Repository navigation
Expand file tree
/
Copy pathcontinuation_geometry_monograph.tex
More file actions
2943 lines (2631 loc) · 144 KB
/
Copy pathcontinuation_geometry_monograph.tex
File metadata and controls
2943 lines (2631 loc) · 144 KB
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]{report}
\usepackage{fontspec}
\setmainfont{TeX Gyre Pagella}
\usepackage[margin=1.15in]{geometry}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{enumitem}
\usepackage{array}
\usepackage{hyperref}
\hypersetup{colorlinks=true, linkcolor=black, urlcolor=black, citecolor=black}
\newtheorem{proposition}{Proposition}[chapter]
\newtheorem{theorem}{Theorem}[chapter]
\newtheorem{lemma}{Lemma}[chapter]
\newtheorem{definition}{Definition}[chapter]
\newtheorem{remark}{Remark}[chapter]
\newtheorem{corollary}{Corollary}[chapter]
\newtheorem{example}{Example}[chapter]
\newtheorem{openproblem}{Open Problem}[chapter]
\newtheorem{caveat}{Caveat}[chapter]
\title{\textbf{Continuation Geometry}}
\author{Flyxion}
\date{July 2026}
\begin{document}
\maketitle
\tableofcontents
\chapter{Why Continuation Matters}
A system that has stopped changing has not necessarily stopped existing,
but a system that has run out of ways to change has. There is a
difference between a room that is quiet because nothing needs to happen
in it and a room that is quiet because nothing can. The first is rest.
The second is the end of a story that may not announce itself as an
ending. Most of what concerns us about a living body, a piece of
software, an institution, or a plan is not its condition at the present
instant but the width of what it can still become. A body in a coma and
a body in deep sleep can look alike from across the room; what
distinguishes them is not visible in a single frame but in the shape of
what still lies ahead of each.
This is a book about that shape. What a system could still do matters
more than what it is doing right now: the whole reachable fan of
futures opening out from wherever it happens to stand. Call that fan a
system's futures, or its options, or its story-space. Whatever you
call it, it's the thing that shrinks when a hinge rusts shut, when a
species loses a gene it won't get back, when an argument commits
itself to a position it can't walk back from, when a bridge corrodes
past the point where any repair schedule could save it. Repair, in the
most general sense of the word, tries to restore it: not a particular
outcome, but the standing
possibility of many outcomes.
Several habits of thought already study pieces of this. Engineers ask
whether a structure will hold under load. Ecologists ask whether a
population can recover from a disturbance. Economists ask whether an
allocation can be sustained, and control theorists ask whether a
system can be steered to stay within safe bounds indefinitely. Each of
these is really the same question in different clothes: whether the
fan of futures remains nonempty, whether there is still at least one
way forward that respects whatever constraints the situation imposes.
That question has a mature and exact mathematical answer, developed
over several decades under the name of viability theory, and this book
leans on that answer throughout.
But existence is not the only thing worth asking about a fan of futures.
Two systems can both have somewhere left to go and still differ enormously
in how much room they have to maneuver, how easily they could be nudged
back from the edge, which of their internal distinctions survive into
their futures and which quietly get erased, and which of their many
possible continuations is the one actually worth choosing. A patient who
can still recover and a patient who can barely still recover are both,
technically, viable; no doctor would treat that technicality as the end
of the conversation. The questions that begin where mere existence leaves off, how much
room, how easily recovered, what is preserved, what is worth choosing,
are what this book is organized around. They
do not yet have as settled a mathematical treatment as the existence
question does. Building one, on top of the existence question rather than
instead of it, is the project.
The order of concerns runs roughly as follows, and will recur throughout
the book. A system draws distinctions: it treats some states as
meaningfully different from others. Those distinctions are carried by
information, and information degrades. This is what entropy names.
Some of that degradation can be undone; the undoing is repair. Repair is
only possible where a future still exists to be repaired into; that
existence is continuation. And among the continuations that exist, some
are better than others by standards that have nothing to do with mere
existence; sorting them out is admissibility. Existence, in other words,
is necessary for everything downstream of it but is far from sufficient
for any of it. This book starts at existence, because it is the
best-understood link in the chain, and spends the rest of its length
building outward from it toward repair, distinction, and admissibility:
the concerns that motivated picking up the chain in the first place.
\section*{The object of study}
We now fix, once and for all, the formal object the rest of the book is
about.
\begin{definition}[State, dynamics, constraint]
Let $X$ be a state space (a metric space; in applications typically
$\mathbb{R}^n$ or a finite-dimensional manifold). Let
$F : X \rightrightarrows X$ be a set-valued map assigning to each state
$x \in X$ a set $F(x) \subseteq X$ of admissible velocities, so that
trajectories of the system are absolutely continuous curves
$\gamma$ satisfying the differential inclusion
$\dot\gamma(t) \in F(\gamma(t))$. Let $K \subseteq X$ be closed; $K$ is
the constraint set, the region within which the system is required to
remain.
\end{definition}
\begin{definition}[Continuation space]
For $x \in K$, the continuation space of $x$ is
\[
\Gamma_K(x) = \{\gamma : [0,\infty) \to K \mid \gamma(0) = x,\ \dot\gamma(t) \in F(\gamma(t))\ \text{for a.e.\ } t \geq 0\}.
\]
For a finite horizon $T > 0$, the finite-horizon continuation space is
\[
\Gamma_K^T(x) = \{\gamma : [0,T] \to K \mid \gamma(0) = x,\ \dot\gamma(t) \in F(\gamma(t))\ \text{for a.e.\ } t \in [0,T]\}.
\]
\end{definition}
$\Gamma_K(x)$ (or simply $\Gamma(x)$ where $K$ is fixed by context) is
the fan of futures referred to above, made precise relative to one
particular reading of ``possible'': a future is possible if it is a
trajectory of the given dynamics that never leaves the given constraint
set. Other readings of possible are available and will matter later:
controllably possible, admissibly possible, epistemically possible. But
this dynamical reading is where the book begins, because it is the one
with an existing, exact theory attached to it.
\begin{proposition}[Restriction is well-defined]
For every $T > 0$, restriction of a trajectory to $[0,T]$ defines a map
\[
\rho_T : \Gamma_K(x) \to \Gamma_K^T(x), \qquad \rho_T(\gamma) = \gamma|_{[0,T]}.
