-
Notifications
You must be signed in to change notification settings - Fork 7
/
Terenin-A-2022-PhD-Thesis.tex
4638 lines (3898 loc) · 315 KB
/
Terenin-A-2022-PhD-Thesis.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
\documentclass[11pt]{book}
\usepackage[T1]{fontenc}
\usepackage{hyperref}
\usepackage{microtype}
\usepackage{lettrine}
\usepackage{xltabular}
\usepackage{standalone}
\usepackage{pgfplots}
\pgfplotsset{compat=1.17}
\usepgfplotslibrary{external}
\usepgfplotslibrary{groupplots}
\usepgfplotslibrary{fillbetween}
\usetikzlibrary{fadings}
\tikzexternalize[prefix=figures/tex/]
\usepackage[commands,environments,enumerate,citations,notes,a4paper]{AVT}
\bibliography{Terenin-A-2022-PhD-Thesis}
\title{Gaussian Processes and Statistical Decision-making in Non-Euclidean Spaces}
\author{Alexander Terenin}
\date{February 2022}
\begin{document}
\begin{titlepage}
\maketitlehooka
\centering
\huge
\null
\vfill
\thetitle
\par
\vfill
\LARGE
\theauthor
\par
\large
Department of Mathematics
\par
Imperial College London
\par
\vfill
\null
\vfill
a dissertation submitted for the degree of
\par
Doctor of Philosophy
\par
\strut
\par
\thedate
\par
\vfill
\null
\maketitlehookd
\end{titlepage}
\chapter*{Declaration}
This dissertation and all research contained in it are a product of my original work, except where indicated otherwise by explicit statement or reference.
All ideas, quotations, and data originating from the works of others, published or otherwise, are fully acknowledged according to standard practices of the academic discipline.
\chapter*{Copyright}
The copyright of this thesis rests with the author. Unless otherwise indicated, its contents are licensed under a Creative Commons Attribution 4.0 International License (CC BY).
Under this license, you may copy and redistribute the material in any medium or format for both commercial and non-commercial purposes.
You may also create and distribute modified versions of the work.
This is on the condition that you credit the author.
When reusing or sharing this work, ensure you make the license terms clear to others by naming the license and linking to the license text.
Where a work has been adapted, you should indicate that the work has been changed and describe those changes.
Please seek permission from the copyright holder for uses of this work that are not included in this license or permitted under UK Copyright Law.
\chapter*{Acknowledgments}
This thesis is dedicated to the several hundred Twitter users who thought what I was working on was interesting enough to give it a read, or a comment.
Thanks to you, my experiment in writing a thesis in an open-source manner visible to the public was a resounding success.
Getting the opportunity to write this thesis has been a dream come true, and I am delighted to share the experience with you.
The original acknowledgments I had planned to write during most of my doctoral work would have started on a far less positive note.
The sheer volume of kind feedback I've received in the weeks during which I was writing this thesis, almost all of it from people I've never met, has convinced me to leave the past behind, and not write about certain people, actions, and events that do not deserve to be remembered.
Instead, I choose to thank those due to whom I had the chance.
Firstly, I am profoundly grateful to Marc Deisenroth for taking me as a student, and giving me a \emph{third} chance to complete a Ph.D., and to Seth Flaxman for taking me as a co-supervisee in order to make this possible.
Most people who do not succeed the first time, whatever the reason may be, don't even get a second chance, much less a third one.
You believed in me in the darkest moments of my academic career, and anything that I ever accomplish in my career will only have been possible because of that belief.
Secondly, I am also profoundly grateful to David Draper for believing in me, and taking me as a student when nobody was willing to talk to me or take my ideas seriously.
It is thanks to you that I have a master's degree, and though I was not able to finish a doctorate at the time under your guidance, I hope this thesis shows that you did the right thing in supporting me, even when it became clear I had no academic future at the place I was at, and later in helping me return to scientific work.
Thirdly, I am grateful to Eric Xing for hiring me when my career was otherwise in the midst of full-blown chaos.
The time I spent at Petuum convinced me that the machine learning community was the right intellectual home for me, and prompted me to look beyond the narrow scope of ideas I had previously been around.
Though I was not truly ready to be an independent researcher at that time, and therefore accomplished little, I am grateful for the opportunity and all that I learned from it.
I thank my thesis committee members, Michael Bronstein and Stefano Ermon, for examining me and for their comments during the viva.
Your careful reading helped significantly improve the quality of this thesis.
I am particularly grateful for Michael's comment that "[t]he introductory chapters, with enough effort, could probably be transformed into a book."
These are among the kindest words my work has ever received.
I also thank my large and ever-growing set of scientific collaborators.
You all give me confidence that, together, we can solve all of the hardest and most important problems!
I am constantly learning from each and every one of you, and none of the work in this thesis would have been possible without your support, ideas, and opportunity to discuss.
I am particularly grateful to Viacheslav Borovitskiy for both the opportunity to collaborate on all of the work presented in this thesis, and to reconnect with my heritage by constantly practicing my Russian.
Together, we were able to write substantially stronger work than any of us, or at least certainly I, would have alone---the awards we've won together speak for themselves.
A large part of why I have gotten better at practicing mathematics over the years---a skill of fundamental importance---is thanks to working with you.
I thank James T. Wilson, Steindór Sæmundsson, Samuel Cohen, and So Takao, for the opportunity to do research and write together.
I am grateful to Yicheng Luo, Sanket Kamthe, Hugh Salimbeni, Yasemin Bekiroğlu, Sicelukwanda Zwane, K. S. Sesh Kumar, and others in the group who have spoken to me at length about ideas during my time.
I feel fortunate to to have been your labmate, and have learned an immense amount from you over the years.
I thank Andreas Hochlehnert, Kai Biegun, and Lucas Cosier for trusting in me as a researcher enough to allow me to supervise your work.
Watching your ideas develop and grow has been a beautiful experience, and I look forward to seeing you shine in the years to come.
I am grateful to Michael Hutchinson, James-Michael Leahy, Peter Mostowsky, Iskander Azangulov, and others who I have collaborated with on papers, or spoken to at length about ideas.
I am particularly grateful to Brandon Amos both for the opportunity to collaborate, and for suggesting I speak to Marc Deisenroth in the weeks leading up to me becoming his student---without this, my path would have been very different.
I thank Nick Sharp for teaching me how to make the three-dimensional figures in this thesis.
I am grateful to Imperial College, and the Department of Mathematics in particular, for both the opportunity in funding my studies, and for support when things did not go as planned.
I am particularly thankful to Henrik Jensen: thank you for treating me with kindness in difficult times.
I am grateful to Pierre Degond, Darryl Holm, and Chris Hallsworth for your ideas in teaching me mathematics.
I joined Imperial because I wanted to get better at mathematics, and it is thanks largely to you that I feel I succeeded at this.
I am thankful to Måns Magnusson, Leif Jonsson, and Shawfeng Dong for collaborating with me in my early days as a researcher, and helping me learn and get my bearings together.
In particular, I thank Peter Drake and Raya Feldman for first setting me on this path.
I am grateful to Matthew Johnson and Chris De Sa for the counterexamples sent to me in the early times, which convinced me that proper mathematics was of utmost importance and that I had to improve at it in order trust myself to say things that are true.
I am grateful to my friends at Carnegie Mellon University, including Dominic Chen, Aurick Qiao, Willie Neiswanger, Kumail Jaffer, Ziv Scully, Sarah Allen Scully, Sol Boucher, Guillaume Didier, Stefan Muller, Priya Donti, Gabriele Farina, Noam Brown, Ben Blum, Evan Cavallo, and others who convinced me to apply to a PhD program by showing me that many that of the people spending their life studying ideas were just like me.
I am grateful to Yuanran Zhu for showing me the same, but in a different place and time.
I am grateful to friends who have supported me over the years, including to Jon Frost for almost a decade of friendship, as well as Paula Siauchó Unriza, Victor Espinosa, Sam Aragon, and others who know me well and have been there for me over the years.
Even as time had gone on, whether we last spoke days or years ago, I have always found nothing about our friendship to have changed whenever we have had the chance to speak again.
Finally, I want to thank my family, including my mom, Irina Terenina.
Without your sacrifices, I would not have spent my youth in the United States, or had almost any of the opportunities I have had.
I am grateful to my grandmother, Olga Novikova, for your wisdom and for keeping our family together.
I wish you could have seen me graduate and write this thesis.
I am grateful to my dad, Vadim Terenin: your authoritative presentation style has at times blinded me to your ideas.
I am grateful to my stepmom, Rheesa Eddings, for both your deeply insightful way of understanding the world, and for your humanity, diplomacy, kindness, helpful ideas, and support over the years.
My gratitude to family includes those who are family by virtue of their friendship and support for my entire adult life, including Jerry Hirsch, John Halper, Lois Morera, Christian Morera, and Peter Morera.
