chunk_uid
stringlengths
40
40
chunk_type
stringclasses
2 values
chunk_index
int64
0
6.71k
total_chunks
int64
1
6.71k
section_title
stringlengths
1
157
embed_text
stringlengths
1
83.3k
spans
dict
paper_doi
stringlengths
0
63
paper_id_arxiv
stringlengths
9
16
title
stringlengths
7
245
authors
listlengths
1
768
categories
listlengths
1
7
year
int64
2k
2.02k
language
stringclasses
2 values
discipline
stringclasses
8 values
dense_vector
listlengths
1.02k
1.02k
e8703c66f3ce4ddf8c07388ae2706af9b510938d
subsection
25
50
Main Lemma: the dynamical system preserves a parabolic orbifold
We have reduced our equation to \beta +\gamma =\frac{3}{2} , with \frac{1}{2}<\beta <\gamma <1, but now the same argument used for \alpha shows that this is possible if and only if \beta =\frac{2}{3} , since otherwise the pole of order \beta k would not be a branched point, which leads us to an absurd. Calling p_1,p_2,...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.0030839319806545973, 0.011458419263362885, -0.0163255762308836, -0.07030678540468216, 0.00715960469096899, -0.05547644570469856, 0.056178294122219086, 0.026166696101427078, 0.04635243117809296, 0.015089715830981731, -0.005168494768440723, 0.0327426977455616, -0.026365043595433235, 0.022...
369a165ea19e5fc09dbd87fb79c8358d60954f57
subsection
26
50
Maps preserving a parabolic orbifold
In this chapter we will discuss the main consequences of Lemma REF . We have a rational map f:\mathbb {P}^1\rightarrow \mathbb {P}^1 that has an invariant orbifold \mathcal {O}, i.e. there exists a map \tilde{f}:\mathcal {O}\rightarrow \mathcal {O} such that the following diagram commutes\begin{}[column sep=2.5pc,row s...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.06836052238941193, 0.057160384953022, 0.02175939828157425, -0.03308160975575447, -0.005153742618858814, -0.03778139501810074, -0.010322744026780128, 0.03772035986185074, 0.041107866913080215, 0.014717349782586098, -0.029648326337337494, 0.020737042650580406, -0.01442742720246315, -0.010...
36706b94873f1a8b3bce3f13b1d4407847f43b27
subsection
27
50
Maps preserving a parabolic orbifold
We have a rational map f:\mathbb {P}^1\rightarrow \mathbb {P}^1 that has an invariant orbifold \mathcal {O}, i.e. there exists a map \tilde{f}:\mathcal {O}\rightarrow \mathcal {O} such that the following diagram commutes\begin{}[column sep=2.5pc,row sep=2pc] \mathcal {O}{r}{\tilde{f}} {d}[swap]{p} & \mathcal {O}{d}{p} ...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.0670648142695427, 0.04381486400961876, 0.03783456236124039, -0.00779193639755249, 0.01183855626732111, -0.05382271483540535, -0.012395395897328854, 0.01923765428364277, 0.042624905705451965, 0.029489601030945778, -0.028604760766029358, -0.010656226426362991, 0.008947568945586681, -0.018...
de5abf8267b9eb5f6ceaf6c5b80fe18beb8145fb
subsection
28
50
Construction of a torsor associated to a torsion line bundle
We begin recalling some general facts concerning line bundles over an orbifold \mathcal {O}. We say that \pi :L \rightarrow \mathcal {O} is a torsion line bundle of order n if L^{\otimes n} is trivial, i.e. if there exists an isomorphism of line bundles on \mathcal {O}\begin{}[column sep=2.5pc,row sep=2pc] L^{\otimes n...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/978-3-540-24899-6", "end": 1470, "openalex_id": "https://openalex.org/W4244873481", "raw": "G. LAUMON, Champs algǸbriques, Springer, 1999.", "source_ref_id": "2cc15835be9cb8eabf9c7f77f3f9d6551414772c", "start": 1041 ...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.03332020714879036, 0.02845338173210621, -0.01603458635509014, 0.003411735873669386, 0.007002735510468483, -0.030162109062075615, 0.0015428140759468079, 0.012136549688875675, 0.03701227903366089, 0.02459348551928997, -0.029017871245741844, -0.0291399247944355, 0.031611476093530655, 0.009...
dcc94f6563babf80d3ee72327dc03d56c22d9a7e
subsection
29
50
Construction of a torsor associated to a torsion line bundle
Note that L\mathop {|}_{U_i}^{\otimes n_i} is the trivial bundle on U_i since such a representation has order dividing n_i.If we set Z=\coprod \limits _i p^{-1}(x_i) and U=\mathcal {O}\setminus Z, then L is completely determined by the triple\left(L\mathop {\mid }\nolimits _{Z}, L \mathop {\mid }\nolimits _{U}, \phi \r...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.029602903872728348, 0.03717147931456566, -0.009491239674389362, 0.00465787947177887, -0.02281254716217518, -0.0029984384309500456, -0.007820354774594307, 0.01611374504864216, 0.02345343306660652, 0.03970451280474663, -0.041993398219347, -0.02281254716217518, 0.02679520472884178, 0.03137...
122211dfa80da0920a3190d82e4ea72ea253e249
subsection
30
50
Construction of a torsor associated to a torsion line bundle
Better still the singular points of \mathcal {E} lie over those of \mathcal {O}, and if the underlying space of \mathcal {O} is \mathbb {P}^1 with y a singular point of \mathcal {E} lying, in the above notation (REF ), over x_i then the local monodromy of \mathcal {E} at y is the kernel of \rho _i.From the exact sequen...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.016276733949780464, 0.02234809100627899, 0.018427642062306404, -0.015681801363825798, 0.008901099674403667, -0.013683439232409, -0.008046838454902172, -0.00536201661452651, 0.015262297354638577, 0.02841944992542267, -0.03963163122534752, 0.00275728153064847, -0.004271307494491339, 0.028...
63f91a6ce94b397042e2b5b6896098987600c41d
subsection
31
50
Construction of a torsor associated to a torsion line bundle
Therefore if L is a torsion line bundle of order n we have the following commutative diagram:\begin{}[column sep=2.5pc,row sep=2pc] L {r}{F_n} {d}[swap]{\pi } & \mathcal {O}\times {ld}{p_1} \\ \mathcal {O}& \end{}and for any \lambda \in * our \mu _n torsor is isomorphic to \mathcal {E}=F_n^{-1}(\mathcal {O}\times \lamb...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.02150946483016014, 0.016948238015174866, -0.02571982704102993, -0.01717706210911274, 0.022150171920657158, -0.026833437383174896, -0.011067458428442478, 0.002398839220404625, 0.03633726388216019, 0.020319579169154167, -0.03197435289621353, -0.03618471696972847, 0.0020594168454408646, 0....
0415c914a529b38d23978927aa79962f3517bfb7
subsection
32
50
Construction of a torsor associated to a torsion line bundle
We are going to prove that this property still holds if X is an orbifold whose underlying space is \mathbb {P}^1.Let us consider an orbifold \mathcal {O} modelled on \mathbb {P}^1 whose set of singular points is \lbrace x_1, \dots ,x_r\rbrace \subset \mathbb {P}^1, each x_i having finite weight n_i. This means that the...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/978-3-540-24899-6", "end": 731, "openalex_id": "https://openalex.org/W4244873481", "raw": "G. LAUMON, Champs algǸbriques, Springer, 1999.", "source_ref_id": "2cc15835be9cb8eabf9c7f77f3f9d6551414772c", "start": 302 }...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.05119387432932854, 0.02977665141224861, -0.02153925783932209, -0.028266463428735733, -0.005926346872001886, -0.015895118936896324, 0.03755640983581543, 0.010106061585247517, 0.015956135466694832, 0.01586460880935192, -0.02873934991657734, -0.0356343537569046, 0.007718742825090885, 0.008...
17036735975126995432a540e286bc7612178a27
subsection
33
50
Construction of a torsor associated to a torsion line bundle
Indeed if each representation \rho _i is trivial, then L\mid _{U_i} is trivial so the maps \phi are just gluing with the trivial line bundle on the moduli of U_i, i.e. the naive quotient of the \mu _{n_i} action, identified with open subset of \mathbb {P}^1, and so the kernel is a line bundle on \mathbb {P}^1. We know,...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.04177402704954147, 0.016386160627007484, 0.015325789339840412, -0.00456951092928648, 0.0069038523361086845, -0.012137048877775669, 0.003482440486550331, -0.007544651627540588, 0.02654740959405899, 0.04442876949906349, -0.036678146570920944, 0.005111139267683029, 0.0212226714938879, 0.05...
d4a036c810187816cc7f25b4b897ff3b78f1edb9
subsection
34
50
Construction of a torsor associated to a torsion line bundle
Therefore if L is a torsion line bundle of order n we have the following commutative diagram:\begin{}[column sep=2.5pc,row sep=2pc] L {r}{F_n} {d}[swap]{\pi } & \mathcal {O}\times {ld}{p_1} \\ \mathcal {O}& \end{}and for any \lambda \in * our \mu _n torsor is isomorphic to \mathcal {E}=F_n^{-1}(\mathcal {O}\times \lamb...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.029677676036953926, 0.005733361002057791, -0.019591843709349632, -0.0322868712246418, 0.008476066403090954, 0.0049399216659367085, -0.023192835971713066, 0.009475494734942913, 0.03555217757821083, 0.03025749698281288, -0.04080108180642128, -0.02886897884309292, -0.009628078900277615, 0....
022eaea07b0f734c41509aa5b07b5ece0739c87a
subsection
35
50
Holomorphic differentials on a parabolic orbifold
In this section we show how the construction of Lemma REF applies to the line bundle \Omega _{\mathcal {O}} of holomorphic differential forms over a parabolic orbifold.Around a non-space like point x_i, in the above notation (REF ), we have that \Omega _{\mathcal {O}}\mathop {|}_{U_i} is the \mu _{n_i}-module \mathcal...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1057, "openalex_id": "", "raw": "P. Deligne, Equations Differentielles a Points Singuliers Reguliers (Lecture Notes in Mathematics) (French Edition), Springer, 1970.", "source_ref_id": "3cb4d7c8ce85cd64adb636900d5c5109d17878...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.026508301496505737, 0.03751145675778389, -0.0067682391963899136, -0.0339098684489727, -0.0008512758649885654, -0.010736091062426567, 0.02054126188158989, 0.026096254587173462, 0.0576864592730999, 0.01272001676261425, -0.037938766181468964, -0.003128498326987028, 0.017702722921967506, 0....