\]
\end{proposition}
\begin{proof}
If $\gamma \in \Gamma_K(x)$ then $\gamma(0) = x$, $\gamma(t) \in K$ for
all $t \geq 0$, and $\dot\gamma(t) \in F(\gamma(t))$ for a.e.\ $t \geq 0$.
Restricting to $[0,T]$ preserves each of these three conditions on the
smaller domain, so $\gamma|_{[0,T]} \in \Gamma_K^T(x)$.
\end{proof}
\begin{proposition}[Consistency across horizons]
For $0 < T_1 < T_2$, restriction further to $[0,T_1]$ defines a map
$\rho_{T_1, T_2} : \Gamma_K^{T_2}(x) \to \Gamma_K^{T_1}(x)$, and the
diagram of restriction maps commutes:
$\rho_{T_1} = \rho_{T_1,T_2} \circ \rho_{T_2}$ for every
$T_1 < T_2 < \infty$.
\end{proposition}
\begin{proof}
Immediate from the definition of restriction: restricting a curve to
$[0,T_1]$ gives the same result whether it is done directly from the
domain $[0,\infty)$ or via an intermediate restriction to $[0,T_2]$,
since restriction of a function to a subset of its domain depends only on
that subset, not on the route by which the subset was reached.
\end{proof}
\begin{remark}
This consistency is what allows $\Gamma_K(x)$ to be recovered, in good
cases, as the projective (inverse) limit of the finite-horizon spaces
$\Gamma_K^T(x)$ as $T \to \infty$. This observation is recorded here
because Chapter 6 will build continuation volume and continuation
distance on the finite-horizon spaces precisely because they are
well-posed where the infinite-horizon trajectory space is not; the
present proposition is what justifies treating the finite-horizon
construction as an approximation to $\Gamma_K(x)$ rather than a
different object entirely.
\end{remark}
\begin{proposition}[Equilibria have nonempty continuation spaces]
If $x^\ast \in K$ satisfies $0 \in F(x^\ast)$, then $\Gamma_K(x^\ast) \neq \varnothing$.
\end{proposition}
\begin{proof}
Let $\gamma(t) = x^\ast$ for all $t \geq 0$. Then $\gamma(0) = x^\ast$,
$\gamma(t) = x^\ast \in K$ for all $t$, and $\dot\gamma(t) = 0 \in F(x^\ast) = F(\gamma(t))$
for all $t$. Hence $\gamma \in \Gamma_K(x^\ast)$.
\end{proof}
\begin{remark}
This is the first, and simplest, example of a nonempty continuation
space, and it is worth pausing on because it already illustrates a
distinction the rest of the book will not let go of: nonemptiness of
$\Gamma_K(x^\ast)$ here is witnessed by a single, unmoving trajectory. It
says nothing about how many other continuations $x^\ast$ has, how close
$x^\ast$ sits to states with no continuations at all, or whether the
distinctions the system cares about survive along the witnessing
trajectory. A full characterization of nonemptiness for arbitrary
$x \in K$, not just equilibria, is the content of Aubin's Viability
Theorem, taken up next.
\end{remark}
\section*{The progression this book follows}
The chain of concerns described in natural language above, distinction,
information, entropy, repair, continuation, admissibility, is not a
theorem and is not offered as one. It is a roadmap, and its purpose is to
say in advance which part of the whole problem each part of the book is
responsible for.
Continuation, in the sense of $\Gamma_K(x) \neq \varnothing$, is the
best-understood link and is addressed first: it is the subject of
viability theory, surveyed and put to use in the chapters immediately
following this one. Repair is addressed once continuation is in hand,
since a repair is, formally, nothing more than a continuation that
reaches a designated target: the machinery for saying this precisely
is capture, and it costs nothing beyond what continuation has already
supplied. Distinction and admissibility are addressed last, and are the
hardest part of the roadmap to formalize, because they require deciding
not merely whether a future exists but which of the many existing
futures should be preferred, and by what standard. The book's later
chapters build that standard in stages rather than assuming it in
advance.
\chapter{Precursors}
No idea arrives without ancestors, and an idea that pretends otherwise
usually turns out to be a poorer version of something already on the
shelf. The temptation for any project that finally finds its central
object is to treat the discovery as an arrival rather than a
continuation of its own, to write the history chapter last, briefly,
and mostly as a courtesy. That temptation is worth resisting here for a
specific reason: the people who spent real years on the questions this
book asks were not working on one shared problem that only now gets
its proper name. They were working on at least two different problems
that happen, from a certain distance, to look like the same problem.
The first tradition asks how a system manages to keep going. Given
where a system stands and how it is allowed to move, will it still be
able to move tomorrow, and the day after, without breaking whatever
rule keeps it a functioning version of itself? This is a question about
persistence, and it has been asked with increasing precision for over a
century: orbits that never quite settle down, equilibria that recover
from being nudged, decisions that must be made in sequence without
knowing the whole future in advance. Each advance in this tradition
sharpened the same underlying question rather than replacing it, which
is why it reads, in retrospect, like a single conversation carried on
by different people.
The second tradition asks a different question, one that does not
reduce to the first even though it sounds similar. Given where a system
stands, what does the whole space of things it could still become look
like? Not whether that space is empty, but what shape it has, how it
grows or shrinks as choices get made, and which of its features are
worth preserving. This tradition is younger, more scattered across
disciplines that rarely cite one another, and considerably less tidy.
Evolving biological form. What innovations become available once other
innovations have already happened. Causal alternatives to what
actually occurred. Even an architect's attempt to say which building
layouts leave a neighborhood room to keep being a good place to live.
None of these projects thought of itself as contributing to a general
theory of possibility-spaces. Read together, they look like exactly
that.
This book sits where these two traditions meet, and owes each of them a
debt it should not obscure. What follows makes some of that debt
precise, showing in each case how much of the earlier work can be
recovered as a special case or a direct ancestor of the object this
book is built around, and where each earlier tradition stops short.