Thank you for continuing to stay in touch, for being the extended family I never otherwise had, and for being with me as my journey has unfolded.
Though this list of acknowledgments has ballooned beyond what might otherwise be expected in a dissertation, I nonetheless feel compelled to write it in full by virtue of my path being what it was.
It is also without a doubt incomplete, so I am grateful to anyone whose name should have been here but has been accidentally omitted, including whoever suggested I limit all paragraphs written to seven lines or less.
To conclude, I offer you, the reader, my exceeding gratitude for taking the time to read my work and ideas.
\chapter*{Abstract}
Bayesian learning using Gaussian processes provides a foundational framework for making decisions in a manner that balances what is known with what could be learned by gathering data.
In this dissertation, we develop techniques for broadening the applicability of Gaussian processes.
This is done in two ways.
Firstly, we develop pathwise conditioning techniques for Gaussian processes, which allow one to express posterior random functions as prior random functions plus a dependent update term.
We introduce a wide class of efficient approximations built from this viewpoint, which can be randomly sampled once in advance, and evaluated at arbitrary locations without any subsequent stochasticity.
This key property improves efficiency and makes it simpler to deploy Gaussian process models in decision-making settings.
Secondly, we develop a collection of Gaussian process models over non-Euclidean spaces, including Riemannian manifolds and graphs.
We derive fully constructive expressions for the covariance kernels of scalar-valued Gaussian processes on Riemannian manifolds and graphs.
Building on these ideas, we describe a formalism for defining vector-valued Gaussian processes on Riemannian manifolds.
The introduced techniques allow all of these models to be trained using standard computational methods.
In total, these contributions make Gaussian processes easier to work with and allow them to be used within a wider class of domains in an effective and principled manner.
This, in turn, makes it possible to potentially apply Gaussian processes to novel decision-making settings.
\tableofcontents
\listoffigures
\chapter*{Principal Notation}
\chaptermark{Principal Notation}
\section*{General}
\newif\ifnodotfill
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\ifnodotfill\else\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc\fi} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$(\Omega,\c{F},\P)$ & Probability space & \labelcpageref{ntn:prob} \\
$\E(\.)$ & Expectation of a real-valued random variable & \labelcpageref{ntn:expectation-cov} \\
$\Cov(\.,\.)$ & Covariance between two real-valued random variables & \labelcpageref{ntn:expectation-cov} \\
$y \~ \pi$ & Random variable $y$ with distribution $\pi$ & \labelcpageref{ntn:rv} \\
$\c{M}_1(X)$ & Space of probability measures over $X$ & \labelcpageref{ntn:space-of-measures} \\
$D_{\f{KL}}(\.||\.)$ & Kullback--Leibler divergence between probability measures & \labelcpageref{ntn:kl-div} \\
$W_{p,d}(\.,\.)$ & Wasserstein distance of $p$th order over distance $d$ & \labelcpageref{ntn:wass-dist} \\
\global\nodotfilltrue
$\N$ & Natural numbers, not including zero \\
$\Z$ & Integers \\
$\R$ & Real numbers \\
$\C$ & Complex numbers \\
$\R^d$ & Euclidean space of dimension $d$ \\
$\R^X$ & Space of functions from $X$ to $\R$ \\
$\v{a}$ & Euclidean vector (bold italic letters) \\
$\m{A}$ & Matrix (bold upface letters) \\
\global\nodotfillfalse
$\oplus$ & Direct sum of vector spaces & \labelcpageref{ntn:direct-sum} \\
$V^*$ & Dual space of a topological vector space & \labelcpageref{ntn:dual-space} \\
$\c{A}^*$ & Adjoint of a bounded linear operator & \labelcpageref{ntn:bounded-operators} \\
$\norm{\.}$ & Norm of a vector in a Banach space & \labelcpageref{ntn:banach-hilbert} \\
$\innerprod{\.}{\.}$ & Inner product between vectors in a Hilbert space & \labelcpageref{ntn:banach-hilbert} \\
$\dualprod{\.}{\.}$ & Duality pairing between topological vector spaces & \labelcpageref{ntn:dual-pair} \\
$L(V;W)$ & Banach space of bounded linear operators & \labelcpageref{ntn:bounded-operators} \\
$C^0(X;\R)$ & Banach space of real-valued continuous functions & \labelcpageref{ntn:conts-fns} \\
$L^p(X;\R)$ & Lebesgue space & \labelcpageref{ntn:leb-space} \\
\end{xltabular}
\section*{Bayesian learning}
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$(\Theta,\mathit\Theta)$ & Measurable space representing quantity of interest & \labelcpageref{ntn:meas} \\
$(Y,\c{Y})$ & Measurable space representing the data & \labelcpageref{ntn:meas} \\
$A_{(\.)}$ & Measurable set & \labelcpageref{ntn:meas} \\
$\theta, y$ & Random variables & \labelcpageref{ntn:rv} \\
$\pi_{(\.)}$ & Probability measure & \labelcpageref{ntn:prob} \\
$f_{(\.)}$ & Probability density & \labelcpageref{ntn:density} \\
$(\.\given\.)$ & Jointly measurable stochastic process & \labelcpageref{ntn:jmsp} \\
$\pi_{(\. \given \.)}$ & Probability kernel & \labelcpageref{ntn:prob-ker} \\
$f_{(\.\given\.)}$ & Conditional probability density & \labelcpageref{ntn:density} \\
\end{xltabular}
\section*{Markov decision processes}
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$S$ & State space & \labelcpageref{ntn:mdp} \\
$A$ & Action space & \labelcpageref{ntn:mdp} \\
$r(\.\given\.,\.)$ & Reward kernel & \labelcpageref{ntn:mdp} \\
$p(\.\given\.,\.)$ & Transition kernel & \labelcpageref{ntn:mdp} \\
$\pi$ & Policy & \labelcpageref{ntn:mdp-policy} \\
$V^{(\pi)}$ & Value function with respect to policy $\pi$ & \labelcpageref{ntn:mdp-value} \\
$R^{(\pi)}$ & Regret of policy $\pi$ & \labelcpageref{ntn:mdp-regret} \\
\end{xltabular}
\section*{Multi-armed bandits}
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$X$ & Space of arms & \labelcpageref{ntn:mab} \\
$K$ & Number of arms & \labelcpageref{ntn:mab-number-arms} \\
$T$ & Maximum time & \labelcpageref{ntn:mab-regret} \\
$y$ & Observed stochastic rewards & \labelcpageref{ntn:mab} \\
$f(\.)$ & Expected rewards & \labelcpageref{ntn:mab} \\
$\eps$ & Observation noise & \labelcpageref{ntn:mab} \\
$R$ & Cumulative stochastic regret of a bandit algorithm & \labelcpageref{ntn:mab-regret} \\
$\Delta(\.)$ & Regret of a specific arm & \labelcpageref{ntn:mab-regret-arms} \\
$n_t$ & Cumulative number of arm pulls up to time $t$ & \labelcpageref{ntn:mab-regret-arms} \\
$f^+_t(\.)$ & Upper confidence bound at time $t$ & \labelcpageref{ntn:mab-ucb} \\
$\mu_t(\.)$ & Empirical mean rewards at time $t$ & \labelcpageref{ntn:mab-means} \\
$\sigma_t(\.)$ & Width of confidence bounds at time $t$ & \labelcpageref{ntn:mab-ucb} \\
$c_t$ & Scaling constant for confidence bound width at time $t$ & \labelcpageref{ntn:mab-ucb} \\
\end{xltabular}
\section*{Gaussian processes}
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$\f{N}(\mu,\sigma^2)$ & Gaussian distribution with mean $\mu$ and variance $\sigma^2$ & \labelcpageref{ntn:norm-dist} \\
$\f{N}(\v\mu,\m\Sigma)$ & Multivariate Gaussian with mean $\v\mu$ and covariance $\m\Sigma$ & \labelcpageref{ntn:mvn-dist} \\
$k(x,x')$ & Positive semi-definite kernel & \labelcpageref{ntn:kernel} \\
$\m{K}_{\v{x}\v{x}'}$ & Kernel matrix with entries $k(x_i, x'_j)$ & \labelcpageref{ntn:kernel} \\
$\f{GP}(\mu,k)$ & Gaussian process with mean $\mu$ and covariance kernel $k$ & \labelcpageref{ntn:gp-dist} \\
$f$ & Prior Gaussian process & \labelcpageref{ntn:gp-model} \\
$\v\eps$ & Likelihood noise & \labelcpageref{ntn:gp-model} \\
$\m\Sigma$ & Covariance of likelihood noise & \labelcpageref{ntn:gp-model} \\
$(\v{x},\v\gamma)$ & Data used to condition the model $y_i = f(x_i) + \eps_i$ & \labelcpageref{ntn:gp-model} \\
$f\given\v{y}$ & Posterior Gaussian process conditioned on $\v{y} = \v\gamma$ & \labelcpageref{ntn:gp-model} \\
$n$ & Number of training data points & \labelcpageref{ntn:gp-complexity} \\
$n_*$ & Number of evaluation locations & \labelcpageref{ntn:gp-complexity} \\
$\tilde{f}$ & Approximate prior defined via a finite basis expansion & \labelcpageref{ntn:gp-approx-prior} \\
$\phi_i(\.)$ & Basis function used by approximate prior & \labelcpageref{ntn:gp-approx-prior} \\
$w_i$ & Random weight of basis function $\phi_i$ & \labelcpageref{ntn:gp-approx-prior} \\
$\ell$ & Number of basis functions used in approximate prior & \labelcpageref{ntn:gp-approx-prior} \\
$k(x_j,\.)$ & Canonical basis function & \labelcpageref{ntn:canonical-basis-fns} \\
$v_j$ & Random weight of canonical basis function $k(x_j,\.)$ & \labelcpageref{ntn:canonical-basis-fns} \\