38457f095a3b92aefe34800f27e078ec7760624e
subsection
36
50
Holomorphic differentials on a parabolic orbifold
H^0(\Omega _{\mathcal {O}}(\log D)^{\otimes k})1) From the definition of n all the local representations (REF ) have order that divides n, so we need only to show that deg(\Omega _{\mathcal {O}})=0 (resp. deg(\Omega _{\mathcal {O}}(\log D))=0, but this follows from the fact that deg(\Omega _{\mathcal {O}})=-\chi (\math...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.03917393088340759, 0.004000531043857336, 0.0011889090528711677, -0.03542128950357437, 0.010449432767927647, -0.014873280189931393, 0.022973498329520226, 0.004400965292006731, 0.03908240422606468, 0.024636253714561462, -0.0040768045000731945, 0.014751243405044079, 0.014530050568282604, 0...
71dd9e680ae2ff4eb4f3a4273069ad16813741fd
subsection
37
50
Holomorphic differentials on a parabolic orbifold
(X,\partial )=(\mathbb {P}^1, 0 + \infty )).We have seen in REF that if n= lcm\lbrace n_1,\dots ,n_r\rbrace , then \Omega _{\mathcal {O}} (resp. \Omega _{\mathcal {O}}(\log D)) is a torsion line bundle of order n and we know from Lemma REF that it defines a unique \mu _n torsor. On the other hand, the representation de...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.0630439966917038, 0.007689780555665493, 0.0157304834574461, -0.027921227738261223, -0.010214896872639656, 0.010817568749189377, -0.01919393613934517, 0.03579409793019295, 0.03905920311808586, 0.040920618921518326, -0.02897399477660656, 0.018781984224915504, 0.021131638437509537, 0.00858...
6ce444aa7e4c45adacd217f563746a4d555a64f3
subsection
38
50
Holomorphic differentials on a parabolic orbifold
Consequently we obtain the following commutative diagram,[column sep=2.5pc,row sep=2pc] E [r, ""] rd f*E d [r,"f"] E [d]O[r,"f"] Oand hence by composition we obtain a \mu _n-equivariant map F: E \rightarrow E.Corollary 4.4 Let f: \mathbb {P}^1\rightarrow \mathbb {P}^1 be a rational map of degree d>1 which satisfies As...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1655, "openalex_id": "", "raw": "J. H. Silverman, The Arithmetic of Elliptic Curves: v. 106 (Graduate Texts in Mathematics), Springer, 1994.", "source_ref_id": "3210686d82b7c7dac9c338de522eb922d690e577", "start": 1274 ...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.04079650714993477, 0.027233269065618515, 0.022335859015583992, -0.0011232770048081875, 0.009787194430828094, -0.08409512042999268, 0.03274095058441162, 0.0305592380464077, 0.011648516170680523, 0.040003154426813126, 0.007208806928247213, -0.012441866099834442, 0.0022694391664117575, 0.0...
d94466e10174ac5281915d767d7627651fcf0e80
subsection
39
50
Holomorphic differentials on a parabolic orbifold
It is a fact that the naive quotient E/G is isomorphic to \mathbb {P}^1, as the map f is necessarily ramified and the Riemann-Hurwitz formula gives \chi (E/G) >0.It follows that \# Ram_f=2\# G, so if the fiber of each p\in E/G consists of n_p distinct elements, each of order e_p=\# Stab_G(p), we can write Riemann-Hurwi...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1706, "openalex_id": "", "raw": "P. Deligne, Equations Differentielles a Points Singuliers Reguliers (Lecture Notes in Mathematics) (French Edition), Springer, 1970.", "source_ref_id": "3cb4d7c8ce85cd64adb636900d5c5109d17878...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.02231433428823948, 0.03846246376633644, -0.003592500928789377, -0.0225738026201725, 0.011653210036456585, -0.040599264204502106, 0.04560549184679985, 0.03748563677072525, 0.03189942240715027, 0.007658150978386402, -0.04206450283527374, -0.011813471093773842, 0.029289470985531807, 0.0064...
c6e09eabf4021c081f4eeb24772242ab093bef56
subsection
40
50
Holomorphic differentials on a parabolic orbifold
\Omega _{\mathcal {O}}(\log D)) is a torsion line bundle of order n:=lcm\lbrace \nu _f(x): x \notin D\rbrace . If we denote by p:\mathcal {O}\rightarrow \mathbb {P}^1 the natural projection, then \tilde{q}:=p^*q\in H^0(\Omega _{\mathcal {O}}^{\otimes k}) (resp. H^0(\Omega _{\mathcal {O}}(\log D)^{\otimes k})1) From th...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.044201429933309555, 0.019902853295207024, -0.0050291335210204124, -0.044293008744716644, 0.004662823397666216, -0.014568462036550045, 0.014736353419721127, -0.009989583864808083, 0.07094207406044006, 0.031319521367549896, -0.011500613763928413, 0.006311219185590744, 0.023993317037820816, ...
e81cb80f2f84de7d07049c9808adf887a4eba5f9
subsection
41
50
Holomorphic differentials on a parabolic orbifold
F:(\mathbb {P}^1, 0+\infty )\rightarrow (\mathbb {P}^1, 0+\infty )) such that the following diagram commutes:\begin{}[column sep=2.5pc,row sep=2pc] (X, \partial ) {r}{F} {d}[swap]{\pi } & (X, \partial ) {d}{\pi } \\ (\mathcal {O}, D) {r}{f} & (\mathcal {O},D) \end{}where (X,\partial )= (E, \emptyset ) (resp. (X,\partia...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.07474770396947861, 0.020166626200079918, 0.015300397761166096, -0.0015960161108523607, 0.0006573793943971395, -0.006696784403175116, -0.02010560780763626, 0.038014546036720276, 0.047441910952329636, 0.031088944524526596, -0.02988382801413536, 0.020837830379605293, 0.02257685735821724, -...
3cd0ac323020badc308fa0c956600092727f342a
subsection
42
50
Holomorphic differentials on a parabolic orbifold
This follows from the fact that the isomorphism f^*\Omega _{\mathcal {O}} \mathop {\rightarrow }\limits ^{\sim } \Omega _{\mathcal {O}} (resp. f^*\Omega _{\mathcal {O}}(\log D) \mathop {\rightarrow }\limits ^{\sim } \Omega _{\mathcal {O}}(\log D)) affords an isomorphism of \mu _n-torsors f^*E \mathop {\rightarrow }\lim...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.019317571073770523, 0.00010263890726491809, 0.013870199210941792, -0.03619374334812164, -0.0017137478571385145, -0.04083240404725075, -0.004665361251682043, 0.028549112379550934, 0.030929476022720337, 0.04925522953271866, -0.01202389132231474, -0.00037121667992323637, 0.01649470441043377,...
cb5ee29d3414701e4a05bd2edcf64ab9902acd9a
subsection
43
50
Holomorphic differentials on a parabolic orbifold
In the case n=2 the condition above is empty, hence is generic.\textbf {Remark} 4.5 The orbifolds listed in Corollary REF are the only one which can be realized as quotients of an elliptic curve E for the action of a group of automorphisms of E. In fact it is well known that the group of automorphisms of an elliptic c...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 381, "openalex_id": "", "raw": "J. H. Silverman, The Arithmetic of Elliptic Curves: v. 106 (Graduate Texts in Mathematics), Springer, 1994.", "source_ref_id": "3210686d82b7c7dac9c338de522eb922d690e577", "start": 0 ...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.045769933611154556, 0.02671438455581665, 0.031459201127290726, -0.0029254616238176823, 0.011587421409785748, -0.07512371987104416, 0.04671584814786911, 0.04634968563914299, 0.0001265824685106054, 0.022442525252699852, -0.010923757217824459, -0.012769811786711216, 0.022106878459453583, -...
58e8cfb810d6bb25f67aba44838bb3a631b3914f
subsection
44
50
Conclusions
We conclude this chapter formulating at first a structural theorem for the maps f with a parallel tensor q and finally discussing the invariant nature of q.Theorem 4.6 Let f: \mathbb {P}^1\rightarrow \mathbb {P}^1 be a rational map of degree d>1 which satisfies Assumption REF .Then, up to conjugation with an element of...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1446, "openalex_id": "", "raw": "A. Douady and J. H. Hubbard, A proof of Thurston's topological characterization of rational functions, Acta Math., 171 (1993), pp. 263–297.", "source_ref_id": "e9525ab135cc9efcce6bd3820525519...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.05086350813508034, 0.03655337914824486, 0.026164041832089424, -0.016674809157848358, 0.0017611145740374923, -0.01530176866799593, 0.018337713554501534, 0.011617444455623627, -0.002974920207634568, 0.040886081755161285, -0.01563740149140358, 0.01418808102607727, 0.006163424346596003, 0.0...
6f0c4c35c2f8300c32445b33dabc4fcb29de0c13
subsection
45
50
Conclusions
Consequently the only translations allowed are the solutions of\theta z \equiv z \;(mod\,\mathbb {Z}[\theta ])where \theta is a primitive n-th root of unity. [table] Table table contains the solutions of (REF ) lying in the fundamental domain of the elliptic curve. For n=2 they consist of the subgroup E[2] of 2-torsion...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1099, "openalex_id": "", "raw": "A. Douady and J. H. Hubbard, A proof of Thurston's topological characterization of rational functions, Acta Math., 171 (1993), pp. 263–297.", "source_ref_id": "e9525ab135cc9efcce6bd3820525519...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.01234625093638897, 0.00886671431362629, 0.019869985058903694, 0.000026528128728386946, 0.024234667420387268, -0.0522235743701458, 0.03937370330095291, 0.0011178774293512106, 0.019061146304011345, 0.019869985058903694, -0.02119770459830761, 0.014147063717246056, -0.005196413490921259, 0....
22e7eec5481b2c5428370c29b6f40bef957be620
subsection
46
50
Conclusions
In case (2) they are (up to sign) the Tchebycheff polynomials P_n(z) defined by P_n(cosz)=cos(nz), since the cosine function is the universal covering map of * which commutes with z \mapsto -z. Note that in each case we have shown that (modulo multiplication by an element of PGL_2() the action of f on \mathbb {P}^1, w...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.03163965418934822, 0.015308771282434464, -0.0031960015185177326, -0.015941869467496872, 0.023844154551625252, -0.024835754185914993, 0.0432947613298893, -0.016307998448610306, -0.011731385253369808, 0.061997853219509125, -0.03176169469952583, 0.02677318826317787, -0.020320162177085876, ...
6a21c21da4b90bfaed45b664e827dd71a079d75f
subsection
47
50
Conclusions
\textit {f} is obtained as quotient of such automorphisms under the action of a discrete group of automorphisms of the complex plane.\\ \begin{}[htb] \begin{}{|c|c|} \hline Orbifold \mathcal {O} preserved by \textit {f} & Automorphism of inducing \textit {f} \\ \hline (\infty ,\infty ) & z \mapsto nz; n\in \mathbb {Z} ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1514, "openalex_id": "", "raw": "J. H. Silverman, The Arithmetic of Elliptic Curves: v. 106 (Graduate Texts in Mathematics), Springer, 1994.", "source_ref_id": "3210686d82b7c7dac9c338de522eb922d690e577", "start": 1282 ...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.018459752202033997, 0.024607917293906212, 0.007155061233788729, 0.002337981015443802, 0.021861838176846504, -0.037743330001831055, 0.049643002450466156, 0.027460789307951927, 0.015309389680624008, 0.024775734171271324, -0.0237840935587883, -0.006804173346608877, -0.018795384094119072, 0...