\section*{The persistence lineage}
\begin{definition}[Recurrent state, after Poincar\'e]
Let $\phi_t : X \to X$ be a measure-preserving flow on a probability
space $(X,\mu)$, and let $A \subseteq X$ have $\mu(A) > 0$. A point
$x \in A$ is recurrent in $A$ if $\phi_t(x) \in A$ for some $t > 0$,
and infinitely often for a.e.\ such $x$.
\end{definition}
\begin{proposition}[Poincar\'e recurrence and nonempty continuation]
If $A$ is invariant under $\phi_t$ in the sense that $\phi_t^{-1}(A)$
has full relative measure in $A$ for all $t$, then for a.e.\ $x \in A$,
the trajectory $\gamma(t) = \phi_t(x)$ lies in $\Gamma_A(x)$, and
moreover returns to $A$ infinitely often.
\end{proposition}
\begin{proof}
This is the Poincar\'e Recurrence Theorem, applied to the flow
restricted to $A$: almost every point of a measure-preserving invariant
set returns to that set infinitely often under forward iteration
\cite{poincare1890}. The trajectory $\gamma(t) = \phi_t(x)$ satisfies
$\gamma(0) = x$ and $\gamma(t) \in A$ for all $t \geq 0$ by invariance
of $A$, and $\dot\gamma(t)$ equals the generator of $\phi_t$ at
$\gamma(t)$, which lies in $F(\gamma(t))$ trivially when $F$ is taken to
be the vector field generating $\phi_t$. Hence $\gamma \in \Gamma_A(x)$.
\end{proof}
\begin{remark}
This is the earliest instance, in the lineage this book claims, of a
proof that a continuation space is nonempty, though Poincar\'e's
concern was not persistence under constraint but the much stronger fact
of eventual return. Recurrence is strictly stronger than mere viability:
every recurrent trajectory witnesses nonemptiness of $\Gamma_A(x)$, but
nonemptiness of $\Gamma_K(x)$ does not require any trajectory to return
to its starting neighborhood at all. The relation between recurrence and
viability, when a viable trajectory can be strengthened to a recurrent
one, is not pursued in this book but is flagged here as a legitimate
question left on the table by folding Poincar\'e into the same lineage
as Aubin.
\end{remark}
\begin{definition}[Lyapunov stability]
A state $x^\ast$ with $0 \in F(x^\ast)$ is Lyapunov stable if for every
$\varepsilon > 0$ there exists $\eta > 0$ such that
$\|x - x^\ast\| < \eta$ implies every $\gamma \in \Gamma_X(x)$ satisfies
$\|\gamma(t) - x^\ast\| < \varepsilon$ for all $t \geq 0$.
\end{definition}
\begin{proposition}[Lyapunov functions certify viability of sublevel sets]
Suppose $V : X \to \mathbb{R}_{\geq 0}$ is continuously differentiable,
$V(x^\ast) = 0$, $V(x) > 0$ for $x \neq x^\ast$, and
$\sup_{v \in F(x)} \nabla V(x) \cdot v \leq 0$ for all $x$. Then for
every $c > 0$, the sublevel set $K_c = \{x : V(x) \leq c\}$ is viable
under $F$: $\Gamma_{K_c}(x) \neq \varnothing$ for every $x \in K_c$.
\end{proposition}
\begin{proof}
Fix $x \in K_c$ and let $\gamma$ solve $\dot\gamma(t) \in F(\gamma(t))$,
$\gamma(0) = x$ (such a solution exists locally under standard
regularity hypotheses on $F$). Along $\gamma$,
$\frac{d}{dt} V(\gamma(t)) = \nabla V(\gamma(t)) \cdot \dot\gamma(t) \leq
\sup_{v \in F(\gamma(t))} \nabla V(\gamma(t)) \cdot v \leq 0$, so
$V(\gamma(t))$ is nonincreasing. Since $V(\gamma(0)) = V(x) \leq c$, it
follows that $V(\gamma(t)) \leq c$ for all $t$, i.e.\ $\gamma(t) \in K_c$
for all $t$. Hence $\gamma \in \Gamma_{K_c}(x)$.
\end{proof}
\begin{remark}
This is the classical Lyapunov stability argument \cite{lyapunov1892},
restated to make explicit what it already was: a direct construction of
a nonempty, in fact tangentially controlled, continuation space. The
tangency criterion of Aubin's Viability Theorem can be read as the
generalization of this proposition obtained by dropping the requirement
that a single global function $V$ certify viability for every sublevel
set at once, and instead checking the weaker, purely local condition
$F(x) \cap T_K(x) \neq \varnothing$ directly on the boundary of whatever
set $K$ is given, without needing $K$ to arise as a sublevel set of
anything.
\end{remark}
\begin{definition}[Bellman's principle of optimality]
For a controlled system with running cost $l$ and terminal cost $c$,
define the value function
\[
J(t,x) = \inf_{u(\cdot)} \left[ \int_t^T l(x(s), u(s))\, ds + c(x(T)) \right]
\]
over admissible controls $u(\cdot)$ steering $\dot x(s) = f(x(s), u(s))$
from $x(t) = x$.
\end{definition}
\begin{proposition}[Dynamic programming as a special case of guaranteed valuation]
Suppose $Q(x) = \{0\}$ (no perturbation) in Aubin's dynamic core
construction. Then Aubin's guaranteed valuation function $V^\sharp(T,x)$
reduces to Bellman's value function $J(t,x)$ under the correspondence
$T \leftrightarrow T - t$, and the Hamilton-Jacobi-Isaacs variational
inequality reduces to the Hamilton-Jacobi-Bellman equation
\[
-\frac{\partial J}{\partial t} + \inf_{u \in P(x)}\left[ \nabla J(t,x) \cdot f(x,u) + l(x,u) \right] = 0.
\]
\end{proposition}
\begin{proof}
When $Q(x) = \{0\}$, the supremum over perturbations $v \in Q(x)$ in the
defining formula for $V^\sharp$ (Aubin, Theorem 1.2) has a single term
to range over, so it drops out, leaving only the infimum over feedbacks
$p_e(x) \in P(x)$. This is exactly Bellman's infimum over controls, and
the discount term $m(x,p,v)$, in the absence of perturbation, plays the
role of a running cost once $l$ is identified with the appropriate
component of the functional $J_u$. Substituting into the
Hamilton-Jacobi-Isaacs inequality and dropping the now-vacuous
$\sup_{v \in Q(x)}$ term gives the stated Hamilton-Jacobi-Bellman
equation \cite{bellman1957}.