$H_k$ & Reproducing kernel Hilbert space with kernel $k$ & \labelcpageref{ntn:rkhs} \\
$\v\phi$ & Finite-dimensional approximate feature map & \labelcpageref{ntn:approx-feature-map} \\
$m$ & Number of inducing points in sparse approximation & \labelcpageref{ntn:gp-num-inducing} \\
\end{xltabular}
\section*{Manifolds and graphs}
\begin{xltabular}{\textwidth}{@{} p{1.625cm} X<{\unskip\kern-1pc\leaders\hbox{\makebox[1pc][r]{\makebox[0pc]{.}}}\hfill\kern0pc} @{} >{\raggedleft\arraybackslash}p{0.625cm} @{}}
$X$ & Manifold & \labelcpageref{ntn:manifold} \\
$TX$ & Tangent bundle & \labelcpageref{ntn:tangent-bdl} \\
$T^*X$ & Cotangent bundle & \labelcpageref{ntn:cotangent-bdl} \\
$C^\infty(X)$ & Space of smooth real-valued functions & \labelcpageref{ntn:smooth-fns} \\
$\Gamma(TX)$ & Space of smooth sections & \labelcpageref{ntn:smooth-sections} \\
$d$ & Dimension of $X$ & \labelcpageref{ntn:manifold} \\
$f_*$ & Pushforward of a function $f$ & \labelcpageref{ntn:vector-pushforward} \\
$g(\.,\.)$ & Metric tensor of a Riemannian manifold & \labelcpageref{ntn:metric-tensor} \\
$\lap_g$ & Laplace--Beltrami operator induced by $g$ & \labelcpageref{ntn:laplace-beltrami} \\
$(\lambda_n, f_n)$ & Laplace--Beltrami eigenpair & \labelcpageref{ntn:laplace-beltrami-eigenpairs} \\
$\Phi(\lap_g)$ & Operator defined via functional calculus & \labelcpageref{ntn:functional-calculus} \\
$\c{W}_g$ & Gaussian white noise induced by $g$ & \labelcpageref{ntn:riemannian-white-noise} \\
$\m\lap$ & Graph Laplacian & \labelcpageref{ntn:graph-laplacian} \\
$\bc{W}$ & Standard Gaussian & \labelcpageref{ntn:graph-white-noise} \\
$\Phi(\m\lap)$ & Matrix defined via functional calculus & \labelcpageref{ntn:matrix-functional-calculus} \\
$F$ & Frame over $X$ & \labelcpageref{ntn:frame} \\
$\m{P}_F(\.)$ & Projection map induced by $F$ & \labelcpageref{ntn:frame-projection} \\
\end{xltabular}
\chapter{Introduction}
\label{ch:intro}
\lettrine{L}{earning} from experience in order to change behavior is one of the defining abilities of biological systems, which differentiates them from other kinds of systems found in the world.
Replicating the processes biological systems use to learn and adapt is a fundamental goal of science and technology.
To this end, the development of mathematical formalisms rich enough to capture the notion of learning is one of the crowning achievements of statistics, machine learning, and artificial intelligence.
One such formalism is the \emph{Bayesian} view of learning.
The idea behind Bayesian learning is to represent the information known about the quantity of interest using probability.
To do so, the relationship between the quantity of interest and the data is formalized as a joint probability distribution.
This gives rise to a conditional probability distribution describing what was learned about the quantity of interest by observing the data.
Bayesian learning fits naturally within a theory of \emph{decision}, which describes how an abstract decision system should select actions in pursuit of a goal.
This is done by learning how different actions affect pursuit of the goal, and selecting optimal actions consistent with what was learned.
By virtue of being probabilistic, such decision systems assess and propagate uncertainty, enabling them to balance what is already known with what could be learned by taking actions---a concept known as the \emph{explore-exploit tradeoff}.
The performance of a decision system can be evaluated by examining how quickly its decisions improve and become optimal.
A decision system's \emph{regret} is the reduction in its quality of decisions by virtue of not knowing the quantity of interest in advance.
In most non-trivial settings, one can show that some regret is inevitable: a decision-making system must make some degree of mistakes in order to learn.
A decision system is considered \emph{optimal} if its regret is within a constant factor of the best possible regret.
Decision systems with optimal or close-to-optimal regret require less data in order to solve their respective tasks, and are called \emph{data-efficient}.
Data-efficiency is a key concern in practical settings, where data-collection takes time and can be expensive.
By virtue of resolving explore-exploit tradeoffs in a manner amenable to regret analysis, the Bayesian formalism gives broad tools for constructing data-efficient decision systems.
The key limitation of the Bayesian approach is that it often leads to computational problems which are intractable.
Conditional distributions generally contain more information than actually needed to make optimal decisions, yet calculating them is largely unavoidable.
Probabilistic decision systems are thus most attractive in settings where their strengths---including data-efficiency, solid technical foundations, and amenability to analysis---can shine, while computational costs are kept under control.
In my view, \emph{Gaussian processes} are one such setting: they are powerful enough to model wide classes of unknown quantities of interest, yet their computational costs are generally polynomial.
Better yet, Gaussian-process-based decision systems have demonstrated excellent performance in real-world scientific applications.
Studying Gaussian processes is therefore a promising avenue towards improved understanding of Bayesian learning and Bayesian decision-making in pursuit of artificial intelligence.
The goals of this dissertation are twofold: (i) to make Gaussian processes easier to work with when used within larger decision systems, and (ii) to expand the set of settings where Gaussian processes can used, enabling construction of decision systems for applications not previously considered.
Contributions toward (i) include path-wise conditioning techniques studied in \Cref{ch:pathwise}, and contributions toward (ii) include non-Euclidean Gaussian processes studied in \Cref{ch:noneuclidean}.
Following these, \Cref{ch:discussion} concludes.
To pursue these goals, it is critically important that all of the concepts described in the preceding paragraphs be made into rigorous mathematics, so that the ideas described in the sequel ultimately reduce to definitions and implications, and not metaphor or opinion.
Together, we therefore begin by defining the key mathematical notions needed.
\subsection*{Contributions}
The work presented in this thesis is published as a series of papers.
My contributions to the individual works are primarily on the theoretical and methodological side, and are described below.
In \Cref{ch:pathwise}, we present pathwise conditioning techniques: this work is published as \textcite{wilson20,wilson21}.
My contributions include (i) development of the random-function-based formalism for describing pathwise conditioning of Gaussian processes, (ii) error analysis of basis-function-approximation-based pathwise sampling, (iii) re-interpretation of inducing point methods, and (iv) review of prior sampling methods.
All of these were developed jointly with the other authors.
In \Cref{ch:noneuclidean}, we present Matérn Gaussian processes in non-Euclidean settings: this work is published as \textcite{borovitskiy20,borovitskiy21,jacquier21,hutchinson21}.
My contributions include (i) developing the differential-geometric and stochastic-partial-differential-equation-based formalisms for defining these processes, and (ii) describing computational techniques for working with these models in practice.
All of these ideas were developed jointly with the other authors.
\section{Bayesian learning}
The first concept we develop in depth along our path towards a mathematically precise understanding of statistical decision-making is \emph{Bayesian learning}---a mathematical formalism for reasoning about unknown quantities of interest on the basis of data.
Bayesian learning is a \emph{probabilistic} theory: a model specifies how the quantity of interest and the data depend on one another within a probability distribution.
Learning entails calculating how the distribution of the quantity of interest changes upon observing the data.
One of the key strengths of Bayesian theory is that it applies in wide generality, owing to the substantial scope of probability theory.
In particular, one can study learning of quantities of interest that are function-valued using observed data consisting of pointwise function evaluations.
This, however, requires a non-elementary treatment, as one cannot rely solely on probability densities to define what conditional distributions are.
We therefore begin by recalling the necessary mathematical formalism.