2b15b84975450c924d2d8dd3d5cc7041fa5d96f0
subsection
48
50
Conclusions
Recall that when \alpha \notin \mathbb {Z} we say that the elliptic curve E has complex multiplication and if multiplication by a complex \alpha , (which must be an integer in some imaginary quadratic field, as \alpha \Lambda \subset \Lambda , see ) is allowed in E, then the complex structure of E is completely determi...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1363, "openalex_id": "", "raw": "A. Douady and J. H. Hubbard, A proof of Thurston's topological characterization of rational functions, Acta Math., 171 (1993), pp. 263–297.", "source_ref_id": "e9525ab135cc9efcce6bd3820525519...
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.03796348720788956, 0.005943238735198975, 0.020583899691700935, -0.018447691574692726, 0.022231832146644592, -0.07000662386417389, 0.023360971361398697, -0.012336607091128826, 0.0033988612703979015, 0.022674333304166794, -0.022323384881019592, 0.02912873588502407, -0.015006869100034237, ...
f1a944c86324382838d75e40b484459f906963c3
subsection
49
50
Conclusions
We have seen in REF that \tilde{q}=p^*q is a holomorphic section of \Omega _{\mathcal {O}}^{\otimes k}, hence \pi ^*\tilde{q} is a constant multiple of dz^{\otimes k}. It follows that the eigenvalue \lambda such that f^*q=\lambda q can be computed explicitly. Referring to [table] Table table , in the first two cases we...
{ "cite_spans": [] }
1806.10056
Dynamical systems with a parallel tensor
[ "Jacopo Garofali" ]
[ "math.DS" ]
2,018
en
Mathematics
[ -0.06100517138838768, 0.04678385704755783, 0.013511775992810726, -0.017776645720005035, 0.0140610970556736, -0.016006609424948692, 0.06396540254354477, 0.013412592932581902, 0.0276949442923069, 0.03924594819545746, -0.026794668287038803, 0.03576691448688507, -0.03042629174888134, 0.0212404...
e9f471e064e44d934a339209f85a5fd7a42e2dfb
abstract
0
121
Abstract
A long-standing problem in the theory of stochastic gradient descent (SGD) is to prove that its without-replacement version RandomShuffle converges faster than the usual with-replacement version. We present the first (to our knowledge) non-asymptotic solution to this problem, which shows that after a "reasonable" numbe...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0536840446293354, 0.04089631512761116, -0.03155730664730072, -0.05136455222964287, -0.012322304770350456, -0.005665208213031292, 0.024308891966938972, 0.005207413341850042, 0.02583487518131733, 0.022477714344859123, -0.004295639228075743, -0.007122520823031664, -0.016328006982803345, 0....
81a4b4118d3a978cd8019a99ba4d7f2292ea9c94
subsection
1
121
Introduction
We consider stochastic optimization methods for the finite-sum problemF\left(x\right) := \frac{1}{n} \sum _{i=1}^n f_i\left(x\right),where each function f_i : \mathbb {R}^d\rightarrow \mathbb {R} is smooth and convex, and the sum F is strongly convex. A classical approach to solving (REF ) is stochastic gradient descen...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 824, "openalex_id": "", "raw": "S. Sra, S. Nowozin, and S. J. Wright. Optimization for machine learning. Mit Press, 2012.", "source_ref_id": "67c860b48b8ff307afa0ce68a8ddcd840f0bac05", "start": 497 }, { "...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06592930853366852, 0.016940169036388397, -0.029393484815955162, -0.04966064170002937, 0.0027756697963923216, 0.008660852909088135, 0.0408395454287529, 0.004318980034440756, 0.010599052533507347, 0.022129051387310028, -0.04114477336406708, 0.010454068891704082, -0.018756277859210968, 0.0...
36d4a2fd581fb72f386ed2ec3b2efca6917f0407
subsection
2
121
Summary of results
We follow the common practice of reporting convergence rates depending on T, the number of calls to the (stochastic / incremental) gradient oracle. For instance, Sgd converges at the rate \mathcal {O}(\frac{1}{T}) for solving (REF ), ignoring logarithmic terms in the bound . The underlying argument is to view Sgd as st...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 275, "openalex_id": "https://openalex.org/W2112269233", "raw": "A. Rakhlin, O. Shamir, K. Sridharan, et al. Making gradient descent optimal for strongly convex stochastic optimization. In ICML. Citeseer, 2012.", "source_ref_...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05694999173283577, 0.04434530436992645, -0.020478801801800728, -0.04114072397351265, 0.0018273744499310851, 0.01000287290662527, 0.032717250287532806, -0.015687189996242523, 0.0230272077023983, 0.026201268658041954, -0.025773992761969566, -0.018342414870858192, -0.025758732110261917, 0....
a313559c50bc110f62395cd29a9112ad951d1a82
subsection
3
121
Related work
conjecture a tantalizing matrix AM-GM inequality that underlies RandomShuffle's superiority over Sgd. While limited progress on this conjecture has been reported , , the correctness of the full conjecture is still wide open. With the technique of transductive Rademacher complexity, shows that Sgd is not worse than Rand...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.48550/arxiv.1202.4184", "end": 102, "openalex_id": "https://openalex.org/W1600587676", "raw": "B. Recht and C. Ré. Beneath the valley of the noncommutative arithmetic-geometric mean inequality: conjectures, case-studies, and consequences...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06965576857328415, 0.01531541720032692, -0.03601831942796707, -0.025014419108629227, 0.010813127271831036, 0.005566814448684454, 0.016544008627533913, -0.004792267922312021, 0.008104122243821621, 0.02625063993036747, -0.03385111689567566, -0.011522809974849224, -0.01214092131704092, 0.0...
94739a9119da92b1518a0544fb149c471e1f10fe
subsection
4
121
Background and problem setup
For problem (REF ), we assume the finite sum function F(x): \mathbb {R}^d\rightarrow \mathbb {R} is strongly convex, i.e.,F(x) \ge F(y) + \left\langle \nabla F(y), x-y \right\rangle + \tfrac{\mu }{2}\left\Vert x-y\right\Vert ^2,where x, y\in \mathbb {R}^d, and \mu >0 is the strong convexity parameter. Furthermore, we a...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.02690083719789982, -0.0017328000394627452, -0.0003666822158265859, -0.021270429715514183, 0.01730320416390896, -0.0008983475854620337, 0.0046347868628799915, -0.014724508859217167, 0.023040423169732094, 0.03671209141612053, -0.021361980587244034, 0.0318293534219265, 0.02079741470515728, ...
df8f7e435282df9bc2dfbe92309f76206fe5cd91
subsection
5
121
The algorithms under study:
For both Sgd and RandomShuffle, we use \gamma as the step size, which is predetermined before the algorithms are run. The sequences generated by both methods are denoted as (x_k)_{k=0}^T; here x_0 is the initial point and T is the total number of iterations (i.e., number of stochastic gradients used).Sgd is defined as ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06996270269155502, -0.006898591294884682, -0.04416929930448532, -0.03379088640213013, 0.0009510364034213126, -0.0004359314334578812, 0.05943166837096214, 0.006654393393546343, 0.001107475720345974, 0.022954603657126427, 0.00866902619600296, -0.0010941211367025971, -0.03125733137130737, ...
0107230f60b2265dcc984918ebc1d2e2fc391b41
subsection
6
121
Convergence analysis of
The goal of this section is to build theoretical analysis for RandomShuffle. Specifically, we answer the following question: when can we show RandomShuffle to be better than Sgd? We begin by first analyzing quadratic functions in Section REF , where the analysis benefits from having a constant Hessian. Subsequently, in...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0752481073141098, 0.022229621186852455, 0.011450466699898243, -0.03622039780020714, 0.0016573043540120125, -0.007582787424325943, 0.014593432657420635, -0.009993411600589752, 0.02605915628373623, 0.05376608297228813, -0.008032872341573238, -0.0021512536332011223, -0.022824648767709732, ...
c5f3b5476696260fce115c239dc9a769382c0dba
subsection
7
121
Body
We first consider the quadratic instance of (REF ), wheref_i(x) = \tfrac{1}{2}x^T A_i x + b_i^Tx,\qquad i=1,\ldots ,n,where A_i\in \mathbb {R}^{d\times d} is positive semi-definite, and b_i \in \mathbb {R}^d. We should notice often in analyzing strongly convex problems, the quadratic case presents a good example when t...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/s10107-019-01440-w", "end": 1925, "openalex_id": "https://openalex.org/W1866529865", "raw": "M. Gürbüzbalaban, A. Ozdaglar, and P. Parrilo. Why random reshuffling beats stochastic gradient descent. arXiv preprint arXiv:1510.08560, 2...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.046550385653972626, 0.0038900095969438553, -0.018559103831648827, -0.03949914872646332, 0.0203600712120533, 0.001112248981371522, 0.03002118319272995, 0.014018533751368523, 0.018681203946471214, 0.017826508730649948, -0.03705716133117676, -0.00849353801459074, -0.0264497771859169, 0.009...
273b1ef3b8640bed882c4ea530ec6e097c18325e
subsection
8
121
Body
Formally, we recover the main result of  as the following:Corollary 2 As T\rightarrow \infty , RandomShuffle achieves asymptotic convergence rate \mathcal {O}\left(\frac{1}{T^2}\right) when run with the proper step size schedule.Next, we consider the more general case where each component function f_i is convex and the...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/s10107-019-01440-w", "end": 229, "openalex_id": "https://openalex.org/W1866529865", "raw": "M. Gürbüzbalaban, A. Ozdaglar, and P. Parrilo. Why random reshuffling beats stochastic gradient descent. arXiv preprint arXiv:1510.08560, 20...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0549333281815052, 0.032654810696840286, -0.019928589463233948, -0.040528587996959686, 0.00861385092139244, 0.000683328602463007, 0.031922366470098495, -0.019531849771738052, 0.014450516551733017, 0.042909033596515656, -0.023102516308426857, -0.0030938144773244858, -0.007980591617524624, ...