\end{proof}
\begin{remark}
Dynamic programming is, on this reading, the special case of guaranteed
valuation in which nature is not permitted to intervene. Aubin's
contribution to the persistence lineage was to reinstate the adversary
that Bellman's formulation sets to zero, without discarding the
optimization structure Bellman had already built. Aubin's viability
kernel and capture basin are the further generalization obtained by
dropping the requirement that a single scalar cost be optimized at all,
and asking only whether \emph{some} admissible control keeps the system
where it needs to be: existence in place of optimality, which is
where this book's own starting point sits.
\end{remark}
\section*{The possibility-space lineage}
\begin{definition}[Adjacent possible, after Kauffman]
Given a set of currently realized configurations $S_t \subseteq X$ and a
one-step generative map $\Phi : X \to 2^X$ (the set of configurations
directly reachable from a given one), the adjacent possible of $S_t$ is
\[
\mathrm{Adj}(S_t) = \bigcup_{x \in S_t} \Phi(x).
\]
\end{definition}
\begin{proposition}[Adjacent-possible growth as reachable-set recursion]
Define $S_{t+1} = S_t \cup \mathrm{Adj}(S_t)$. Then, taking
$\Phi$ as a discrete-time one-step reachability map and $t$ discrete
steps,
\[
S_t = \bigcup_{x \in S_0} R_t(x),
\]
where $R_t(x)$ is the discrete-time finite-horizon reachable set
generated by $\Phi$ from $x$.
\end{proposition}
\begin{proof}
By induction on $t$. The base case $t=0$ is immediate:
$R_0(x) = \{x\}$ and $S_0 = \bigcup_{x \in S_0} \{x\}$. For the
inductive step, assume $S_t = \bigcup_{x \in S_0} R_t(x)$. Then
\[
S_{t+1} = S_t \cup \mathrm{Adj}(S_t) = \bigcup_{x \in S_0} R_t(x) \cup \bigcup_{y \in S_t} \Phi(y) = \bigcup_{x \in S_0} \left( R_t(x) \cup \bigcup_{y \in R_t(x)} \Phi(y) \right) = \bigcup_{x \in S_0} R_{t+1}(x),
\]
using that $\Phi(y)$ for $y \in R_t(x)$ is exactly the one-step
extension of a $t$-step reachable trajectory from $x$, which is what
$R_{t+1}(x)$ collects.
\end{proof}
\begin{remark}
Kauffman's adjacent possible \cite{kauffman2000} is, under this
correspondence, an early and purely combinatorial instance of the
finite-horizon reachable-set construction that later chapters use to
build continuation volume. What it lacks, by design, is any constraint
set $K$: the adjacent possible is not asked to remain within bounds,
only to grow, so it corresponds to the unconstrained case $K = X$. It is
this unconstrained framing, uncommon in the persistence lineage but
common in Kauffman's and Holland's own writing \cite{holland1995}, that
marks the possibility-space lineage as a genuinely separate tradition
rather than a duplicate of the first.
\end{remark}
\begin{definition}[Structural causal model and intervention, after Pearl]
A structural causal model is a tuple $(V, U, \mathcal{F})$ of endogenous
variables $V$, exogenous variables $U$, and structural equations
$\mathcal{F}$, each $f_i \in \mathcal{F}$ determining $V_i$ as a function
of its parents and relevant exogenous noise. An intervention
$\mathrm{do}(V_j = v)$ replaces $f_j$ with the constant function $v$,
producing a modified model $(V, U, \mathcal{F}_{\mathrm{do}(V_j=v)})$.
\end{definition}
\begin{proposition}[Counterfactual reachability as continuation under intervened dynamics]
Let $F$ correspond to the structural equations $\mathcal{F}$ in the
sense that $F(x)$ encodes the admissible next-states consistent with
$\mathcal{F}$ from configuration $x$, and let $F_{\mathrm{do}(V_j=v)}$
be the corresponding map under the intervened model. Then the
counterfactual ``what would have happened to $x$ had $V_j$ been set to
$v$'' corresponds exactly to $\Gamma_K^{F_{\mathrm{do}(V_j=v)}}(x)$, the
continuation space of $x$ computed under the intervened dynamics rather
than the original one.
\end{proposition}
\begin{proof}
By construction: a counterfactual trajectory in Pearl's sense is
obtained by holding the exogenous noise $U$ fixed at its actual values
and re-evaluating the structural equations under the modified
$\mathcal{F}_{\mathrm{do}(V_j=v)}$ \cite{pearl2000}. This is precisely
the definition of a trajectory of the continuation space
$\Gamma_K^{F_{\mathrm{do}(V_j=v)}}(x)$: an admissible evolution from $x$
under the dynamics $F_{\mathrm{do}(V_j=v)}$ rather than $F$.
\end{proof}
\begin{remark}
This is the cleanest correspondence in the possibility-space lineage,
and it points at something later chapters will need directly: comparing
continuations across different choices of dynamics $F$, not merely
across different starting states $x$ under one fixed $F$. Every
construction introduced so far has been stated for a single fixed
$(F,K)$. Pearl's framework is a reminder that the more general
comparison, across interventions on $F$ itself, is both meaningful and
already partly worked out elsewhere; this book does not take it up in
full, but the treatment of multiple agents sharing a dynamics later on
is the nearest point of contact.
\end{remark}
\begin{remark}[Levin: morphological and bioelectric continuation]
Levin's work on bioelectric signaling as a substrate for the storage and
recall of anatomical target-morphologies \cite{levin2021} is included in
this lineage on different grounds than the constructions above: not
because it yields a theorem restatable in the $\Gamma_K(x)$ vocabulary
without loss, but because it is the clearest existing example of a
system whose continuation space genuinely depends on more than its
instantaneous state: a regenerating organism's available futures
depend on bioelectric pattern memory that persists through and beyond
local tissue damage, in a way not recoverable from cell-by-cell
instantaneous configuration alone. This is, in the vocabulary already
fixed, exactly the history-dependence flagged as an open problem for the
continuation-operator construction. No proposition is offered here
forcing Levin's biology into the present formalism; the honest claim is
narrower, that this is the strongest existing example motivating why
history-dependence needs to be taken seriously rather than assumed away.