For an overview of Bayesian methods from a model-building perspective, see \textcite{gelman14}, and for an overview from a mathematical perspective see \textcite{ghosal17,gine15}.
\subsection{Review of probability theory}
We adopt the language of measure-theoretic probability, which we now describe.
To ease presentation, we state the definitions together with useful ways of thinking about them.
We use \textcite{kallenberg06} as our standard reference, along with certain results from \textcite{bogachev07a,bogachev07b}.
\label{ntn:meas}
We say that \emphmarginnote{measurable space} is a pair $(Y,\c{Y})$ consisting of a set $Y$ and a $\sigma$-algebra $\c{Y}$ over $Y$.
A \emph{$\sigma$-algebra} is a set of subsets of $Y$ containing the space itself which is closed under countable unions, intersections, complements, and therefore set-theoretic monotone limits.
These can be reinterpreted as Boolean logical operations, so $\c{Y}$ can be thought of as the set of all true-false questions one can ask about elements of the set $Y$.
These questions, then, are closed under \emph{and/or/not} operations and monotone limits thereof.
\parmarginnote{Product of measurable spaces}
Given two measurable spaces $(Y,\c{Y})$ and $(\Theta,\mathit\Theta)$, if we form the Cartesian product $Y \x \Theta$, then we can define the \emph{product $\sigma$-algebra} $\c{Y}\ox\mathit\Theta$ as the smallest $\sigma$-algebra containing all sets of the form $A_y \x A_\theta$ with $A_y \in\c{Y}$ and $A_\theta\in\mathit\Theta$.
\marginnote{Measurable subspace}
For a measurable subset $Y' \subseteq Y$, we can define $\c{Y}' = \{A_y \^ Y' : A_y \in\c{Y}\}$, which is also a $\sigma$-algebra, thereby making $(Y',\c{Y}')$ into a measurable space.
We call $\c{Y}'$ the \emph{subset $\sigma$-algebra}.
If $Y$ is a topological space, the \emphmarginnote{Borel $\sigma$-algebra} $\c{B}(Y)$ is defined as the smallest $\sigma$-algebra containing the open sets in the topology.
Topological spaces, in turn, are sets equipped with additional structure enabling them to admit notions such as locality and convergence---we review these spaces in additional detail in \Cref{ch:noneuclidean}.
\parmarginnote{Measurable function}
A map $f : Y \-> Y'$ between measurable spaces $(Y,\c{Y})$ and $(Y',\c{Y}')$ is said to be \emph{measurable} if its preimage defines a map $f^{-1} : \c{Y}' \-> \c{Y}$ between the respective $\sigma$-algebras.
This intuitively means that true-false questions for the space $Y'$ can be asked and answered relative to those in $Y$.
On product spaces, a map $f : Y \-> \Theta \x \Theta'$ is measurable if its components are measurable, but a map $f : Y \x Y' \-> \Theta$ \emph{need not} automatically be measurable if $f(\.,y') : Y \-> \Theta$ and $f(y,\.) : Y' \-> \Theta$ are measurable for all $y$ and $y'$.
\label{ntn:prob}
A \emphmarginnote{probability measure} is a non-negative countably additive map $\pi_y : \c{Y} \-> \R$ satisfying $\pi_Y(Y) = 1$.
This can be thought of as a map that takes a true/false question, and assigns a number indicating how close to true or false its answer is according to the measure---in this view, probability measures describe uncertainty.
A \emphmarginnote{probability space} $(\Omega,\c{F},\P)$ consists of a measurable space and probability measure.
The set $\Omega$ can be thought of as a space of abstract random numbers, with a random number generator described by $\P$.
\label{ntn:rv}
Given a probability space, we say that a \emphmarginnote{random variable} is a measurable map $y : \Omega \-> Y$.
A random variable, then, maps random numbers $\omega\in\Omega$ into the space $Y$.
The \emph{distribution}\marginnote{Distribution} of a random variable is defined as the \emph{pushforward measure} $\pi_y = y_* \P$ where $y_*$ is defined as $(y_* \P)(A_y) = \P(y^{-1}(A_y))$ for all $A_y\in\c{Y}$.
The probability of an event $A_y$, then, is determined by measuring the probability of random numbers under which $A_y$ occurs---measurability guarantees this is possible.
\parmarginnote{Notation for random variables}
In this work, we will generally \emph{not} adopt the standard convention of suppressing $\omega$-arguments of random variables from our notation, and will write such arguments explicitly in cases that other function arguments are also used.
Though this makes expressions slightly denser, it also avoids ambiguity and presents the mathematics more precisely.
\label{ntn:expectation-cov}
The \emphmarginnote{expectation} $\E(y)$ of a real-valued random variable $y$, if it exists, is defined as its integral with respect to $\P$.
This can be thought of as the average value of $y$ under the random numbers generated from $\P$, with averaging performed via the algebraic structure of the reals.
This notion extends to vector-valued random variables.
The \emphmarginnote{covariance} of two real-valued random variables, if it exists, is defined as $\Cov(y,y') = \E(yy') - \E(y)\E(y')$.
From this, we define the covariance of a finite-dimensional random vector componentwise.
\parmarginnote{Equality of random variables}
There are multiple senses in which we can say random variables are equal.
We say that $y = y'$ \emph{surely} if they are equal as mathematical functions.
This notion is essentially never used, because it is exceedingly strong.
We say that $y = y'$ \emph{almost surely} if $\P(y \neq y') = 0$, in which case $y$ and $y'$ are equal as functions, except possibly on a set of probability zero.
We say that $y = y'$ \emph{in distribution} if $y_* \P = y'_* \P$, meaning their distributions are equal.
\parmarginnote{Existence of a random variable with given distribution}
For a given probability space $(\Omega,\c{F},\P)$, a random variable $y : \Omega \-> Y$ with distribution $\pi_y$ need not exist.
However, if it does not, there always exists another probability space $(\Omega',\c{F}',\P')$ and a measurable map $i : \Omega' \-> \Omega$ such that $\P = i_* \P'$, and for which a $\pi_y$-distributed random variable $y : \Omega' \-> Y$ \emph{does} exist.
We thus implicitly assume all probability spaces are large enough to ensure all random variables with prescribed distributions exist.
\label{ntn:prob-ker}
\parmarginnote{Probability kernel}
We will be interested in probability measures extended to allow them to be parameterized by other quantities.
A \emph{probability kernel} is defined to be a map $\pi_{y\given\theta} : \c{Y} \x \Theta \-> \R$ satisfying two conditions: (i) the map $\pi_{y\given\theta}(\.,\vartheta) : \c{Y} \-> \R$ is a probability measure for all $\vartheta\in\Theta$, and (ii) the map $\pi_{y\given\theta}(A_y,\.) : \Theta \-> \R$ is measurable for all $A_y\in\c{Y}$.
In particular, this means that a probability kernel can be integrated against other measures.
\label{ntn:jmsp}
\parmarginnote{Conditional distribution}
Random variables can also be extended to parameterize them by other quantities.
A $\c{F}\ox \mathit\Theta$-measurable map $y\given\theta : \Omega \x \Theta \-> Y$ is called a \emph{jointly measurable stochastic process}---we largely eschew this terminology to emphasize the Bayesian formalism.
In particular, unlike most presentations of stochastic processes, we \emph{never} think of $\Theta$ as representing time.
Define the \emph{conditional distribution} of $y\given\theta$ to be $(y\given\theta)_*\P$, with the pushforward taken in the first argument---by \Cref{lem:rcrv-rvm-equiv}, this is a probability kernel.
\subsection{Bayes' Rule for probability measures}
The key idea behind Bayesian learning is to formalize \emph{learning} using the concept of \emph{conditional probability}.
This entails (i) building a model quantifying the relationship between the \emph{quantity of interest} $\theta$, and the \emph{data} $y$, and (ii) quantifying what is learned using Bayes' Rule.
We now begin to make these concepts precise in the language of measure-theoretic probability.
\begin{definition}[Bayesian model]
Let $Y$ and $\Theta$ be sets.
A \emph{Bayesian model} is a probability measure defined on $\Theta \x Y$.
\end{definition}
A model can be constructed in a number of different ways.
The most common technique is to specify two components: (i) a probability distribution describing what is known about the quantity of interest external to the data, and (ii) how the data relates to the quantity of interest.
These are called the \emph{prior distribution} and \emph{likelihood}, respectively.
For example, a simple Gaussian model is given by
\[
y_i \given \theta &\~[N](\theta, 1)
&
\theta &\~[N](0,1)
.
\]
In this model, $\theta$ is an unknown scalar assigned a standard normal prior, while $y_i\given\theta$ is assigned a Gaussian likelihood centered at $\theta$.
Our first goal is to make this notation into proper mathematics.
\begin{proposition}
A Bayesian model can be constructed in the following ways.
