values=actual_trajectory(m)
vals=np.array([float(x) for x in values])
dec=np.array([float(values[i]-values[i+1]) for i in range(M)])
curv=np.array([float(values[i]-2*values[i+1]+values[i+2]) for i in range(m)])
fig,ax=plt.subplots(1,3,figsize=(8.4,2.65),layout='constrained')
ax[0].plot(vals,color=TEAL,lw=2);ax[0].set(title='Loss decreases',xlabel='Iteration k',ylabel=r'$H_a(x_k)$')
ax[1].plot(dec*1e3,color=PLUM,lw=2);ax[1].set(title='Improvements increase',xlabel='Step k',ylabel=r'$10^3 D_k$')
ax[2].plot(curv*1e5,color=TEAL,lw=2,label='Exact rational trajectory')
ax[2].axhline(-float(a*a)*885/512*1e5,color=PLUM,ls='--',label='Proved upper bound')
ax[2].set(title='Strict negative curvature',xlabel='Index k',ylabel=r'$10^5\kappa_k$')
ax[2].legend(fontsize=6.8,loc='center',bbox_to_anchor=(.50,.59))
for q in ax:q.grid(axis='y',alpha=.17)
fig.suptitle(f'm = 32; condition number = {128*M:,}; normalized step = {float(h):.8f}')
save(fig,'witness_m32')
with (DATA/'witness_m32_plot_data.csv').open('w',newline='') as f:
w=csv.writer(f);w.writerow(['k','loss_decimal_approx','decrement_decimal_approx','second_difference_decimal_approx'])
for k,v in enumerate(vals):w.writerow([k,v,dec[k] if k<len(dec) else '',curv[k] if k<len(curv) else ''])
def information():
mu=.25;L=1.;b=.5;v=1.;eta=.5;t=5
x=np.linspace(-4.5,4.5,700)
f=lambda x,a: np.where(x<0,v*x+a*x*x/2,v*x+b*x*x/2)
fig,ax=plt.subplots(1,2,figsize=(8.2,3.0),layout='constrained')
ax[0].plot(x,f(x,mu),color=TEAL,lw=2,label='Hidden curvature 1/4')
ax[0].plot(x,f(x,L),color=PLUM,lw=2,ls='--',label='Hidden curvature 1')
ax[0].axvspan(0,4.5,color='#E8EEF0',zorder=-1)
ax[0].axvline(0,color=GREY,lw=.8)
ax[0].set(title='Identical observed half-line',xlabel='Parameter x',ylabel='Objective value')
ax[0].text(1.0,5.8,'Same values\nand gradients',fontsize=8)
ax[0].legend(fontsize=7,loc='upper left')
n=np.arange(0,21)
for a,col,lab in [(mu,TEAL,'Remaining gap = 2'),(L,PLUM,'Remaining gap = 1/2')]:
gain=v*v/(2*a)*(1-(1-eta*a)**(2*n))
ax[1].plot(n,gain,color=col,lw=2,label=lab)
ax[1].set(title='Different future improvement',xlabel='Additional gradient steps',ylabel='Cumulative improvement')
ax[1].legend(fontsize=8,loc='lower right')
for q in ax:q.grid(alpha=.14)
save(fig,'information_barrier')
def smoothing():
m=np.array([1,2,4,8,16,32,64,128,256,512])
M=m+1.;a=1/(8*M)
sigma=a/(2*np.sqrt(2*np.log(64/a**2)))
fig,ax=plt.subplots(1,2,figsize=(8.1,3),layout='constrained')
ax[0].loglog(M,sigma,'o-',color=TEAL,label='Quadratic-chamber theorem')
ax[0].set(xlabel='M = m + 1',ylabel='Gaussian standard deviation',title='Certified analytic smoothing width')
ax[1].loglog(M,128*M,'o-',color=PLUM)
ax[1].set(xlabel='M = m + 1',ylabel='Exact condition number',title='Linear conditioning for long runs')
for q in ax:q.grid(which='both',alpha=.15)
save(fig,'analytic_scales')
def boundaries():
hc=1.75;hi=(19-math.sqrt(17))/8;hs=1.863425364363717532
fig,ax=plt.subplots(1,2,figsize=(8.4,3.2),layout='constrained',
gridspec_kw={'width_ratios':[1.35,1]})
names=['Every curvature nonnegative','Two-step compensation',
'No adjacent negative curvatures','Stable objective descent']
values=[hc,hi,hs,2.]
colors=[TEAL,PLUM,GREEN,GREY]
for i,(name,v,col) in enumerate(zip(names,values,colors)):
y=3-i
ax[0].barh(y,v-1.65,left=1.65,color=col,height=.28,alpha=.85)
ax[0].text(1.653,y+.24,name,fontsize=7.8,va='bottom')
ax[0].text(v+.004,y,f'{v:.6f}' if i in (1,2) else f'{v:g}',
fontsize=7.1,va='center')
ax[0].set(xlim=(1.65,2.065),ylim=(-.45,3.7),yticks=[],
xlabel=r'Normalized step $h=\eta L$',title='Different universal guarantees')
ax[0].set_xticks([1.7,1.8,1.9,2.])
for spine in ['left','top','right']:ax[0].spines[spine].set_visible(False)
ax[1].axvspan(1.855,hi,color=PLUM,alpha=.15)
ax[1].axvspan(hi,hs,color=GREEN,alpha=.18)
ax[1].axvspan(hs,1.868,color=GREY,alpha=.2)
ax[1].axvline(hi,color=PLUM,lw=1.5);ax[1].axvline(hs,color=GREEN,lw=1.5)
ax[1].text((1.855+hi)/2,.63,'Compensation\nguaranteed',ha='center',fontsize=8)
ax[1].text((hi+hs)/2,.30,'Compensation\nmay fail;\nadjacent negative\npairs still excluded',ha='center',fontsize=7.3)
ax[1].text((hs+1.868)/2,.63,'Adjacent\nnegatives\npossible',ha='center',fontsize=8)
ax[1].text(hi,.98,r'$h_{\mathrm{iso}}$',ha='center',va='top',color=PLUM)
ax[1].text(hs,.98,r'$h_*$',ha='center',va='top',color=GREEN)
ax[1].set(xlim=(1.855,1.868),ylim=(0,1),yticks=[],xlabel=r'Normalized step $h$',
title='The two nearby boundaries')
ax[1].set_xticks([1.856,1.860,1.864,1.868])
ax[1].tick_params(axis='x',labelsize=7.8)
for spine in ['left','top','right']:ax[1].spines[spine].set_visible(False)
save(fig,'phase_boundaries')
if __name__=='__main__':
witness();information();smoothing();boundaries()
print('Created witness_m32, information_barrier, analytic_scales, phase_boundaries (PDF and PNG).')
|