\end{remark}
\begin{remark}[Alexander: pattern languages]
Christopher Alexander's pattern language \cite{alexander1977} is the
loosest member of this lineage and is kept loose deliberately. A
pattern, in Alexander's sense, is a reusable solution to a recurring
design problem, stated together with the conditions under which it
applies and the further patterns it tends to invite or foreclose. Read
architecturally rather than dynamically, a choice of pattern is a move
that changes which further patterns remain available, which is a
possibility-space claim in substance, even though Alexander never
formalized it as one and this book will not retroactively supply the
formalization on his behalf. What is worth taking from Alexander is not
a theorem but a discipline: that a possibility-space can be studied
through the patterns that compose it and the compatibility relations
between them, prior to and independent of any measure, metric, or
dynamics being placed on it. That discipline anticipates the
lattice-based treatment of admissibility taken up later, where
composability and compatibility, rather than a single scalar score, are
again the primary structure.
\end{remark}
\section*{Where the lineages meet}
The two lineages traced here answer different questions and were not, in
general, in conversation with one another. The persistence lineage
answers whether a future exists; its instruments, recurrence,
stability, dynamic programming, viability, get sharper across a
century of work without changing what they are fundamentally asking. The
possibility-space lineage answers what the space of futures looks like
once one exists; its instruments are more heterogeneous because the
question is less settled, and in at least two cases above (Levin,
Alexander) the correspondence to the present formalism is offered as
motivation rather than as an established equivalence. The chapters that
follow begin by finishing what the persistence lineage started: stating
Aubin's Viability Theorem properly and using it as the foundation
everything else stands on, before turning to the harder task of giving
the possibility-space lineage's questions the same kind of precision
the persistence lineage has already earned.
\chapter{Viability Theory}
Walk along the top of a sea wall and there is a moment, well before you
lose your footing, when the only directions left that keep you on the
wall are a narrower and narrower band around straight ahead. You do not
fall the instant that band closes to nothing; you fall the instant
after, when a gust or a stumble sends you in some direction the band no
longer contains. Whether you are in danger, in other words, is not
really a question about your current position. Position alone cannot
distinguish the middle of a wide, flat wall from its crumbling edge if
both happen to put your feet at the same height. What matters is
narrower than position: it is whether the directions physics is about
to hand you and the directions that keep you on the wall still overlap.
This is a more useful fact than it first appears, because it turns a
question about the entire rest of your walk (will I, at any point in
the future, stumble off this wall) into a question you can answer by
looking only at where your feet are right now and which way you could
step next. The whole of the future collapses into a local, immediate
test. That collapse is not a metaphor; it is a theorem, discovered
first for smooth boundaries by Nagumo in the 1940s and given its
general, set-valued form by Aubin and Frankowska decades later. It is
the single fact this book leans on hardest, because everything from
here on, volume, distance, repair, admissibility, is built on top of a
boundary between what remains possible and what does not, and this is
the theorem that tells us how to find that boundary without having to
simulate every possible future first.
\section*{Standing hypotheses}
Throughout this chapter, $X = \mathbb{R}^n$, $K \subseteq X$ is closed,
and $F : X \rightrightarrows X$ is a set-valued map satisfying the
following standing hypotheses, in force unless stated otherwise:
\begin{enumerate}[label=(H\arabic*)]
\item \emph{Nonempty compact convex values.} $F(x)$ is nonempty,
compact, and convex for every $x \in X$.
\item \emph{Upper semicontinuity.} For every $x \in X$ and every open
$U \supseteq F(x)$, there is a neighborhood $V$ of $x$ such that
$F(y) \subseteq U$ for all $y \in V$.
\item \emph{Linear growth.} There is $c > 0$ such that
$\sup\{\|v\| : v \in F(x)\} \leq c(1 + \|x\|)$ for all $x \in X$.
\end{enumerate}
\begin{remark}
These are the hypotheses under which the classical existence theory for
differential inclusions, and hence the theorem below, holds without
further qualification \cite{aubincellina1984}. They are restrictive
enough to rule out pathological $F$, and permissive enough to cover
every example used in this book.
\end{remark}
\section*{Contingent cones}
\begin{definition}[Contingent cone]
For $K \subseteq X$ closed and $x \in K$, the contingent (Bouligand)
cone to $K$ at $x$ is
\[
T_K(x) = \left\{ v \in X : \liminf_{h \to 0^+} \frac{\operatorname{dist}(x + hv, K)}{h} = 0 \right\}.
\]
Equivalently, $v \in T_K(x)$ iff there exist sequences $h_n \to 0^+$ and
$v_n \to v$ with $x + h_n v_n \in K$ for all $n$.
\end{definition}
\begin{proposition}[Contingent cone at a smooth boundary]
Suppose $K = \{x : g(x) \leq 0\}$ for some continuously differentiable
$g : X \to \mathbb{R}$ with $\nabla g(x) \neq 0$ whenever $g(x) = 0$.
Then for $x$ with $g(x) = 0$,
\[
T_K(x) = \{v \in X : \nabla g(x) \cdot v \leq 0\}.
\]
\end{proposition}
\begin{proof}
($\supseteq$) Suppose $\nabla g(x) \cdot v \leq 0$. By Taylor expansion,
$g(x + hv) = g(x) + h \nabla g(x) \cdot v + o(h) = h \nabla g(x) \cdot v + o(h) \leq o(h)$,
so for $h$ small enough $g(x+hv) \leq 0$ up to an $o(h)$ correction,
giving $\operatorname{dist}(x+hv, K) = o(h)$ and hence $v \in T_K(x)$.
($\subseteq$) Suppose $\nabla g(x) \cdot v > 0$. Then by the same Taylor
expansion, $g(x+hv) = h\nabla g(x)\cdot v + o(h) > 0$ for $h$ small
enough, and since $\nabla g(x) \neq 0$, the distance from $x+hv$ to
$\{g \leq 0\}$ is bounded below by a constant multiple of
$g(x+hv)$ for $h$ small, which is bounded below by a constant multiple
of $h$. Hence $\operatorname{dist}(x+hv,K)/h$ does not tend to $0$, so
$v \notin T_K(x)$.