\1 Using measures: integrate the following two components.
\1 \emph{Prior}: a probability measure $\pi_\theta : \mathit\Theta \-> \R$.
\2 \emph{Likelihood}: a probability kernel $\pi_{y\given\theta} : \c{Y}\x\Theta \-> \R$.
\0
\2 Using random variables: compose the following two components.
\1 \emph{Prior}: a random variable $\theta:\Omega\->\Theta$.
\2 \emph{Likelihood}: a jointly measurable stochastic process $y\given\theta: \Omega\x\Theta\->Y$.
\0
\0
Moreover, if one takes the pushforward of the composition of the obtained random variables with respect to $\P\x\P$, then the two constructions coincide.
\end{proposition}
\begin{proof}
For the former, define
\[
\pi_{\theta,y}(A_\theta\x A_y) = \int_{A_\theta} \pi_{y\given\theta}(A_y\given\theta) \d\pi_\theta(\theta)
\]
which extends to the full product $\sigma$-algebra by \Cref{lem:cyl-prod}, giving the desired probability measure.
For the latter, define the random variable
\[
\gamma : \Omega\x\Omega &\-> \Theta \x Y
&
\gamma : (\omega,\omega') &\|> (\theta(\omega), (y\given\theta)(\omega',\theta(\omega)))
\]
and take $\pi_{\theta,y} = \gamma_* (\P\x\P)$.
We now prove these expressions coincide.
Write
\[
\pi_{\theta,y}(A_\theta\x A_y) &= \int_{A_\theta} \pi_{y\given\theta}(A_y\given\theta) \d\pi_\theta(\theta)
\\
&= \int_\Omega\1_{\theta(\omega)\in A_\theta} \pi_{y\given\theta}(A_y\given\theta(\omega)) \d\bb{P}(\omega)
\\
&= \int_\Omega\1_{\theta(\omega)\in A_\theta} \int_\Omega \1_{(y\given\theta)(\omega',\theta(\omega)) \in A_y} \d\bb{P}(\omega') \d\bb{P}(\omega)
\\
&= \int_\Omega\int_\Omega \1_{(\theta(\omega),(y\given\theta)(\omega',\theta(\omega))) \in A_\theta \x A_y} \d\bb{P}(\omega') \d\bb{P}(\omega)
\\
&= (\P\x\P)(\gamma \in A_\theta \x A_y)
\]
which follows using Tonelli's Theorem and \Cref{lem:cyl-prod}.
\end{proof}
These two components should be interpreted as follows: $\theta$ describes what is known about the quantity of interest external to the data, and $y\given\theta$ describes how the data $y$ relates to the quantity of interest.
More specifically, the likelihood describes how $y$ would be distributed if $\theta$ was known and equal to the conditioned value.
The condition that the pushforward of the composition of the prior and likelihood needs to be taken with respect to the product measure $\P\x\P$ should be interpreted as a kind of \emph{non-duplicate dependence} condition which ensures that $y\given\theta$ depends on $\theta$ only through its second argument, and not through the abstract random numbers $\omega$.
Given a Bayesian model, we can formalize the notion of what is \emph{learned} about $\theta$ from observing $y$ as its conditional probability distribution---note that this is meant in a \emph{distributional} sense, not a random-variable-based sense.
The formulation we use requires topological assumptions: we say that a topological space is \emph{Polish} if it is homeomorphic to a complete separable metric space.
Note that any two uncountable Polish spaces are Borel-isomorphic: see \textcite[Chapter 1]{villani08}, for discussion.
The key result is given as follows.
\begin{figure}
\begin{subfigure}{0.55\textwidth}
\input{figures/tex/bayes.tex}
\end{subfigure}
\begin{subfigure}{0.44\textwidth}
\tikzset{external/export next=false}
\begin{tikzpicture}
\begin{axis}[axis lines={none}, height={7.5cm}, axis equal, view={{310}{22.5}}, xmin={-0.6}, xmax={1.75}, ymin={-1.75}, ymax={1.75}, zmin={0}, zmax={1}]
\node at (0,0.03,0.31) {\includegraphics[scale=0.25]{figures/png/bayes.png}};
\addplot3[no markers, domain={-1.75:1.75}, smooth, samples={32}, samples y={0}, thick, color={black}, opacity={0}] ({-0.5},{x},{0.9 * exp(-x^2) * exp(-(x + 0.5)^2)});
\addplot3[no markers, very thick, color={black}, line cap={round}, dashed] coordinates {(-0.5,-1.75,0) (-0.5,1.75,0)};
\end{axis}
\end{tikzpicture}
\end{subfigure}
\caption[Bayesian models]{Here, we illustrate a Bayesian model. For all possible values of the quantity of interest, the likelihood describes the respective distribution of the data, shown on the left. This is combined with the prior, also shown on the left, to form a joint distribution, shown on the right. From this, the posterior distribution, also shown on the right, is obtained by conditioning. Note that the two plots are shown not to scale, and that the joint and posterior live on different spaces, hence are normalized differently.}
\label{fig:bayes-rule}
\end{figure}
\begin{result}[Bayes' Rule]
Suppose that $\Theta$ is a Polish topological space, and let $\pi_y = \pi_{\theta,y}(\Theta\x\.)$.
Then for every Bayesian model $\pi_{\theta,y}$ there is a $\pi_y\ae[-]$ unique probability kernel $\pi_{\theta\given y}$ satisfying
\[
\pi_{\theta,y}(A_\theta \x A_y) = \int_{A_y} \pi_{\theta\given y}(A_\theta \given y) \d\pi_y(y)
\]
which we call the \emph{posterior distribution}.
\end{result}
\begin{proof}
The claim follows directly from \textcite[Corollary 10.4.15]{bogachev07b}.
See also \textcite[Theorem 5.3 and Theorem 5.4]{kallenberg06}, \textcite[Theorem 5.3.1]{ambrosio08}, and \textcite{chang97} for variations of this result.
\end{proof}
This result is illustrated in \Cref{fig:bayes-rule}, and shows that given a Bayesian model, the posterior distribution describing what was learned from the data exists.
In its full abstract formulation, however, Bayes' Rule is \emph{non-constructive}, and it is not at all clear how to calculate any kind of useful formula from it.
Moreover, the null sets can be problematic: to ensure that $\pi_{\theta\given y}$ is defined pointwise, one needs further properties such as continuity.
Fortunately, in many settings these issues are resolved by virtue of additional structure.
\label{ntn:density}
The simplest such structure occurs when $\pi_{\theta,y}$ admits a density with respect to a product measure $\lambda$.
In this case, it follows from \Cref{lem:ref-density} that $\pi_\theta$ admits the density $f_\theta$ with respect to $\lambda(\.\x Y)$, and similarly for $\pi_y$ and $f_y$.
Define a \emph{conditional density} as the ratio of joint and marginal densities, for instance $f_{\theta\given y} = \frac{f_{\theta,y}}{f_y}$, and let $\propto$ denote equality up to a multiplicative proportionality constant.
Then, we have the following.
\begin{proposition}[Bayes' Rule for densities]
Suppose that $\pi_{\theta,y}$ admits the density $f_{\theta,y}$ with respect to $\lambda_{\theta,y}$.
Then we have
\[
f_{\theta\given y} \propto f_{y\given\theta}f_\theta
.
\]
\end{proposition}
\begin{proof}
We have
\[
f_{\theta,y} &= \frac{f_{\theta,y}}{f_\theta} f_\theta = f_{y\given\theta}f_\theta
&
f_{\theta,y} &= \frac{f_{\theta,y}}{f_y} f_y = f_{\theta\given y}f_y
\]
which, when combined, give the result.
\end{proof}
Sometimes, the knowledge of $f_{\theta\given y}$ up to proportionality is enough to fully deduce its form.
For example, if the likelihood is Gaussian with unknown mean and known variance, and the prior is Gaussian, then posterior is also Gaussian.
\marginnote{Conjugacy}
In cases such as this, where the prior and posterior land in the same parameterized class of distributions, the pair $(\pi_{y\given\theta}, \pi_\theta)$ are called \emph{conjugate}.
Bayes' Rule for densities is remarkable in its generality yet restrictiveness.
On the one hand, we require no direct assumptions about the spaces $\Theta$ and $Y$, and in particular allow for real spaces, discrete spaces, and Riemannian manifolds.
On the other hand, other than in the aforementioned settings, where we can employ the Lebesgue, counting, and Riemannian volume measures, respectively, finding a suitable reference measure can be difficult.
We will work with posterior distributions in infinite-dimensional vector spaces in the sequel---there, densities are either not available or not convenient.
In those cases, it is enough to calculate the posterior on arbitrary finite-dimensional marginal projections to uniquely determine its value on the full infinite-dimensional space.
\begin{proposition}[Conditioning and marginalization]
Conditioning and marginalization commute: given a probability measure $\pi_{\theta,\theta',y}$, if Bayes' Rule holds, then we have $\pi_y\ae[-]$ that
\[
\pi_{\theta \given y} = \pi_{\theta,\theta' \given y}(\. \x \Theta')
.