840e95eeb9693e6c906fbd7cfe556e1535e37962
subsection
9
121
Body
We have the following extension of our previous result:Theorem 5 Under the Polyak-Łojasiewicz condition, define condition number \kappa = L/\mu . So long as \frac{T}{\log T} > 16\kappa ^2 n, with step size \eta = \frac{2\log T}{ T\mu }, RandomShuffle achieves convergence rate:\operatorname{\mathrm {\mathbb {E}}}[\left...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 778, "openalex_id": "https://openalex.org/W2557191228", "raw": "O. Shamir. Without-replacement sampling for stochastic gradient methods. In Advances in Neural Information Processing Systems, pages 46–54, 2016.", "source_ref_...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.03790159896016121, 0.03179828077554703, -0.02516092360019684, -0.02888394705951214, 0.005264110863208771, 0.029906252399086952, 0.004798733163625002, 0.02073602005839348, 0.012748303823173046, 0.01698247902095318, -0.0023612207733094692, 0.004363871645182371, -0.024306459352374077, 0.00...
fc69e206ee7ca8dbf4f7dbc062437ea4a429c926
subsection
10
121
Body
Then there isF(\bar{x}) - F(x^*) \le \frac{2D\sqrt{nD\left(\left\Vert \Delta \right\Vert +L_H LD^2 + 2L_HDG\right)}}{\sqrt{T}} + \mathcal {O}\left(\left(\frac{n}{T}\right)^\frac{2}{3}\delta ^\frac{1}{3} + \left(\frac{n}{T}\right)^\frac{3}{4}\right).We have some discussion of this result:Firstly, it is interesting to se...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04843058064579964, 0.040953900665044785, -0.04028252512216568, -0.04086235165596008, -0.006507764104753733, -0.0034770381171256304, -0.00534048629924655, 0.014434571377933025, 0.0005407242570072412, 0.007503383327275515, -0.02174340933561325, 0.014121771790087223, -0.009841753169894218, ...
831c27f63bffc33755f868f002bb391a70c09841
subsection
11
121
Body
We leave this possibility (either improving upon the {1}{\sqrt{T}} dependence on T under existence of noise, or recovering {1}{T} when there is no noise) as an open question.For the completeness of the paper, we include the following analysis of Sgd under Polyak-Łojasiewicz condition.Theorem 8 For finite sum problem sa...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04876530542969704, 0.02920425310730934, -0.040129147469997406, -0.042784079909324646, 0.02256692387163639, 0.011497076600790024, 0.021636171266436577, -0.022811055183410645, 0.0060575176030397415, 0.0232535433024168, -0.018248843029141426, 0.009231229312717915, -0.008392026647925377, -0...
29affe39a368fd0c8d3d1fed328cb0aa0621a2be
subsection
12
121
Understanding the dependence on
Since the motivation of building our convergence rate analysis is to show that RandomShuffle behaves better than Sgd, we would definitely hope that our convergence bounds have a better dependence on T compared to the \mathcal {O}(\frac{1}{T}) bound for Sgd. In an ideal situation, one may hope for a rate of the form \ma...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.07659368962049484, 0.01260286569595337, -0.030805310234427452, -0.023466352373361588, 0.0022161819506436586, 0.03198015317320824, -0.004207312595099211, 0.007434317842125893, 0.017195435240864754, 0.01974347233772278, -0.023435838520526886, -0.011786578223109245, 0.0056834956631064415, ...
d060657c769ee9ea17b022418efa3d8ed51f7217
subsection
13
121
Understanding the dependence on
We directly have the following corollary:Corollary 3 Given the information of \mu , L, G, under the assumption T\ge n and constant step size, there is no step size choice that leads to a convergence rate \mathcal {O}\left(\frac{1}{T^{1+\delta }}\right) for \delta >0.This result indicates that in order to achieve a bett...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1391, "openalex_id": "https://openalex.org/W2107438106", "raw": "R. Johnson and T. Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In Advances in neural information processing systems, pages 31...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04340814799070358, 0.02051355503499508, -0.03797449544072151, -0.014553328976035118, 0.01886514388024807, 0.04249236360192299, 0.002428734675049782, 0.031716641038656235, 0.021337758749723434, -0.0011609465582296252, 0.0010226251324638724, -0.011935713700950146, -0.00021511377417482436, ...
2f543b881f24ca47fbf7309db765b944d974cce1
subsection
14
121
Sparse functions
In the literature on large-scale machine learning, sparsity is a common feature of data. When the data are sparse, each training data point has only a few non-zero features. Under such a setting, each iteration of Sgd only modifies a few dimensions of the decision variables. Some commonly occurring sparse problems incl...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1401, "openalex_id": "", "raw": "B. Recht, C. Re, S. Wright, and F. Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pages 693–701, 2011.", ...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05054118111729622, -0.010559933260083199, -0.04632941633462906, -0.0060734874568879604, 0.005405861884355545, -0.023851409554481506, 0.018113642930984497, 0.00976641196757555, 0.008232779800891876, 0.04294169321656227, -0.031069397926330566, -0.009781671687960625, -0.00609637750312686, ...
de3f6e1fcf16f7e72db7c70796a94b0a0b275932
subsection
15
121
Sparse functions
The key to proving Theorem REF lies in constructing a tighter bound for the error term in the main recursion (see §) by including a discount due to sparsity.We end this section by noting the following simple corollary:Corollary 4 When \rho = \mathcal {O}\left(\frac{1}{n}\right), there is some constant C only dependent ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.03710327669978142, 0.03438765928149223, -0.002259836532175541, -0.008238391950726509, -0.0029025075491517782, -0.000016328987840097398, 0.0005816456978209317, 0.024257486686110497, -0.011518492363393307, 0.00560286920517683, -0.02312852256000042, -0.02622554637491703, 0.012693225406110287...
f0edebac84cae55f6febf6ddf7dff8cd0b54a239
subsection
16
121
Proof sketch of Theorem
In this section we provide a proof sketch for Theorem REF . The central idea is to establish an inequality\operatorname{\mathrm {\mathbb {E}}}\bigl [\left\Vert x^{t+1}_0 - x^*\right\Vert ^2\bigr ] \le (1-n\gamma \alpha _1)\left\Vert x^t_0 - x^*\right\Vert ^2 + n\gamma ^3 \alpha _2 + n^4\gamma ^4 \alpha _3,where x^t_0 a...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06039887294173241, 0.00423463573679328, -0.023668179288506508, -0.044498007744550705, 0.004410124849528074, 0.02429383620619774, 0.030214697122573853, 0.028597142547369003, 0.02136392705142498, 0.0032808887772262096, -0.02453799545764923, -0.003998106345534325, -0.0350978784263134, 0.02...
39a897db87b952e3f78d73780635a4cca9e2d956
subsection
17
121
Proof sketch of Theorem
Thus, the terms A^t_2 and A^t_4 are also random variables that depend on \sigma _t(\cdot ), and require taking expectations. The main body of our analysis involves bounding each of these terms separately.The term A_1^t can be easily bounded by exploiting the strong convexity of F, using a standard inequality (Theorem 2...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 483, "openalex_id": "https://openalex.org/W3141595720", "raw": "Y. Nesterov. Introductory lectures on convex optimization: A basic course, volume 87. Springer Science & Business Media, 2013.", "source_ref_id": "ac12aab6cc5bb...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.008108648471534252, 0.03077167272567749, -0.01646139658987522, -0.0032800743356347084, 0.003753015073016286, -0.023219874128699303, -0.006117719691246748, 0.0009458818822167814, 0.015683332458138466, 0.0013177507789805532, -0.05031481385231018, 0.034051746129989624, -0.008909597061574459,...
282607955d6d37045d837f999ccefcb1197f6d2f
subsection
18
121
Proof sketch of Theorem
Specifically, by introducing second-order information and somewhat involved analysis, we obtain the following bound for A_2^t:Lemma 1 Over the randomness of the permutation, we have the inequality:-2\gamma \langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}[R^t] \rangle &\le \frac{1}{2}\gamma \mu (n-1)\left\Vert...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05185435339808464, 0.0074012246914207935, -0.027560023590922356, 0.00011665750935208052, -0.007649203762412071, 0.02873506397008896, 0.013955505564808846, 0.032595910131931305, 0.028124654665589333, 0.02748372219502926, -0.050755616277456284, 0.010422755964100361, 0.010895823128521442, ...
792a9279eb55df24270e5ef9b91beea02e46a52a
subsection
19
121
Discussion of results
We discuss below our results in more detail, including their implications, strengths, and limitations.
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.01777699962258339, 0.020798327401280403, -0.049409378319978714, 0.014557302929461002, -0.022965136915445328, -0.03884999454021454, -0.0058137658052146435, 0.01593063399195671, 0.01609848439693451, 0.031525563448667526, -0.0018129869131371379, 0.045930273830890656, -0.03955191746354103, 0...
2fe0928d3bb4061ccab747e47f3729abd6c5911c
subsection
20
121
Comparison with
It is well-known that under strong convexity Sgd converges with a rate of \mathcal {O}\bigl (\frac{1}{T}\bigr ) . A direct comparison indicates the following fact: RandomShuffle is provably better than Sgd after \mathcal {O}(\sqrt{n}) epochs. This is an acceptable amount of epochs for even some of the largest data sets...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 113, "openalex_id": "https://openalex.org/W2112269233", "raw": "A. Rakhlin, O. Shamir, K. Sridharan, et al. Making gradient descent optimal for strongly convex stochastic optimization. In ICML. Citeseer, 2012.", "source_ref_...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06859561800956726, 0.015623200684785843, -0.025402957573533058, -0.05724438279867172, -0.012228511273860931, -0.02164972946047783, 0.021802298724651337, 0.004531338345259428, 0.008971135132014751, 0.03338238596916199, -0.01557742990553379, -0.014333982020616531, -0.017835469916462898, 0...
e43533b0c1e2ca66339f995867fbf05a4352eabe
subsection
21
121
Deterministic variant.
When the algorithm is run in a deterministic fashion, i.e., the functions f_i are visited in a fixed order, better convergence rate than Sgd can also be achieved as T becomes large. For instance, a result in  translates into a \mathcal {O}\bigl (\frac{n^2}{T^2}\bigr ) bound for the deterministic case. This directly imp...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 302, "openalex_id": "", "raw": "M. Gürbüzbalaban, A. Ozdaglar, and P. Parrilo. Convergence rate of incremental gradient and newton methods. arXiv preprint arXiv:1510.08562, 2015a.", "source_ref_id": "77a1eaebdd70c8b40c822000...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05914851650595665, 0.018358014523983, -0.01132306456565857, -0.011422256007790565, -0.022493528202176094, 0.004120252560824156, -0.009583402425050735, -0.011063641868531704, -0.009781785309314728, 0.027117367833852768, -0.03833360970020294, 0.0008583859889768064, -0.000028374424800858833,...
bfbd89a532b2212bb863915e3b7cc9cd948ae454
subsection
22
121
Epochs required.