\end{proof}
\begin{remark}
This recovers the Lyapunov-style computation of the previous chapter as
the special case $g = V - c$ on the sublevel set $K_c = \{V \leq c\}$:
the tangency condition $F(x) \cap T_K(x) \neq \varnothing$ becomes
$\min_{v \in F(x)} \nabla V(x) \cdot v \leq 0$, exactly the condition
used there. The Viability Theorem below is what allows this
computation to be done boundary point by boundary point, without
requiring a single function $V$ to work for the whole of $K$ at once.
\end{remark}
\section*{The Viability Theorem}
\begin{theorem}[Aubin's Viability Theorem]
Under the standing hypotheses, $K$ is viable under $F$, that is,
$\operatorname{Viab}_F(K) = K$, if and only if
\[
F(x) \cap T_K(x) \neq \varnothing \qquad \text{for every } x \in K.
\]
\end{theorem}
\begin{proof}[Proof of necessity]
Suppose $\operatorname{Viab}_F(K) = K$, and fix $x \in K$. Then
$\Gamma_K(x) \neq \varnothing$; let $\gamma \in \Gamma_K(x)$. For a.e.\
$t$, $\dot\gamma(t) \in F(\gamma(t))$; since $\gamma$ is absolutely
continuous, it is differentiable at a.e.\ point, and in particular there
is a sequence $h_n \to 0^+$ of times at which $\gamma$ is differentiable
with $\dot\gamma(0^+) := \lim_{h_n \to 0^+} \frac{\gamma(h_n) - x}{h_n}$
existing and lying in $F(x)$ by upper semicontinuity of $F$ and
continuity of $\gamma$ at $0$. Since $\gamma(h_n) \in K$ for every $n$
(as $\gamma \in \Gamma_K(x)$), the sequence $v_n = \frac{\gamma(h_n)-x}{h_n}$
satisfies $x + h_n v_n = \gamma(h_n) \in K$, which is exactly the
sequential definition of membership in $T_K(x)$ for the limit
$v = \dot\gamma(0^+)$. Hence $v \in F(x) \cap T_K(x)$, so the
intersection is nonempty.
\end{proof}
\begin{proof}[Proof sketch of sufficiency]
Suppose $F(x) \cap T_K(x) \neq \varnothing$ for every $x \in K$; fix
$x_0 \in K$. We construct a viable trajectory from $x_0$ by an Euler
polygon argument, following the classical method \cite{aubin1991}.
Fix $\varepsilon > 0$ and a step size $h > 0$. At $x_0$, choose
$v_0 \in F(x_0) \cap T_K(x_0)$; by definition of the contingent cone,
there is a point within distance $\varepsilon h$ of $x_0 + h v_0$
lying in $K$; call it $x_1$. Repeat from $x_1$: choose
$v_1 \in F(x_1) \cap T_K(x_1)$, find $x_2 \in K$ within $\varepsilon h$
of $x_1 + h v_1$, and so on, producing a sequence $x_0, x_1, x_2, \dots
\in K$ and a piecewise-linear curve $\gamma_{h,\varepsilon}$ interpolating
them.
Under the linear growth hypothesis (H3), a discrete Gr\"onwall argument
bounds $\|x_k\|$ uniformly on any fixed time interval $[0,M]$,
independently of $h$ and $\varepsilon$ (for $h,\varepsilon$ small
enough), which in turn bounds $\|v_k\|$ uniformly on $[0,M]$ by (H3)
again. The curves $\gamma_{h,\varepsilon}$ are therefore uniformly
bounded and equi-Lipschitz on $[0,M]$ for each fixed $M$, so by the
Arzel\`a--Ascoli theorem, a subsequence converges uniformly on $[0,M]$
as $h, \varepsilon \to 0$; a diagonal argument over $M = 1,2,3,\dots$
produces a single limit curve $\gamma : [0,\infty) \to X$.
Since $K$ is closed and each $x_k \in K$ with interpolation error
vanishing as $\varepsilon \to 0$, the limit satisfies $\gamma(t) \in K$
for all $t \geq 0$. Since $F$ has closed graph (a consequence of upper
semicontinuity together with compact values) and each $F(x)$ is convex,
a standard weak-convergence argument (Mazur's lemma, applied to the
difference quotients of $\gamma_{h,\varepsilon}$) upgrades the a.e.\
subsequential convergence of the derivatives to
$\dot\gamma(t) \in F(\gamma(t))$ for a.e.\ $t$. Hence
$\gamma \in \Gamma_K(x_0)$, so $x_0 \in \operatorname{Viab}_F(K)$.
The technical steps compressed here, the discrete Gr\"onwall bound and
the passage from weak convergence of difference quotients to inclusion
in $F$ via Mazur's lemma, are standard in the differential inclusions
literature and are given in full in \cite{aubincellina1984} and
\cite{aubinframkowska1990}.
\end{proof}
\begin{remark}
Necessity did not need convexity of $F(x)$; sufficiency does, precisely
at the step converting weak convergence of difference quotients into
membership in $F(\gamma(t))$. This asymmetry is typical of existence
theorems for differential inclusions and is one reason (H1) is stated
as a standing hypothesis rather than derived.
\end{remark}
\section*{Finite horizons and the viability kernel}
\begin{definition}
$V_T(K) = \{x \in K : \Gamma_K^T(x) \neq \varnothing\}$, the set of
states admitting a trajectory remaining in $K$ for at least time $T$.
\end{definition}
By Proposition~1 of the opening chapter (restriction is well-defined),
$V_{T_2}(K) \subseteq V_{T_1}(K)$ whenever $T_1 \leq T_2$: a trajectory
valid on a longer horizon restricts to one valid on a shorter horizon,
so the finite-horizon sets are nested and shrinking in $T$.
\begin{theorem}[Viability kernel as a nested intersection]
Under the standing hypotheses,
\[
\operatorname{Viab}_F(K) = \bigcap_{T > 0} V_T(K).
\]
\end{theorem}
\begin{proof}
($\subseteq$) If $x \in \operatorname{Viab}_F(K)$, any
$\gamma \in \Gamma_K(x)$ restricts to $\Gamma_K^T(x)$ for every $T$, so
$x \in V_T(K)$ for every $T$.