\]
\end{proposition}
\begin{proof}
Let $\pi_{\theta,\theta'\given y}$ be the $\pi_y\ae[-]$ unique probability kernel given by the Disintegration Theorem satisfying
\[
\pi_{\theta,\theta',y}(A_{\theta,\theta'} \x A_y) = \int_{A_y} \pi_{\theta,\theta'\given y}(A_{\theta,\theta'} \given y) \d\pi_y(y)
.
\]
Plugging in $A_{\theta} \x \Theta'$ for $A_{\theta,\theta'}$ gives
\[
\pi_{\theta,y}(A_\theta \x A_y) = \int_{A_y} \pi_{\theta,\theta'\given y}(A_\theta \x \Theta' \given y) \d\pi_y(y)
.
\]
On the other hand, applying the Disintegration Theorem to $\pi_{\theta,y}$ gives a $\pi_y\ae[-]$ unique probability kernel satisfying
\[
\pi_{\theta,y}(A_\theta \x A_y) = \int_{A_y} \pi_{\theta\given y}(A_\theta \given y) \d\pi_y(y)
.
\]
Since both $\pi_{\theta,\theta'\given y}$ and $\pi_{\theta,y}$ are defined with respect to the same null sets, we conclude by uniqueness that they coincide, and the claim follows.
\end{proof}
This result makes densities into a substantially more powerful tool than they would be otherwise, since it enables us to use them for calculating posterior distributions even where they are not directly available.
In particular, we can map an infinite-dimensional function space into finite-dimensional vector spaces induced by pointwise function evaluations at arbitrary points, enabling us to calculate posterior distributions in such settings, in spite of no suitable densities existing directly in the space of interest.
\label{ntn:space-of-measures}
We now introduce the \emph{variational formulation} of Bayes' Rule, which expresses the posterior as the solution to an infinite-dimensional optimization problem.
Let $\c{M}_1(\Theta)$ be the space of all probability measures over $\Theta$.
\label{ntn:kl-div}
\begin{proposition}[Bayes' Rule in variational form]
\label{prop:variational-bayes}
Let $\pi_{y,\theta}$ be a Bayesian model, and assume it admits the density $f_{\theta,y}$ with respect to a product measure $\lambda$.
Then for every $\gamma\in Y$, the posterior distribution satisfies
\[
\pi_{\theta\given y}(\.\given \gamma) = \argmin_{\bb{q}_\theta \in \c{M}_1(\Theta)} D_{\f{KL}}(\bb{q}_\theta \from \pi_\theta) - \operatorname*{\E}_{\vartheta\~\bb{q}_\theta} \ln f_{y\given\theta}(\gamma\given\vartheta)
\]
where $D_{\f{KL}}$ is the Kullback--Leibler divergence between probability measures, and the minima does not depend on the choice of $\lambda$.
\end{proposition}
\begin{proof}
Write the quantity being minimized as
\[
D_{\f{KL}}(\bb{q}_\theta \from \pi_\theta) - \operatorname*{\E}_{\vartheta\~\bb{q}_\theta} \ln f_{y\given\theta}(\gamma\given\vartheta) &= \operatorname*{\E}_{\vartheta\~\bb{q}_\theta} \ln \frac{f_{\bb{q}}(\vartheta)}{f_\theta(\vartheta) f_{y\given\theta}(\gamma\given\vartheta)}
\\
&= \operatorname*{\E}_{\vartheta\~\bb{q}_\theta} \ln \frac{f_{\bb{q}}(\vartheta)}{f_{\theta\given y}(\vartheta\given\gamma) f_y(\gamma)}
\\
&= D_{\f{KL}}(\bb{q}_\theta \from \pi_{\theta\given y}(\.\given \gamma)) - \ln f_y(\gamma)
\]
Since $\ln f_y(\gamma)$ does not depend on $\bb{q}_\theta$, and since $D_{\f{KL}}(\mu \from \nu) = 0$ implies $\mu = \nu$, we conclude that the objective is minimized by taking $\bb{q}_\theta = \pi_{\theta\given y}(\.\given\gamma)$.
Since the minima does not depend on the choice of the measure $\lambda$ with respect to which the densities are defined, the claim follows.
\end{proof}
For any given dataset, this result shows that Bayes' Rule can be viewed in \emph{information-theoretic} manner: among all probability measures, the posterior maximizes predictive power, while retaining as many bits as possible from the prior, in the sense given by the Kullback--Leibler divergence.
It's also possible to prove the result in a different way using the calculus of variations, with variations taken in the Banach space of signed measures under the total variation norm.
That argument also employs Bayes' Rule for densities for its key step, making it largely similar in spirit.
The result suggests a way to approximately compute posterior distributions: restrict the optimization problem to a suitably chosen subspace of the space of all probability measures $\c{M}_1(\Theta)$.
This strategy will be particularly fruitful in the later-described setting of Gaussian processes, where techniques for constructing such approximations with well-understood and favorable accuracy will be considered.
We \emph{always} view variational approximations as minimization of Kullback--Leibler divergences, as this is mathematically sound.
We will eschew standard presentations involving \emph{evidence lower bounds}: the only widely-known mathematical explanations for why maximizing these bounds should improve model performance appeal to Kullback--Leibler divergences---so, then, why talk about evidence lower bounds in the first place?
Once a posterior is calculated, the next step is to extract the relevant quantities from it.
Traditionally, this is often done by displaying summary statistics to be evaluated and interpreted by a person with statistical training, often with a focus on assessing uncertainty.
We will instead focus on settings where the posterior is given as input to an upstream decision-making algorithm: we explore these next.
\subsection{Technical lemmas}
We now prove a number of technical lemmas used in the preceding text, which for completeness are presented here in order to avoid disrupting the reader's flow.
\begin{lemma}
\label{lem:rcrv-rvm-equiv}
Let $b:\Omega\x A\->B$ be a jointly measurable stochastic process.
Then the map $b_*\P : \c{B} \x A \-> \R$, where the pushforward is taken in the first argument, is a probability kernel.
\end{lemma}
\begin{proof}
It is clear that $b_* \P$ is a probability measure for all $a'\in A$, so we only need to prove that the map $a \|> (b(\.,a)_*\P)(A_b)$ is measurable for all $A_b\in\c{B}$.
First, write
\[
(b_*\P)(A_b) = \P(b(\.,a)^{-1}(A_b)) = \int_\Omega \1_{b(\omega,a)\in A_b} \d\bb{P}(\omega)
.
\]
Now, note that the map $\Omega\x A\->\R$ given by $(\omega,a) \|> \1_{b(\omega,a)\in A_b}$ is bounded measurable, since $b$ is measurable in both arguments, and indicators of measurable functions on measurable sets are bounded measurable.
Finally, since for any bounded measurable $f : \Omega \x A \-> \R$, the map $a \|> \int_\Omega f(\omega,a) \d\bb{P}(\omega)$ is measurable by Fubini's Theorem, the claim follows.
\end{proof}
\begin{lemma}
\label{lem:semi-algebra}
Let $\c{S}$ be a \emph{semi-algebra of sets}, which is defined as a family of sets satisfying the following conditions.
\1 $\c{S}$ contains the empty set.
\2 $\c{S}$ is closed under finite intersections.
\3 The complement of any set in $\c{S}$ can be written as a finite union of disjoint sets in $\c{S}$.
\0
Then any bounded countably additive non-negative function $\mu : \c{S} \-> \R$ satisfying $\mu(\emptyset) = 0$ extends uniquely to a measure defined on the smallest $\sigma$-algebra containing $\c{S}$.
\end{lemma}
\begin{proof}
Define the \emph{algebra of sets} generated by $\c{S}$ to be
\[
a(\c{S}) = \cbr{\U_{i=1}^n S_i : S_i \in \c{S} \t{disjoint}}
\]
and note that the smallest $\sigma$-algebra generated by $\c{S}$ obviously coincides with the smallest $\sigma$-algebra generated by $a(\c{S})$.
We claim every function $\mu : \c{S} \-> \R$ satisfying the given assumptions extends uniquely to a function on $a(\c{S})$.
Every element of $a(\c{S})$ can be written as a finite union of disjoint sets---using this, define
\[
\mu^{(a)} : a(\c{S}) &\-> \R
&
\mu^{(a)}\del{\U_{i=1}^n S_i} &= \sum_{i=1}^n \mu(S_i)
.
\]
To ensure this is well-defined, we check that the definition is independent of the choice of which disjoint subsets to take the union of---suppose that $\U_{i=1}^n S_i = \U_{j=1}^m T_j$.