It is also a limitation that our bound only holds after a certain number of epochs. Moreover, this number of epochs is dependent on \kappa (e.g., \mathcal {O}(\kappa ) epochs for the quadratic case). This limits the interest of our result to cases when the problem is not too ill-conditioned. Otherwise, such a number of...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.0026639350689947605, 0.03404955193400383, -0.020930100232362747, 0.02983912266790867, -0.04109744355082512, -0.06382765620946884, 0.005972401238977909, 0.01932830736041069, -0.024484556168317795, -0.01181513350456953, -0.049762383103370667, 0.03240199387073517, 0.010472677648067474, 0.00...
e5081ba69926cbe205fc042592736506b2cbe8aa
subsection
23
121
Dependence on
It should be noticed that \kappa can be large sometimes. Therefore, it may be informative to view our result in a \kappa -dependent form. In particular, we still assume D, L, L_H are constant, but no longer \mu . We use the bound G\le \max _i{\left\Vert \nabla f_i(x^*)\right\Vert + DL} and assume \max _i{\left\Vert \na...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1410, "openalex_id": "https://openalex.org/W2112269233", "raw": "A. Rakhlin, O. Shamir, K. Sridharan, et al. Making gradient descent optimal for strongly convex stochastic optimization. In ICML. Citeseer, 2012.", "source_ref...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.025069909170269966, 0.02984585613012314, -0.03518637269735336, -0.020263444632291794, 0.017837325111031532, -0.03817706182599068, 0.022994738072156906, 0.02380344457924366, 0.015624824911355972, 0.004730171523988247, -0.020965341478586197, 0.029800081625580788, -0.004707283806055784, 0....
644b5a2bfa0612a7a1efbf3d4640946eacce5ed9
subsection
24
121
Sparse data setting.
Notably, in the sparse setting (with sparsity factor \rho = \mathcal {O}\bigl (\frac{1}{n}\bigr )), the proven convergence rate is strictly better than the \mathcal {O}\bigl (\frac{1}{T}\bigr ) rate of Sgd. This result follows the following intuition: when each dimension is only touched by several functions, letting th...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.054721176624298096, 0.026307670399546623, -0.028657659888267517, -0.008675121702253819, -0.013962597586214542, -0.0022946728859096766, 0.02272164821624756, 0.00661125173792243, -0.0029165043961256742, 0.024552809074521065, -0.0251174159348011, -0.014969735406339169, 0.015572492964565754, ...
963838aa88fe1743829b9a2900bcde4bd19dbd1a
subsection
25
121
Extensions
In this section, we provide some further extensions before concluding with some open problems.
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.033669985830783844, 0.016102373600006104, -0.014232666231691837, 0.020391251891851425, -0.036783620715141296, -0.016712889075279236, 0.02858743630349636, 0.01542317308485508, -0.016728153452277184, 0.016361841931939125, 0.01004299707710743, 0.024802234023809433, -0.021123871207237244, 0...
fa52c60d38db1187398e19192eedad7dbab8ca7e
subsection
26
121
Vanishing variance
Our previous results show that RandomShuffle converges faster than Sgd after a certain number of epochs. However, one may want to see whether it is possible to show faster convergence of RandomShuffle after only one epoch, or even within one epoch. In this section, we study a specialized class of strongly convex proble...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 619, "openalex_id": "https://openalex.org/W2156779765", "raw": "E. Moulines and F. R. Bach. Non-asymptotic analysis of stochastic approximation algorithms for machine learning. In Advances in Neural Information Processing Systems,...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05158405750989914, 0.02927166223526001, -0.009462164714932442, -0.04050416871905327, -0.026433013379573822, 0.023411225527524948, 0.013056260533630848, -0.008271763101220131, 0.016879281029105186, 0.027486061677336693, -0.016009371727705002, -0.007394223473966122, -0.015055524185299873, ...
8144af9e05982d21228d5e0dc690c1c598d27097
subsection
27
121
Vanishing variance
For step size \eta \le \min \limits _i\lbrace \frac{2}{L_i+\mu _i}\rbrace and any T\ge 1, there is\max \limits _{P\in \mathcal {P}, x_0 \in D_R(x^*)} \operatorname{\mathrm {\mathbb {E}}}[\left\Vert X_{RS}(T, x_0, \gamma , P) - x^*\right\Vert ^2] \le \max \limits _{P\in \mathcal {P}, x_0 \in D_R(x^*)}\operatorname{\math...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06639712303876877, 0.023922493681311607, -0.022167973220348358, -0.021084748208522797, -0.034449610859155655, 0.051170945167541504, -0.009092988446354866, 0.034266531467437744, 0.01919291913509369, 0.020367683842778206, 0.004439697600901127, 0.022167973220348358, -0.012464717030525208, ...
567b390a1374d700ff4255b517eb0e6f3defee59
subsection
28
121
Conclusion and open problems
A long-standing problem in the theory of stochastic gradient descent (Sgd) is to prove that RandomShuffle converges faster than the usual with-replacement Sgd. In this paper, we provide the first non-asymptotic convergence rate analysis for RandomShuffle. We show in particular that after \mathcal {O}(\sqrt{n}) epochs, ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.059987738728523254, 0.029765024781227112, -0.011098952032625675, -0.03286205232143402, -0.006018607411533594, 0.010908248834311962, 0.021206246688961983, -0.006506808567792177, 0.029780281707644463, 0.028361447155475616, -0.022609824314713478, -0.014447708614170551, -0.02602723427116871, ...
ddd66e76622526a2f2e271f7f52257c5aa2852a4
subsection
29
121
Conclusion and open problems
For one epoch of RandomShuffle, We have the following inequality \left\Vert x^{t}_n - x^*\right\Vert ^2 &= \left\Vert x^t_0 - x^*\right\Vert ^2 - 2\gamma \left\langle x^t_0 - x^*, \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) \right\rangle + \gamma ^2 \left\Vert \sum _{i=1}^n \nabla f_{\sigma...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1359, "openalex_id": "https://openalex.org/W3141595720", "raw": "Y. Nesterov. Introductory lectures on convex optimization: A basic course, volume 87. Springer Science & Business Media, 2013.", "source_ref_id": "ac12aab6cc5b...
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06925797462463379, 0.026284469291567802, -0.012120146304368973, -0.03957162797451019, -0.029060890898108482, 0.032340727746486664, 0.003310348140075803, 0.06132534518837929, -0.00783347338438034, -0.0033198825549334288, -0.025399677455425262, -0.009290331974625587, -0.047931402921676636, ...
a845bb4cc1630854953548a4421d2b498a082fb2
subsection
30
121
Conclusion and open problems
Take the expectation of (REF ) over randomness of permutation \sigma _t\left(\cdot \right), we have \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x^t_n - x^*\right\Vert ^2\right] &\le \left(1-2n\gamma \frac{L\mu }{L+\mu }\right)\left\Vert x^t_0 - x^*\right\Vert ^2 -\left(2n\gamma \frac{1}{L+\mu }-2n^2\gamma ^2\r...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05427094176411629, 0.01895514875650406, -0.0018495387630537152, -0.0015128253726288676, -0.011904199607670307, 0.006238258443772793, 0.021732795983552933, 0.04419815540313721, 0.007413416635245085, -0.01162948738783598, -0.03617045283317566, 0.027959607541561127, -0.023289497941732407, ...
681301327b1b30d6ac938950be0beb168b8ef2da
subsection
31
121
Conclusion and open problems
Firstly, we give a bound on the norm of R^t: \left\Vert R^t\right\Vert &= \left\Vert \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)}\left(x^t_0\right)\right\Vert \\ &\le \sum _{i=1}^n \left\Vert \nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.034629661589860916, 0.03188369795680046, -0.025461191311478615, 0.002078542485833168, 0.016430020332336426, -0.00032822854700498283, 0.006498782429844141, 0.026773152872920036, -0.006967884488403797, 0.0263765137642622, -0.049579910933971405, 0.01777249202132225, -0.0037928633391857147, ...
64a1e98aa57032148613b5ff55c6f4a2d2240d48
subsection
32
121
Conclusion and open problems
We begin with the following decomposition: R^t &= \sum _{i=1}^n \left[\nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \nabla f_{\sigma _t\left(i\right)}\left(x^t_{0}\right)\right] \\ &= \sum _{i=1}^n \left[H_{\sigma _t\left(i\right)}\left(x^t_{i-1} - x^t_0\right)\right] \\ &= \sum _{i=1}^n \left\lbrace H_{...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.012390484102070332, 0.022675197571516037, -0.011604635044932365, 0.007904274389147758, -0.010696711018681526, 0.007591460831463337, 0.013634110800921917, 0.03357027843594551, 0.005447540432214737, 0.010475452989339828, -0.03796493262052536, -0.008850346319377422, -0.03111354447901249, 0....
c7842c759f7634017820718f6af9bba059e62788
subsection
33
121
Conclusion and open problems
Here we define random variables A^t = -\gamma \sum _{i=1}^n \left[H_{\sigma _t\left(i\right)}\sum _{j=1}^{i-1} \nabla f_{\sigma _t\left(j\right)} \left(x^t_0\right)\right], B^t = - \gamma \sum _{i=1}^n \left\lbrace H_{\sigma _t\left(i\right)}\sum _{j=1}^{i-1}\left[\nabla f_{\sigma _t\left(j\right)} \left(x^t_{j-1}\righ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05455244332551956, -0.0042776502668857574, -0.019201727584004402, -0.006784712430089712, 0.0009282234241254628, -0.01521790400147438, -0.0022628428414463997, 0.034404367208480835, -0.004827143624424934, 0.028069933876395226, -0.05061441287398338, 0.03883083909749985, -0.000046625817049061...
e10ade45b7ea5b57cde1c77149459a4d63e81eae
subsection
34
121
Conclusion and open problems
Using (REF ) and  (REF ), we can decompose the inner product of x^t_0 - x^* and \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] into: -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] \right\rangle &= -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\lef...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.017502684146165848, 0.015023009851574898, -0.01939486525952816, 0.014595742337405682, 0.002018073108047247, -0.010544336400926113, 0.008842898532748222, 0.03695858642458916, 0.005981736816465855, 0.03128204122185707, -0.043062400072813034, 0.023759091272950172, -0.012344961054623127, -0...
c07e17930e223b70e248f1d558b8734d1106b470
subsection
35
121
Conclusion and open problems
For the first term in (REF ), there is &\ \ \ \ \gamma ^2 n\left(n-1\right) \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i \nabla f_j\left(x^t_0\right)\right\rangle \\ &= \gamma ^2 n\left(n-1\right) \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.056741006672382355, 0.010392853058874607, -0.0218082033097744, -0.015932654961943626, -0.011751096695661545, -0.007199454121291637, 0.010644662193953991, 0.05432974547147751, 0.018633881583809853, 0.018252352252602577, -0.0576261542737484, 0.0015881148865446448, -0.033544037491083145, -...
0fe6e7b7beab2ef9a78c747359b63e11e8dddc97
subsection
36
121
Conclusion and open problems
For the second term in (REF ), we use the bound -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\left[B^t\right]\right\rangle &\le 2\gamma \left[\frac{\lambda _2}{2}\left\Vert x^t_0- x^*\right\Vert ^2 + \frac{1}{2\lambda _2}\left\Vert \operatorname{\mathrm {\mathbb {E}}}\left[B^t\right]\right\Ve...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05859992653131485, 0.02411142736673355, -0.004452221095561981, -0.02321106381714344, -0.025835853070020676, 0.006657348480075598, 0.0161149799823761, 0.057073887437582016, 0.027300851419568062, 0.005100787617266178, -0.028338557109236717, 0.02829277701675892, -0.019121278077363968, 0.00...