($\supseteq$) Suppose $x \in \bigcap_T V_T(K)$; for each $n = 1,2,3,\dots$
choose $\gamma_n \in \Gamma_K^n(x)$. As in the sufficiency proof above,
the linear growth hypothesis (H3) gives a discrete Gr\"onwall bound
making $\{\gamma_n\}_{n \geq M}$ uniformly bounded and equi-Lipschitz on
$[0,M]$ for each fixed $M$. Arzel\`a--Ascoli and a diagonal argument over
$M$ produce a subsequence converging uniformly on every compact
$[0,M]$ to a limit curve $\gamma : [0,\infty) \to X$. Closedness of $K$
gives $\gamma(t) \in K$ for all $t$; the same weak-convergence argument
via Mazur's lemma gives $\dot\gamma(t) \in F(\gamma(t))$ for a.e.\ $t$.
Hence $\gamma \in \Gamma_K(x)$ and $x \in \operatorname{Viab}_F(K)$.
\end{proof}
\begin{remark}
This is the genuine sense in which finite-horizon objects determine the
infinite-horizon viability kernel, and it is the fact that licenses
building continuation volume and continuation distance out of
finite-horizon reachable sets, as the appendix already commits to doing:
the finite-horizon approximation is not a separate, weaker theory but a
nested sequence whose intersection recovers the object of real interest,
under exactly the hypotheses already in force throughout this book.
\end{remark}
\section*{Capture basins}
\begin{definition}[Capture basin]
For a target $C \subseteq K$ and horizon $T > 0$, define
\[
\Gamma^T_{K,C}(x) = \left\{ \gamma : [0,T] \to K \;\middle|\;
\begin{array}{l}
\gamma(0) = x,\ \dot\gamma(t) \in F(\gamma(t)) \text{ a.e.}, \\
\exists\, t^\ast \in [0,T] \text{ with } \gamma(t^\ast) \in C \text{ and } \gamma(t) \in K\ \forall t \leq t^\ast
\end{array}
\right\},
\]
and $\operatorname{Capt}_F^T(K,C) = \{x \in K : \Gamma^T_{K,C}(x) \neq \varnothing\}$.
\end{definition}
\begin{remark}
This is Aubin's own formulation, allowing capture at any time
$t^\ast \leq T$ rather than requiring it exactly at the endpoint; it is
slightly more general than the simplified fixed-endpoint version used
provisionally in the appendix, which is recovered as the special case
in which capture is required to occur exactly at $t^\ast = T$. This
is rule (a), the prescribed-final-time rule, among the three
rules distinguished in Aubin's treatment of dynamic cooperative games;
rules (b) and (c) there correspond to different choices of when capture
is permitted to count.
\end{remark}
\begin{proposition}[Monotonicity in the horizon]
$\operatorname{Capt}_F^{T_1}(K,C) \subseteq \operatorname{Capt}_F^{T_2}(K,C)$
whenever $T_1 \leq T_2$.
\end{proposition}
\begin{proof}
If $\gamma \in \Gamma^{T_1}_{K,C}(x)$ witnesses capture at some
$t^\ast \leq T_1$, the same $\gamma$, viewed on the longer domain
$[0,T_2]$ by any admissible extension past $T_1$ (which exists, since
$\gamma(t^\ast) \in C \subseteq K$ and the trajectory need not continue
past $t^\ast$ to satisfy the defining condition), already witnesses
$t^\ast \leq T_1 \leq T_2$, so $x \in \operatorname{Capt}_F^{T_2}(K,C)$.
\end{proof}
\begin{remark}[The degenerate case $C = K$]
When $C = K$, every $x \in K$ trivially satisfies the capture condition
at $t^\ast = 0$, since $\gamma(0) = x \in K = C$ by hypothesis. Hence
$\operatorname{Capt}_F^T(K,K) = K$ for every $T > 0$, regardless of
whether $K$ itself is viable. This is not the interesting connection
between capture basins and the viability kernel; that connection is
supplied by the nested-intersection theorem above, using proper
finite-horizon persistence sets $V_T(K)$ rather than a capture basin
with a degenerate target. Flagging this explicitly is worth the space,
since the superficially similar-sounding claim, that capture basins
with $C = K$ reduce to the viability kernel, is false as stated and
should not be repeated uncritically.
\end{remark}
\section*{Regulation maps}
\begin{definition}[Regulation map]
For $x \in \operatorname{Viab}_F(K)$, the regulation map is
\[
\operatorname{Reg}_K(x) = F(x) \cap T_K(x),
\]
the set of admissible velocities at $x$ that keep the system tangent to
$K$.
\end{definition}
\begin{remark}
$\operatorname{Reg}_K(x)$ is exactly the continuation operator of the
opening chapter, differentiated: it names, at each viable state, which
velocities are available for the next instant, rather than which whole
trajectories are available from here on. A continuous selection
$x \mapsto v(x) \in \operatorname{Reg}_K(x)$, where one exists, generates
an actual trajectory by solving $\dot\gamma(t) = v(\gamma(t))$; this is
the mechanism by which the existence statement of the Viability Theorem
is turned into an actual rule for staying viable, rather than a bare
guarantee that some rule exists. The next chapter takes up the case
where $\operatorname{Reg}_K(x)$ contains more than one admissible
velocity and a further criterion is needed to choose among them. This is the
point at which Aubin's own dynamic core, and this book's account of
selection, begin.
\end{remark}
\chapter{The Selection Already Inside Viability}
A ship's captain who has confirmed that some route home exists has
confirmed very little. There may be a dozen routes, and most captains
worth the title do not stop at the first one that avoids the rocks;
they ask which route survives the worst weather the season could
plausibly throw at them, and they plot that one. The difference between
those two captains is not a difference in what they know about the
sea. It is a difference in what they are willing to leave to chance
once they know a safe passage is out there somewhere.