Then we have $S_i = \U_{j=1}^m S_i \^ T_j$ and similarly $T_j = \U_{i=1}^n S_i \^ T_j$: plugging these in to $\mu$ yields a double sum of disjoint sets and affirms well-definedness.
Next, note that $\mu^{(a)}(\emptyset) = \mu(\emptyset) = 0$.
To see that $\mu$ inherits countable additivity, take a sequence $S_1,S_2,.. \in a(\c{S})$ and note that
\[
\U_{n=1}^\infty S_n = \U_{n=1}^\infty \U_{m=1}^{p_n} T_{nm}
\]
where $T_{nm} \in \c{S}$ and $p_n \in \N$.
Plugging this into $\mu^{(a)}$ and applying countable additivity of $\mu$ gives the desired property.
We have thus obtained a uniquely defined countably additive function $\mu^{(a)} : a(\c{S}) \-> \R$ defined on an algebra of sets $a(\c{S})$ which extends $\mu$.
This function satisfies the assumptions of Carathéodory's Extension Theorem---see \textcite[Theorem A1.1]{kallenberg06} or \textcite[Theorem 3.1]{billingsley08}---applying this result gives the claim, where we note that uniqueness follows since the range of $\mu^{(a)}$ is bounded.
\end{proof}
\begin{lemma}
\label{lem:cyl-prod}
Let $\pi : \c{A} \x \c{B} \-> \R$ be a countably additive function with $\pi(\emptyset) = 0$.
Then $\pi$ extends uniquely to a measure on the product $\sigma$-algebra.
\end{lemma}
\begin{proof}
By \Cref{lem:semi-algebra}, it suffices to show that $\c{A} \x \c{B}$ is a semi-algebra of sets.
The first required property is immediate, the second and third properties follow by $(A \x B) \^ (A' \x B') = (A \^ A') \x (B \^ B')$ and $(A \x B)^c = (A^c \x B) \u (A \x B^c) \u (A^c \x B^c)$.
The claim follows.
\end{proof}
\begin{lemma}
\label{lem:ref-density}
Let $\pi_{a,b}$ be a measure which admits the density $f_{a,b}$ with respect to a product measure $\lambda_{a,b}$.
Then $\pi_a = \pi_{a,b}(\.\x B)$ admits the density $f_a$ with respect to $\lambda_a = \lambda_{a,b}(\.\x B)$, and similarly in the other argument.
\end{lemma}
\begin{proof}
By assumption and Tonelli's Theorem, we have
\[
\pi_a(A_a) &= \pi_{a,b}(A_a \x B)
\\
&= \int_{A_a \x B} f_{a,b}(\alpha,\beta) \d\lambda_{a,b}(\alpha,\beta)
\\
&= \int_{A_a} \int_B f_{a,b}(\alpha,\beta) \d\lambda_b(\beta)\d\lambda_a(\alpha)
\\
&= \int_{A_a} f_a(\alpha) \d\lambda_a(\alpha)
\]
where the desired probability density is $f_a(\alpha) = \int_B f_{a,b}(\alpha,\beta) \d\lambda_b(\beta)$.
\end{proof}
\section{Statistical decision-making}
We now use the Bayesian formalism to construct a probabilistic theory of decision-making.
To begin, we formalize the very general concept of an abstract agent making decisions in an environment in pursuit of some goal.
We use \textcite{sutton18,bertsekas19} as our references.
\subsection{Markov decision processes}
A Markov decision process is a stochastic system consisting of a set of states, actions, and transitions between states, together with a rewards that vary depending on states and actions.
This is defined as follows.
\label{ntn:mdp}
\begin{definition}[Discrete-time Markov decision process]
A \emph{discrete-time Markov decision process} is a $4$-tuple consisting of the following.
\1 State space: a measurable space $S$.
\2 Action space: a measurable space $A$.
\3 Reward: a probability kernel $r : \c{B}(\R) \x S \x A \-> \R$.
\4 Transition kernel: a probability kernel $p : \c{S} \x S \x A \-> \R$.
\0
\end{definition}
This is a very broad notion---however, a number of variations are also possible.
For instance, one can consider continuous-time, purely deterministic, and partially observed analogs---each of these involve their own subtleties and deserve study in their own right, but we do not pursue them here.
\begin{figure}
\tikzset{external/export next=false}
\begin{tikzpicture}
\begin{scope}[shift={(-1.5,2.125)}, xscale=-0.0085, yscale=0.0085, thick]
\clip[preaction = {draw,ultra thick}] (300, -395) -- (285, -355) .. controls (285, -355) and (269.142, -342.071) .. (255, -335) .. controls (245, -330) and (227.237, -322.04) .. (215, -320) .. controls (185, -315) and (175, -295) .. (175, -295) .. controls (175, -295) and (150, -295) .. (135, -285) .. controls (129.117, -281.078) and (100, -250) .. (100, -235) .. controls (100, -215) and (105, -205) .. (115, -195) .. controls (135.274, -174.726) and (144.617, -159.069) .. (186.751, -142.257) .. controls (230, -125) and (245, -125.009) .. (255, -125) .. controls (270.117, -124.987) and (289.478, -127.154) .. (305.786, -129.965) .. controls (335, -135) and (347.49, -144.511) .. (364.446, -158.01) .. controls (378.439, -169.15) and (389.863, -183.998) .. (397.853, -200) .. controls (405.048, -214.411) and (409.778, -230.664) .. (410, -246.769) .. controls (410.144, -257.217) and (410.264, -270.263) .. (402.959, -277.316) .. controls (395, -285) and (385, -290) .. (385, -290) .. controls (385, -290) and (377.69, -310.274) .. (375.386, -315.536) .. controls (371.018, -325.511) and (360.595, -330.936) .. (350.937, -334.628) .. controls (340.799, -338.503) and (318.451, -336.817) .. (318.451, -336.817) -- (326.08, -380.928) -- cycle;
\draw (300, -400) -- (300, -275);
\draw (275, -350) -- (275, -250) -- (225, -250);
\draw (250, -350) -- (250, -275);
\draw (225, -325) -- (225, -300) -- (200, -275) -- (100, -275);
\draw (325, -350) -- (325, -250) -- (300, -250);
\draw (400, -300) -- (350, -300) -- (350, -275) -- (375, -250) -- (375, -200) -- (400, -200);
\draw (350, -250) -- (350, -225) -- (250, -225) -- (250, -175) -- (225, -150) -- (150, -150);
\draw (350, -200) -- (300, -200) -- (300, -175);
\draw (400, -175) -- (325, -175) -- (300, -150) -- (300, -125);
\draw (200, -250) -- (150, -250) -- (125, -225) -- (75, -225);
\draw (175, -200) -- (225, -200) -- (225, -225) -- (175, -225);
\draw (225, -175) -- (150, -175) -- (150, -200) -- (100, -200);
\draw (275, -200) -- (275, -125);
\draw[fill=white] (200, -250) circle (7.5);
\draw[fill=white] (225, -250) circle (7.5);
\draw[fill=white] (250, -275) circle (7.5);
\draw[fill=white] (300, -275) circle (7.5);
\draw[fill=white] (300, -250) circle (7.5);
\draw[fill=white] (350, -250) circle (7.5);
\draw[fill=white] (350, -200) circle (7.5);
\draw[fill=white] (175, -225) circle (7.5);
\draw[fill=white] (175, -200) circle (7.5);
\draw[fill=white] (225, -175) circle (7.5);
\draw[fill=white] (300, -175) circle (7.5);
\draw[fill=white] (275, -200) circle (7.5);
\end{scope}
\begin{scope}[shift={(2.5,-1)}, scale=0.0125, very thick, line cap=round]
\draw (2, 2) -- (66, 2);
\draw (82, 2) -- (162, 2);
\draw (50, 18) -- (114, 18);
\draw (146, 18) -- (162, 18);
\draw (18, 34) -- (50, 34);
\draw (98, 34) -- (114, 34);
\draw (2, 50) -- (34, 50);
\draw (50, 50) -- (66, 50);
\draw (114, 50) -- (146, 50);
\draw (18, 66) -- (34, 66);
\draw (66, 66) -- (114, 66);
\draw (130, 66) -- (146, 66);
\draw (34, 82) -- (66, 82);
\draw (98, 82) -- (130, 82);
\draw (2, 98) -- (18, 98);
\draw (34, 98) -- (50, 98);
\draw (82, 98) -- (98, 98);
\draw (146, 98) -- (162, 98);
\draw (50, 114) -- (66, 114);
\draw (82, 114) -- (98, 114);
\draw (2, 130) -- (18, 130);
\draw (34, 130) -- (50, 130);
\draw (98, 130) -- (114, 130);
\draw (130, 130) -- (146, 130);
\draw (18, 146) -- (34, 146);
\draw (50, 146) -- (66, 146);
\draw (114, 146) -- (130, 146);
\draw (146, 146) -- (162, 146);
\draw (2, 162) -- (82, 162);
\draw (98, 162) -- (162, 162);
\draw (2, 2) -- (2, 162);
\draw (18, 18) -- (18, 34);
\draw (18, 66) -- (18, 82);
\draw (18, 98) -- (18, 114);
\draw (18, 130) -- (18, 146);
\draw (34, 2) -- (34, 18);
\draw (34, 50) -- (34, 66);
\draw (34, 82) -- (34, 114);
\draw (50, 18) -- (50, 34);
\draw (50, 50) -- (50, 82);
\draw (50, 114) -- (50, 130);
\draw (50, 146) -- (50, 162);
\draw (66, 34) -- (66, 66);
\draw (66, 82) -- (66, 114);
\draw (66, 130) -- (66, 146);
\draw (82, 18) -- (82, 50);
\draw (82, 66) -- (82, 98);
\draw (82, 114) -- (82, 162);
\draw (98, 34) -- (98, 66);
\draw (98, 98) -- (98, 114);
\draw (98, 130) -- (98, 162);
\draw (114, 18) -- (114, 34);
\draw (114, 82) -- (114, 130);
\draw (130, 2) -- (130, 34);
\draw (130, 50) -- (130, 146);
\draw (146, 34) -- (146, 50);
\draw (146, 82) -- (146, 98);
\draw (146, 114) -- (146, 130);
\draw (162, 2) -- (162, 162);
\end{scope}
\node[minimum size=85] at (-3.5,0) (a) {};
\node[minimum size=85] at (3.5,0) (e) {};
\node[anchor=north] at (a.south) {Agent};
\node[anchor=north] at (e.south) {Environment};
\draw[-latex,very thick] (a) to[bend left=22.5] node[midway, above] {action} (e);
\draw[-latex,very thick] (e) to[bend left=22.5] node[midway, below] {reward, next state} (a);
\end{tikzpicture}
\caption[Markov decision processes]{Illustration of the feedback loop induced by a Markov decision process. Here, the agent chooses an action, which results in the environment changing to a new state. The agent observes the new state, as well as the reward given by the previous state and action. The agent's goal is to select actions to maximize long-term rewards.}
\label{fig:mdp}
\end{figure}
The idea behind this definition is that, at every point in time, the agent observes the current state, chooses an action $a \in A$, and obtains another state $s' \~ p(s\given a)$.