1293ebc949d0136b56d8c69bd3e72e56c2ef50d3
subsection
37
121
Conclusion and open problems
Toward this end, we use the following important fact: \left\Vert \Delta \right\Vert &= \left\Vert \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i \nabla f_j\left(x^*\right)\right\Vert \\ &= \left\Vert \frac{1}{n\left(n-1\right)} \sum _{i\ne j} H_i \nabla f_j\left(x^*\right)\right\Vert \\ &= \left\Vert \frac{...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0626225695014, 0.028778305277228355, -0.015327575616538525, 0.016281258314847946, -0.032226819545030594, -0.016998426988720894, -0.011436553671956062, 0.02461262419819832, 0.006988581269979477, 0.007572234608232975, -0.06347706913948059, 0.011390777304768562, -0.012718302197754383, 0.00...
13438f69120773b11054785bafc2a11f1803da99
subsection
38
121
Conclusion and open problems
Substituting (REF ) (REF ) back to (REF ) and using (REF ) , we finally get a recursion bound for one epoch: &\ \ \ \ \operatorname{\mathrm {\mathbb {E}}}\left\Vert x^t_n - x^*\right\Vert ^2 \\ &\le \left(1-2n\gamma \frac{L\mu }{L+\mu } +\frac{1}{2}\gamma \mu \left(n-1\right)\right)\left\Vert x^t_0 - x^*\right\Vert ^2...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0334845595061779, 0.02286229282617569, -0.01831425353884697, -0.01781061291694641, -0.01639125682413578, -0.020511964336037636, -0.0050135268829762936, 0.03699479252099991, 0.011675337329506874, 0.006326048634946346, -0.03278251364827156, 0.03614012897014618, -0.026143597438931465, 0.03...
7759eb4ac56bd7a9ffb8ec67a3055599b3b26e8b
subsection
39
121
Conclusion and open problems
Expanding (REF ) over all epochs leads to a final bound of RandomShuffle: \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - x^*\right\Vert ^2\right] \le \left(1-n\gamma \frac{L\mu }{L+\mu }\right)^{\frac{T}{n}}\left\Vert x_0 - x^*\right\Vert ^2 + \frac{T}{n} \left(\gamma ^3nC_1 + \gamma ^{5}n^{5}C_2 + \gamma ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.056337323039770126, 0.02763458527624607, -0.014725221320986748, -0.038941726088523865, -0.011085879057645798, 0.019852343946695328, -0.0037633241154253483, 0.02031012438237667, 0.020218567922711372, -0.00421156594529748, -0.005420865025371313, 0.0006513812113553286, -0.032532818615436554,...
3650462780c1746f740f6dac7eef330ff9a9f4b4
subsection
40
121
Conclusion and open problems
Or in the expanding version with constant dependence, we have \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - x^*\right\Vert ^2\right] \le \frac{\left(\log T\right)^2}{T^2}\left(D^2 + 128\frac{L^2G^2}{\mu ^4}\right) + \frac{n^3\left(\log T\right)^4}{T^3}128\frac{L^2G^2}{\mu ^4} + \frac{n^4\left(\log T\right...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05399071425199509, 0.008973019197583199, -0.011345985345542431, 0.03222961723804474, 0.004066852852702141, 0.01638949289917946, -0.020570797845721245, 0.016740478575229645, 0.011376505717635155, 0.018220722675323486, -0.021013345569372177, -0.0034678885713219643, 0.01837332360446453, 0....
e57dc4c1915810e81b3a91814c51cfbacd1a04f4
subsection
41
121
Conclusion and open problems
Again, define error term R^t = \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)}\left(x^t_0\right).
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.0006196077447384596, 0.01813318580389023, -0.018484249711036682, 0.02347545512020588, 0.00022036857262719423, 0.012050631456077099, 0.006712178699672222, 0.030756203457713127, -0.016789987683296204, -0.015301783569157124, -0.026024479418992996, -0.01008925586938858, -0.027779797092080116, ...
cd868780aa7ca4193b7900c874cb402f2cae38fb
subsection
42
121
Conclusion and open problems
We have the following decomposition for the error term: R^t &= \sum _{i=1}^n \left[\nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \nabla f_{\sigma _t\left(i\right)}\left(x^t_{0}\right)\right]\\ &= \sum _{i=1}^n \left[\int _{x^t_0}^{x^t_{i-1}} H_{\sigma _t\left(i\right)}\left(x\right) dx\right]\\ &= \sum _...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.014037642627954483, 0.005763825494796038, -0.018081093207001686, 0.01695197820663452, 0.0019339903956279159, -0.006603032350540161, 0.0014628901844844222, 0.022475486621260643, -0.0050504994578659534, 0.01001707836985588, -0.03207296133041382, -0.005878262687474489, -0.010902060195803642, ...
d719d72593d892e3e2767616793c55f295f0f994
subsection
43
121
Conclusion and open problems
Here we define random variables A^t = -\gamma \sum _{i=1}^n \left[H_{\sigma _t\left(i\right)}\sum _{j=1}^{i-1} \nabla f_{\sigma _t\left(j\right)} \left(x^t_0\right)\right], B^t = - \gamma \sum _{i=1}^n \left\lbrace H_{\sigma _t\left(i\right)}\sum _{j=1}^{i-1}\left[\nabla f_{\sigma _t\left(j\right)} \left(x^t_{j-1}\righ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.03512922301888466, -0.01084244716912508, -0.02678183652460575, 0.00470017408952117, 0.007366912439465523, -0.0340915210545063, -0.010575391352176666, 0.030780036002397537, 0.005577641539275646, 0.02484377659857273, -0.044468529522418976, 0.0207540150731802, 0.005619607400149107, -0.0093...
36983d35f55e541c1ba437e7cd12512c8e8d997b
subsection
44
121
Conclusion and open problems
\left\Vert C^t\right\Vert &\le \sum _{i=1}^n \left[\int _{0}^{\left\Vert x^t_{i-1}-x^t_0\right\Vert } \left\Vert H_{\sigma _t\left(i\right)}\left(x^t_0+t\frac{x^t_{i-1}-x^t_0}{\left\Vert x^t_{i-1}-x^t_0\right\Vert }\right)-H_{\sigma _t\left(i\right)}\right\Vert dt\right]\\ &\le \sum _{i=1}^n \left[L_H\max \left\lbrace ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.027158524841070175, 0.016600266098976135, -0.030484680086374283, -0.01763778366148472, -0.009185073897242546, -0.0038105850107967854, -0.021208060905337334, 0.03844914585351944, 0.003619865048676729, -0.00912404339760542, -0.03262074291706085, 0.005713970400393009, 0.0030705914832651615, ...
6cf7de84087a875374cf40919e23478366600544
subsection
45
121
Conclusion and open problems
Using (REF ) (REF ), we can decompose the innerproduct of x^t_0 - x^* and \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] as following: -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] \right\rangle &= -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\l...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.02541893720626831, 0.014563250355422497, -0.01617291197180748, 0.007563888095319271, -0.0036160191521048546, -0.00850603636354208, 0.021329935640096664, 0.037258729338645935, 0.00494341878220439, 0.04287347570061684, -0.03844881057739258, 0.018049580976366997, -0.011092176660895348, -0....
7ecd9577b06285f3713f0af600f80f06458e3577
subsection
46
121
Conclusion and open problems
For the first term in the (REF ), we have further bound: &\ \ \ \ \gamma ^2 n\left(n-1\right) \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i \nabla f_j\left(x^t_0\right)\right\rangle \\ &= \gamma ^2 n\left(n-1\right)\operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} \left\langle...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04946403577923775, 0.021222848445177078, 0.0065110353752970695, -0.007651515770703554, -0.029431253671646118, -0.008086347952485085, -0.009345071390271187, 0.04568023607134819, 0.01280084066092968, 0.01960557885468006, -0.0411335714161396, 0.01904105953872204, -0.03469500690698624, -0.0...
78f923f15147b456fe0726c44ad4cb2160fe0c78
subsection
47
121
Conclusion and open problems
Note that here H_i\left(x^t_0 - x^*\right) is the matrix H_i(x^*) times vector x^t_0 - x^*, not the Hessian at point x^t_0 - x^*. The last inequality is because of \left\Vert H_i\left(x^t_0 - x^*\right) - \left(\nabla f_i\left(x^t_0\right) - \nabla f_i\left(x^*\right)\right)\right\Vert &=\left\Vert H_i\left(x^t_0 - x^...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.023717762902379036, -0.005696688778698444, -0.014735842123627663, -0.006448362022638321, -0.01610182598233223, 0.026816029101610184, -0.019016943871974945, 0.051831092685461044, 0.023931436240673065, 0.015842366963624954, -0.03907173126935959, 0.007524360902607441, 0.011958086863160133, ...
84d651b79cfc5beabbb367476dc2c2c1e06563d1
subsection
48
121
Conclusion and open problems
Substituting (REF ) (REF ) (REF ) back to (REF ), we get -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] \right\rangle &\le \gamma ^2n^2 \left\Vert \nabla F\left(x^t_0\right)\right\Vert ^2 + \frac{1}{2}\gamma \mu \left(n-1\right)\left\Vert x^t_0-x^*\right\Vert ^2 + \gamma ^{3} \...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04845882207155228, 0.01401430368423462, -0.015395136550068855, -0.03902949392795563, -0.023436006158590317, -0.004588788375258446, 0.009398814290761948, 0.05172399431467056, 0.006312921177595854, -0.001487637055106461, -0.03368925675749779, 0.018767112866044044, -0.037442680448293686, 0...
8336cb623d5fe2400be917140e7cdb91a6c73951
subsection
49
121
Conclusion and open problems
Now assume \frac{3}{2}n\gamma \frac{L\mu }{L+\mu }>\frac{1}{2}\gamma \mu \left(n-1\right)+\gamma ^2n^2\left(L_H LD + 3L_H G\right), and 2n\gamma \frac{1}{L+\mu }-3\gamma ^2n^2>0, which we call assumption 1 and assumption 2, (REF ) can be further turned into: \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x^t_n -...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.07263966649770737, 0.03683868795633316, -0.0056654359214007854, -0.020632106810808182, -0.014459260739386082, 0.012887435965240002, -0.012628008611500263, 0.0381816066801548, 0.001049154787324369, -0.01561905350536108, -0.021517211571335793, 0.028811698779463768, -0.02653789333999157, 0...
189fd83691a0838f21df2ee7dcfe0561bedb1677
subsection
50
121
Conclusion and open problems
Let \gamma = \frac{8\log T}{T\mu }, there is \operatorname{\mathrm {\mathbb {E}}}\left\Vert x_T - x^*\right\Vert ^2 &\le \left(1-\frac{2n\log T}{T}\right)^{\frac{T}{2n\log T} 2\log T}\left\Vert x_0 - x^*\right\Vert ^2 + \frac{T}{n} \left(\gamma ^3nC_1 + \gamma ^{4}n^{4}C_2 + \gamma ^5n^5C_3\right)\\ &\le \frac{1}{T^2}...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.040472161024808884, 0.016802355647087097, -0.03482559323310852, -0.005337533075362444, -0.00569235160946846, -0.02395976334810257, -0.008996357209980488, 0.02229631505906582, -0.020236080512404442, 0.008836116641759872, -0.016649747267365456, -0.0016081273788586259, -0.03433724120259285, ...
c94dc3e3454de0b9450a6161dbb9b35fef2ccdc3
subsection
51
121
Conclusion and open problems
Assumption 3 is equivalent to \frac{T}{\log T} > \frac{8L}{L+\mu }n, which is satisfied when \frac{T}{\log T} > 8n. Since 12\left(1+\frac{L}{\mu }\right)>8, we only need \frac{T}{\log T} > \max \left\lbrace \frac{32}{\mu ^2}\left(L_H LD + 3L_H G\right)n, 12\left(1+\frac{L}{\mu }\right) n \right\rbrace . So whenever \fr...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.027503108605742455, 0.0034607823472470045, -0.025015313178300858, -0.0021424798760563135, 0.0043727196753025055, 0.019795522093772888, -0.011965836398303509, 0.0045215291902422905, -0.021016526967287064, 0.014148380607366562, 0.008195986971259117, 0.023061707615852356, -0.0213675647974014...
8c0410d23535946cc38c87ff268ac6bc8c28bcd5
subsection
52
121
Conclusion and open problems
With (REF ) (REF ), we have close-formed expression on the final error: \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - x^*\right\Vert ^2\right] &= \left\Vert \operatorname{\mathrm {\mathbb {E}}}\left[x_T\right] - x^*\right\Vert ^2 + \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - \operatorname{\...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.008079609833657742, 0.019943438470363617, -0.025406138971447945, -0.02381921000778675, -0.0003204378008376807, 0.010543929412961006, 0.039581697434186935, 0.032532066106796265, 0.01306165475398302, 0.011169546283781528, -0.017410453408956528, 0.029434500262141228, 0.015197906643152237, ...
a43ffdb0943a27fa0f53477f5eab175a4eee0ee5
subsection
53
121
Conclusion and open problems
Since \left\langle e_i, e_j\right\rangle = 0 for i\ne j, we can simplify the last term in (REF ): \left\Vert \sum _{t=1}^T (-1)^{\sigma (t)} \gamma (I-\gamma A)^{T-t} A b\right\Vert ^2 &= \left\Vert \sum _{t=1}^T (-1)^{\sigma (t)} \gamma (I-\gamma A)^{T-t} A (\sum _{i=1}^d b_ie_i)\right\Vert ^2\\ &= \left\Vert \sum _{...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.0420553982257843, 0.03497495502233505, -0.015747884288430214, -0.02301144413650036, 0.000318543694447726, -0.02293514646589756, 0.016251450404524803, 0.04828130826354027, 0.0008578744600526989, 0.01683131605386734, -0.011292087845504284, 0.03698921948671341, -0.009819538332521915, 0.000...
5da8850efbf01ac6786af715ad6c4bb074f3f12a
subsection
54
121
Conclusion and open problems
Then (REF ) can simplified as \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - x^*\right\Vert ^2\right] &= \sum _{i=1}^d (1-\gamma \lambda _i)^{2T} a_i^2 + \gamma ^2 \sum _{i=1}^d b_i^2 \lambda _i^2 \operatorname{\mathrm {\mathbb {E}}}\left[\left[ \sum _{t=1}^T (-1)^{\sigma (t)} (1-\gamma \lambda _i)^{T-t}\r...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05561777204275131, 0.02930382266640663, -0.028891952708363533, -0.01845790073275566, 0.020212164148688316, 0.005243721418082714, 0.008389954455196857, 0.03517679125070572, 0.02048674412071705, 0.033864907920360565, -0.015727350488305092, 0.05314654856920242, -0.008473854511976242, 0.011...
6550ccf93b0e39ed522327ee832ea73d967fce6d
subsection
55
121
Conclusion and open problems
For contradiction, we assume for any T, there is a \gamma dependent on T such that \operatorname{\mathrm {\mathbb {E}}}\left[\left\Vert x_T - x^*\right\Vert ^2\right] \le o({1}{T}). Now we determine the specific requirement of A and b. The only requirement is: A has at least three different positive eigenvalues \lambda...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.03445831686258316, 0.015260547399520874, -0.032260797917842865, -0.019655585289001465, 0.021929405629634857, -0.0110486363992095, 0.018358439207077026, 0.01931985281407833, 0.014192309230566025, 0.012216067872941494, -0.040043674409389496, -0.009873573668301105, -0.009446279145777225, 0...
8682fdcf12247823e2cc1b0426036b2a2bb56201
subsection
56
121
Conclusion and open problems
Since (REF ), there is (1-\gamma \lambda _i)^{T} = o(1) , so \gamma ^2\left[- \frac{1}{T-1} \frac{-2(1-\gamma \lambda _i)^T+(1-\gamma \lambda _i)^{2T}}{\gamma ^2\lambda _i^2}\right] = o(\frac{1}{T}). Again, since |1-\gamma \lambda _1|<1, for i=2,3 there is |\frac{\gamma ^2}{2\gamma \lambda _i - \gamma ^2\lambda _i^2}|\...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06084893271327019, 0.03726005181670189, -0.024290747940540314, -0.02636583521962166, 0.00568742211908102, -0.045560408383607864, 0.0065723867155611515, 0.0424172580242157, -0.006862288806587458, 0.004764312878251076, -0.04522473365068436, 0.003732489887624979, -0.0260454174131155, 0.009...
32d1defbce188f9d782aa311ff83346d06b83cb4
subsection
57
121
Conclusion and open problems
As a result, no step size can leads to convergence of o(\frac{1}{T}). Proof of Theorem REF The idea is similar to the proof of theorem REF , with a slightly different analysis on the R^t term adopting the sparsity parameter. For any i, we use H_i to denote H_i(x^*). Again, we have the following decomposition for the e...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.006781651172786951, 0.029262559488415718, -0.04516015201807022, 0.0002382684324402362, -0.0035815357696264982, 0.010771307162940502, -0.013830296695232391, 0.019376147538423538, 0.021878262981772423, 0.014623651280999184, -0.028835367411375046, -0.00887183379381895, 0.0012424762826412916,...
3e3bf0e436a80349282d81f8495d739b8fe7aef9
subsection
58
121
Conclusion and open problems
Here we define random variables A^t = -\gamma \sum _{i=1}^n \left[H_{\sigma _t(i)}\sum _{j=1}^{i-1} \nabla f_{\sigma _t(j)} (x^t_0)\right], B^t = - \gamma \sum _{i=1}^n \left\lbrace H_{\sigma _t(i)}\sum _{j=1}^{i-1}\left[\nabla f_{\sigma _t(j)} (x^t_{j-1})-\nabla f_{\sigma _t(j)} (x^t_0)\right]\right\rbrace , C^t = \su...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.032189708203077316, -0.009878121316432953, -0.02868087776005268, 0.001072672544978559, -0.008344914764165878, -0.020747870206832886, 0.00836779922246933, 0.019649453461170197, -0.007006220053881407, 0.033044032752513885, -0.04665219411253929, 0.030221713706851006, 0.013394580222666264, ...
c90f0a402ca7fb800f828686810fd636866d8f4c
subsection
59
121
Conclusion and open problems
Here the introduction of \rho in (REF ) is because: if f_{\sigma _t(k)} and f_{\sigma _t(j)} depend on disjoint dimensions of variables and k<j, then there must be \nabla f_{\sigma _t(j)} (x^t_k) = \nabla f_{\sigma _t(j)} (x^t_{k-1}). The introduction of \rho in (REF ) is similar: if f_{\sigma _t(i)} and f_{\sigma _t(j...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.018778612837195396, 0.0007732257945463061, -0.025902587920427322, 0.02504832111299038, 0.002396904630586505, 0.011830068193376064, 0.04048614203929901, 0.020151542499661446, -0.0021165984217077494, 0.025841567665338516, -0.01604801043868065, 0.00964101031422615, -0.03450627252459526, 0....
2e3598ea9b81db1bf578d8b68cce00b325bff12c
subsection
60
121
Conclusion and open problems
For the first term in the (REF ), there is &\ \ \ \ \gamma ^2 n(n-1) \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i \nabla f_j(x^t_0)\right\rangle \\ &= \gamma ^2 n(n-1)\operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} \left\langle H_i (x^t_0 - x^*), \nabla f_j(x^t_0) - \nabla ...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.055429618805646896, 0.018756333738565445, -0.004586068447679281, -0.009210290387272835, -0.02946987748146057, -0.006470858585089445, -0.0005880430690012872, 0.04584542289376259, -0.0013573160395026207, 0.017657507210969925, -0.0346435122191906, 0.016375545412302017, -0.0311638992279768, ...
9c4f3340de37aad517f238fcb5ff38bda0c864a0
subsection
61
121
Conclusion and open problems
Where the last inequality is because of \left\Vert H_i(x^t_0 - x^*) - (\nabla f_i(x^t_0) - \nabla f_i(x^*))\right\Vert &=\left\Vert H_i(x^t_0 - x^*) - \int _{x^*}^{x^t_0} H_i(x) dx\right\Vert \\ &=\left\Vert \int _{x^*}^{x^t_0} (H_i - H_i(x))dx\right\Vert \\ &\le \int _{0}^{\left\Vert x^t_0-x^*\right\Vert } \left\Vert...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.018445106223225594, 0.003827473847195506, 0.000636482029221952, -0.023311927914619446, -0.023800136521458626, 0.029582347720861435, -0.009367489255964756, 0.04177229106426239, 0.012876483611762524, -0.00045388081343844533, -0.040063563734292984, 0.012456930242478848, -0.007204880937933922...
ec05e0081d070d7fae27f0a298d7c2965d1e91ba
subsection
62
121
Conclusion and open problems
Substituting (REF ) (REF ) (REF ) back to (REF ), we get -2\gamma \left\langle x^t_0 - x^*, \operatorname{\mathrm {\mathbb {E}}}\left[R^t\right] \right\rangle &\le \gamma ^2n^2 \left\Vert \nabla F(x^t_0)\right\Vert ^2 + \frac{1}{2}\gamma \mu (n-1)\left\Vert x^t_0-x^*\right\Vert ^2 + \gamma ^{3} \mu ^{-1}n^{2}(n-1)\lef...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.04340474680066109, 0.015589995309710503, -0.020099876448512077, -0.03556014597415924, -0.019138379022479057, -0.014926104806363583, -0.0033461640123277903, 0.045114073902368546, 0.0024113748222589493, -0.00186290149576962, -0.03378976881504059, 0.021473444998264313, -0.039131421595811844,...