\Cref{fig:mdp} illustrates this.
Note that \emph{time} can, and often will, be part of the state $s$.
The agent's goal is to choose actions so as to control the full trajectory of states in order to obtain maximum rewards.
The choice of action in every state is called a \emph{policy}, and is formalized as follows.
\label{ntn:mdp-policy}
\begin{definition}[Policy]
Define the following.
\1 A measurable function $\pi : S \-> A$ is called a \emph{deterministic policy}.
\2 A probability kernel $\pi : \c{A} \x S \-> \R$ is called a \emph{Markov policy}.
\0
\end{definition}
Markov policies include deterministic policies as a special case, by taking the conditional distributions to be Dirac and re-interpreting the given expressions appropriately.
As with Markov decision processes, one can also consider more general classes of policies, but we do not do so here.
Different policies yield different state trajectories, and therefore different rewards.
Of these, some obtain more rewards than others.
We therefore introduce notions for distinguishing between policies according to the rewards they obtain.
\label{ntn:mdp-value}
\begin{definition}[Optimal policy]
Let $T \in \N$ be the \emph{time horizon}.
A policy is called \emph{optimal} if it maximizes the \emph{value function}
\[
V^{(\pi)}(s_0) = \E \sum_{t=0}^T r_t
\]
where $r_t \given s_t,a_t \~ r(s_t,a_t)$, $a_t \given s_t \~ \pi(s_t)$, and $s_{t+1} \given s_t,a_t \~ p(s_t, a_t)$.
\end{definition}
Finding an optimal policy therefore amounts to selecting the best possible actions to maximize expected total rewards.
Note that closely-related alternative notions of optimality, such as minimizing infinite discounted sums, or limits of averages, are also possible.
The most important distinction between different decision problems for finding optimal policies is given by what is assumed known.
\1 If $p$ and $r$ are known, we say we have an \emph{optimal control} problem.
\2 Otherwise, we say we have a \emph{reinforcement learning} problem.
\0
These classes differ fundamentally from one another.
Optimal control problems can be viewed as a class of structured optimization problems, where the goal is to compute $\pi$ by evaluating $r$ and $p$ as necessary.
Here, one generally proceeds by proving that $V^{(\pi)}$ and $\pi$ satisfy certain recursive equations, and developing schemes for solving them directly.
Reinforcement learning problems are more complex.
Due to the rewards or dynamics being \emph{unknown}, they cannot simply be maximized and their expectation must be \emph{learned}.
This forces one to consider whether to take advantage of actions known to be good, or to try others in case they might be better---this is known as the \emph{explore-exploit tradeoff}.
For such settings, we need an appropriate solution concept---to obtain one, define the following.
\label{ntn:mdp-regret}
\begin{definition}[Regret]
The \emph{regret} of a policy $\pi$ is defined as
\[
R^{(\pi)}(s_0) = V^*(s_0) - V^{(\pi)}(s_0)
\]
provided that the value function $V^*$ with respect to an optimal policy exists.
\end{definition}
Minimizing regret is equivalent to maximizing value, but when $p$ and $r$ are unknown doing so directly is impossible.
Instead, the goal is to find an \emph{algorithm}---that is, a way of updating the policy based on observed data---so as to limit growth of regret.
In most settings, one can prove that every algorithm which does not know $p$ and $r$ necessarily incurs some level of regret.
This is done by exhibiting a randomized set of problems and rewards over which regret is lower-bounded in expectation for any algorithm.
In such a class, actions that perform well on one problem will necessarily perform badly on another problem.
Such arguments show that some degree of regret is inevitable.
On the other hand, some algorithms incur more regret than others.
The obviously-bad algorithm which learns nothing and chooses the exact same action over and over again incurs at most linear regret.
An algorithm is said to \emph{solve} a decision problem if its asymptotic regret rate with respect to $T$ matches the respective regret lower bound in the given problem class.
Finding such algorithms is of key interest.
One way to construct algorithms for solving decision problems is to employ a \emph{model-based} approach, which loosely speaking works as follows.
\1 Learn the unknown transitions and/or rewards from observed data using a supervised learning approach.
\2 Use the learned model(s) to find a policy satisfying some criteria.
\0
The details of such approaches depend on the setting.
We distinguish between two key kinds of reinforcement learning problems.
\1 If $|S| = 1$, we say we have a \emphmarginnote{multi-armed bandit} problem.
\2 Otherwise, we say we have a general reinforcement learning problem.
\0
Multi-armed bandits possess no variable state $S$, and thus only require one to learn the rewards to determine optimal actions.
In general reinforcement learning, actions can influence transitions between states---these problems require long-term planning, making them much more general, difficult, and important.
One can argue that as a mathematical theory, reinforcement learning is powerful enough to describe many aspects of human and animal intelligence, making it fundamentally interesting to study and develop.
We now restrict ourselves to the bandit setting, which is substantially easier to study and so can be understood more deeply.
Here, even when $A$ is a finite set and the rewards are Gaussian, the model-based approach consisting of (i) estimating rewards using empirical risk minimization and (ii) choosing the policy which maximizes rewards is known to be non-optimal.
This approach fails to explore, and can get stuck chasing inferior rewards.
One way to fix this problem is to replace empirical risk minimization with Bayesian learning, and adopt an appropriate decision rule for selecting actions.
Doing this yields approaches which can be shown optimal in many settings.
We therefore proceed to study multi-armed bandits in more detail.
\subsection{Multi-armed bandits}
The multi-armed bandit problem takes its name from a casino analogy.
In the 1950s, slot machines often had levers one could pull instead of buttons one could press, and were called \emph{one-armed bandits} for their ability to empty gamblers' wallets.
A \emph{multi-armed bandit} is a slot machine, which, for a fixed cost, allows one to pull an arm $x \in X$ and receive a random reward whose distribution depends on $x$.
The goal is to minimize expected loss, or, equivalently, maximize total expected rewards.
Multi-armed bandits can be viewed as discrete-time Markov decision processes with a one-element state space, but this is not necessarily the most fruitful way to think about them.
We thus begin by introducing formalism and notation better suited to the given setting, which can be viewed as a special case of the notions considered previously.
We use \textcite{slivkins19,lattimore20} as references.
\label{ntn:mab}
\begin{definition}[Multi-armed bandit]
Let $f : X \-> \R$ be a bounded above measurable function, let $\eps : \Omega \x X \-> \R$ be a jointly measurable stochastic process such that $\E(\eps(x)) = 0$ for all $x$, and let $y(\omega,x) = f(x) + \eps(\omega,x)$.
Define the following.