2a7269b0559faaeac0aa9b97643760c9d967fc9f
subsection
63
121
Conclusion and open problems
Here the last inequality is because \left\Vert R^t\right\Vert &= \left\Vert \sum _{i=1}^n \nabla f_{\sigma _t(i)} (x^t_{i-1}) - \sum _{i=1}^n \nabla f_{\sigma _t(i)}(x^t_0)\right\Vert \\ &\le \sum _{i=1}^n \left\Vert \nabla f_{\sigma _t(i)} (x^t_{i-1}) - \nabla f_{\sigma _t(i)}(x^t_0)\right\Vert \\ &= \sum _{i=1}^n \l...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05330732837319374, -0.015127136372029781, -0.008589589037001133, -0.009329545311629772, 0.003120019566267729, 0.054710954427719116, 0.0033412433695048094, 0.04290217533707619, 0.004485504701733589, 0.012167313136160374, -0.02230546995997429, -0.00034590071300044656, 0.019421931356191635, ...
30fcac426276c491886d14bb7a776b7265effb82
subsection
64
121
Conclusion and open problems
Again, define error term R^t = \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \sum _{i=1}^n \nabla f_{\sigma _t\left(i\right)}\left(x^t_0\right). Assume F^* being the minimum of function F(\cdot ). For one epoch of RandomShuffle, we have F(x^{t+1}_0) - F^* &\le F(x^t_0) - F^* - \gamma \left\...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05161808803677559, 0.02029924839735031, -0.018910352140665054, -0.020848700776696205, 0.008157855831086636, 0.02594640664756298, 0.019383491948246956, 0.015193911269307137, -0.004716141149401665, 0.0025965485256165266, -0.030937274917960167, -0.005410589277744293, -0.016269924119114876, ...
48ecee59b99c6932edbe444d784366a063786ce3
subsection
65
121
Conclusion and open problems
We have the following decomposition for the error term: R^t &= \sum _{i=1}^n \left[\nabla f_{\sigma _t\left(i\right)} \left(x^t_{i-1}\right) - \nabla f_{\sigma _t\left(i\right)}\left(x^t_{0}\right)\right]\\ &= \sum _{i=1}^n \left[\int _{x^t_0}^{x^t_{i-1}} H_{\sigma _t\left(i\right)}\left(x\right) dx\right]\\ &= \sum _...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ 0.011490252800285816, 0.008789356797933578, -0.01725826784968376, 0.018433233723044395, -0.002781007206067443, -0.00651572085916996, 0.004623567685484886, 0.017227748408913612, 0.0016737544210627675, 0.008484170772135258, -0.03448601812124252, -0.003318897681310773, -0.009979581460356712, ...
a0918220f3233e0b1cde6a0f466f58b16ce994ac
subsection
66
121
Conclusion and open problems
Here we define random variables A^t = -\gamma \sum _{i=1}^n \left[H_{\sigma _t\left(i\right)}(x^t_0)\sum _{j=1}^{i-1} \nabla f_{\sigma _t\left(j\right)} \left(x^t_0\right)\right], B^t = - \gamma \sum _{i=1}^n \left\lbrace H_{\sigma _t\left(i\right)}(x^t_0)\sum _{j=1}^{i-1}\left[\nabla f_{\sigma _t\left(j\right)} \left(...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.044100336730480194, -0.0014925827272236347, -0.021729717031121254, -0.0047533754259347916, -0.00006306519935606048, -0.009438082575798035, -0.001283716526813805, 0.029496869072318077, -0.0001350715901935473, 0.023805025964975357, -0.042360737919807434, 0.028688108548521996, 0.009155779145...
f872f45095a0ab080def0a2cc55b341c1a1c9767
subsection
67
121
Conclusion and open problems
\left\Vert C^t\right\Vert &\le \sum _{i=1}^n \left[\int _{0}^{\left\Vert x^t_{i-1}-x^t_0\right\Vert } \left\Vert H_{\sigma _t\left(i\right)}\left(x^t_0+t\frac{x^t_{i-1}-x^t_0}{\left\Vert x^t_{i-1}-x^t_0\right\Vert }\right)-H_{\sigma _t\left(i\right)}(x^t_0)\right\Vert dt\right]\\ &\le \sum _{i=1}^n \left[L_H \left\Vert...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.012270919978618622, 0.019917352125048637, -0.024740250781178474, 0.0009376799571327865, 0.00083609001012519, -0.005765348672866821, -0.012339601293206215, 0.022740885615348816, -0.007428943179547787, 0.028143752366304398, -0.06074410676956177, 0.005074727814644575, 0.0117214759811759, -...
7e0bfcfd729e1bd86aec80034ae5664fb5419ad2
subsection
68
121
Conclusion and open problems
For the first term in the (REF ), we have further bound: &\ \ \ \ \frac{1}{2}\gamma ^2 n\left(n-1\right) \left\langle \nabla F(x^t_0), \operatorname{\mathrm {\mathbb {E}}}\limits _{i\ne j} H_i(x^t_0) \nabla f_j\left(x^t_0\right)\right\rangle \\ &= \frac{1}{2}\gamma ^2 n^2 \left\langle \nabla F(x^t_0), \operatorname{\m...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.048631586134433746, 0.018168199807405472, 0.004786123055964708, 0.0004216459929011762, -0.01885465532541275, -0.025887014344334602, -0.019434329122304916, 0.03938731551170349, -0.009587501175701618, 0.007055241148918867, -0.0578758604824543, 0.005045450758188963, -0.03307192027568817, 0...
301f86d4798c8239d1af74339429579400da998d
subsection
69
121
Conclusion and open problems
For the third term in (REF ), we use the bound -\gamma \left\langle \nabla F(x^t_0), \operatorname{\mathrm {\mathbb {E}}}\left[C^t\right]\right\rangle &\le \gamma \left\Vert \nabla F(x^t_0)\right\Vert \cdot \left( n^3\gamma ^2L_H G^2 \right)\\ &= \gamma ^3 n^3 \left\Vert \nabla F(x^t_0)\right\Vert L_H G^2\\ &\le \frac...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.044103704392910004, 0.02721000276505947, -0.0044904896058142185, -0.043706923723220825, -0.013650783337652683, -0.023226933553814888, 0.010751231573522091, 0.039464421570301056, 0.00040751099004410207, 0.009072544053196907, -0.04068528488278389, 0.038182515650987625, -0.009255673736333847...
f4ac45c00335d24c1753f0146159b7b162e6f56d
subsection
70
121
Conclusion and open problems
Now assume \frac{1}{2}n\mu \gamma > 4L^2n^2\gamma ^2, which we call assumption 1, (REF ) can be further turned into: &\ \ \ \ \operatorname{\mathrm {\mathbb {E}}}[F(x^t_0) - F^*] \\ &\le (1-n\mu \gamma )\left[F(x^t_0) - F^* \right] + \gamma ^3 n C_1 + n^4\gamma ^4C_2 + n^5\gamma ^5C_3. where C_1 =\frac{1}{2} \mu ^{...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05925506353378296, 0.05614279955625534, -0.01018351037055254, -0.018887169659137726, -0.001701065804809332, -0.012990650720894337, -0.021206112578511238, 0.041252754628658295, 0.006068151909857988, -0.0038636315148323774, -0.033411066979169846, 0.004328945651650429, -0.02051958255469799, ...
8c1897a8e6d414e70a1d2e581ec232c0b8d3461a
subsection
71
121
Conclusion and open problems
What remains to determine is to satisfy the two assumptions: (1) \frac{1}{2}n\mu \gamma > 4L^2n^2\gamma ^2, (2) n\gamma \mu <1. The first is satisfied when \frac{T}{\log T} > 16 \frac{L^2}{\mu ^2} n. The second assumption is satisfied when \frac{T}{\log T} > 2 n. Since 2<\frac{L}{\mu }, the theorem is proved. Proof of...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.06640666723251343, 0.04501371085643768, 0.0005822218372486532, -0.020813118666410446, -0.011108468286693096, 0.01577768661081791, -0.0054130894131958485, 0.020111210644245148, -0.0188294630497694, 0.004650144837796688, -0.016952620819211006, -0.009475767612457275, -0.03390524163842201, ...
0501852ce870cdae3a6e37eb677241f3cf931d2f
subsection
72
121
Conclusion and open problems
Therefore, taking expectation of (REF ) leads to: \operatorname{\mathrm {\mathbb {E}}}[\left\Vert x^{t}_n - x^*\right\Vert ^2] \le \left\Vert x^t_0-x^*\right\Vert ^2 - (2n\gamma -2n^2\gamma ^2L)\left[F(x^t_0) - F(x^*)\right] -2\gamma \operatorname{\mathrm {\mathbb {E}}}\left\langle x^t_0 - x^*, R^t \right\rangle + 2\g...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.05783936008810997, 0.015947792679071426, -0.020907631143927574, -0.021518073976039886, -0.00892771128565073, 0.02589799277484417, 0.012887952849268913, 0.03311646729707718, -0.01180441863834858, 0.013162651099264622, -0.02963694930076599, 0.03543614596128464, -0.040563859045505524, 0.02...
45ca66ca7b74db8580e38a8d82644186383e02b0
subsection
73
121
Conclusion and open problems
Then for i+1, there is \left\Vert \nabla f_{id}(x^t_{i+1}) - \nabla f_{id}(x^t_0)\right\Vert &\le \left\Vert \nabla f_{id}(x^t_{i}) - \nabla f_{id}(x^t_0)\right\Vert + \left\Vert \nabla f_{id}(x^t_{i+1}) - \nabla f_{id}(x^t_i)\right\Vert \\ &\le \left\Vert \nabla f_{id}(x^t_{i}) - \nabla f_{id}(x^t_0)\right\Vert + L\g...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.028889337554574013, -0.0038171925116330385, -0.034062862396240234, -0.012941446155309677, -0.012697268277406693, 0.012788834981620312, 0.019915787503123283, 0.036107856780290604, -0.0030026291497051716, -0.02359372191131115, -0.034673310816287994, 0.002739374525845051, -0.0542686134576797...
ff6ea079ed7cb8c31231179f5c391ee8461bf508
subsection
74
121
Conclusion and open problems
Therefore, we have \left\Vert \nabla f_{id}(x^t_i) - \nabla f_{id}(x^t_0)\right\Vert &\le \left[\sum _{j=0}^{i-1} (1+L\gamma )^j\right] L\gamma (\left\Vert \nabla F(x^t_0)\right\Vert + \delta ) \\ &\le \left[n (1+\frac{1}{n})^n\right] L\gamma (\left\Vert \nabla F(x^t_0)\right\Vert + \delta )\\ &\le 3nL\gamma (\left\Ve...
{ "cite_spans": [] }
1806.10077
Random Shuffling Beats SGD after Finite Epochs
[ "Jeff Z. HaoChen", "Suvrit Sra" ]
[ "math.OC", "stat.ML" ]
2,018
en
Mathematics
[ -0.02238948829472065, 0.011660237796604633, -0.040108777582645416, -0.034858617931604385, -0.01398770697414875, 0.03256930410861969, 0.03180619701743126, 0.012652273289859295, -0.022862613201141357, -0.012949884869158268, -0.027609122917056084, 0.000638623139820993, -0.042031798511743546, ...