This is worth dwelling on because it is easy to hear ``does a way
forward exist'' and ``which way forward should I take'' as two stages
of the same question, the second simply refining the first. They are
not quite that. Existence is a yes-or-no fact about a whole space of
possibilities; selection requires a standard for comparing the members
of that space against one another, and no amount of sharpening the
existence question will produce that standard on its own. What makes
the sea captain's practice worth studying formally is that the
standard she actually uses (survive the worst plausible weather) is
not an ad hoc addition bolted onto navigation from outside. It falls
directly out of taking the existence question seriously under
uncertainty, once uncertainty is treated as something an adversary
controls rather than something that merely happens. Follow the
existence question that far, and a selection criterion arrives
whether it was asked for or not.
\section*{From existence to selection}
Recall the regulation map of the previous chapter,
$\operatorname{Reg}_K(x) = F(x) \cap T_K(x)$: the set of admissible
velocities at $x$ that keep a trajectory tangent to $K$. The Viability
Theorem guarantees this set is nonempty at every viable state, but says
nothing about which of its elements to use when it contains more than
one. Viability theory, as developed so far, answers only whether
$\operatorname{Reg}_K(x)$ is empty; any measurable selection from it
generates an equally valid viable trajectory, and the theory is
indifferent among them.
To ask which selection is best requires enough additional structure to
make ``best'' mean something, and the minimal such structure separates
the velocity at each state into a part the system controls and a part
it does not.
\begin{definition}[Controlled dynamical game]
Let $P : X \rightrightarrows U$ and $Q : X \rightrightarrows V$ be
set-valued maps (the available controls, or allotments, and the
available perturbations, or tyches, at each state), and let
$f : X \times U \times V \to X$. A feedback is a map
$u_e : X \to U$ with $u_e(x) \in P(x)$ for all $x$. For a feedback
$u_e$, the trajectory set is
\[
C_{u_e}(x) = \left\{ (\gamma, v(\cdot)) \;\middle|\;
\dot\gamma(t) = f(\gamma(t), u_e(\gamma(t)), v(t)),\ v(t) \in Q(\gamma(t)),\ \gamma(0) = x \right\}.
\]
\end{definition}
\begin{proposition}[Uncontrolled viability is the trivial case]
If $P(x) = \{\star\}$ is a singleton for every $x$ (no genuine control
freedom), then setting $F(x) := \{f(x,\star,v) : v \in Q(x)\}$, the
trajectory set $C_{u_e}(x)$ under the unique feedback $u_e \equiv \star$
coincides with $\Gamma_{K}(x)$ of Chapter~1 for $K = X$: a curve
$\gamma$ appears in $C_{u_e}(x)$ (via some choice of $v(\cdot)$) if and
only if $\gamma \in \Gamma_X(x)$ under $F$.
\end{proposition}
\begin{proof}
If $(\gamma, v(\cdot)) \in C_{u_e}(x)$, then
$\dot\gamma(t) = f(\gamma(t), \star, v(t)) \in F(\gamma(t))$ for a.e.\
$t$ by definition of $F$, and $\gamma(0) = x$, so $\gamma \in \Gamma_X(x)$.
Conversely, if $\gamma \in \Gamma_X(x)$, then for a.e.\ $t$,
$\dot\gamma(t) \in F(\gamma(t)) = \{f(\gamma(t),\star,v) : v \in Q(\gamma(t))\}$,
so a measurable selection $v(t)$ with
$\dot\gamma(t) = f(\gamma(t),\star,v(t))$ exists by standard measurable
selection theorems, giving $(\gamma, v(\cdot)) \in C_{u_e}(x)$.
\end{proof}
\begin{remark}
This is the sense in which everything proved in the previous two
chapters is a special case of what follows, not a separate theory:
viability theory is controlled dynamics with the control stripped out.
Reintroducing it is not a generalization for its own sake; it is the
minimal move needed to ask which viable trajectory a system should
follow, rather than merely whether one exists.
\end{remark}
\section*{Guaranteed valuation}
Fix a horizon $T$ and a terminal payoff $u : X \to \mathbb{R}$, to be
minimized in the worst case over perturbations. Following Aubin's
treatment of dynamic cooperative games \cite{aubin2001}, and taking the
discount term $m \equiv 0$ for simplicity, define
\[
V^\sharp(T,x) = \inf_{u_e(\cdot) \in P(\cdot)} \sup_{(\gamma,v(\cdot)) \in C_{u_e}(x)} u(\gamma(T)).
\]
$V^\sharp(T,x)$ is the smallest terminal cost the system can guarantee
against every admissible perturbation, once the best available feedback
has been chosen.
\begin{definition}[The menu and the guaranteed-optimal selections]
$\Gamma(x) := P(x)$, the full menu of controls available at $x$.
$\Gamma^\star(T,x) := \{p \in P(x) : \text{using } p \text{ at } x
\text{ achieves the value } V^\sharp(T,x) \text{ against every
admissible continuation of the worst-case perturbation}\}$, made
precise via Aubin's dynamic core: when $V^\sharp$ is differentiable,
\[
\Gamma^\star(T,x) = \left\{ p \in P(x) : \sup_{v \in Q(x)} \left[ \left\langle \frac{\partial V^\sharp}{\partial x}(T,x) - p,\, f(x,v) \right\rangle \right] \leq -\frac{\partial V^\sharp}{\partial T}(T,x) \right\}.
\]
\end{definition}
\begin{remark}[A notational warning]
$\Gamma$ is being reused here for something different in kind from
$\Gamma_K(x)$ of Chapter 1: there, $\Gamma_K(x)$ was a set of whole
trajectories; here, $\Gamma(x) := P(x)$ is a set of instantaneous
control values available at $x$. The two are related, since choosing
$p \in P(x)$ at every state along a trajectory is what generates a
member of $\Gamma_K(x)$ in the first place, but they are not the same
kind of object, and conflating them would be a mistake. The symbol is
reused rather than replaced because the pattern being tracked, a menu
$\Gamma$ narrowed by some criterion to a selected subset $\Gamma^\star$,
is the same pattern each time it appears, in this chapter, in
Chapter 8, and in Chapter 9, and giving each instance its own letter
would obscure that repetition rather than clarify it. Context (whether
the surrounding discussion is about trajectories or about
instantaneous controls) disambiguates which $\Gamma$ is meant at each
occurrence from here on.
\end{remark}
\begin{remark}
This is Aubin's own dynamic-core condition, restated here with $m\equiv
0$; that the achieving-the-value characterization and the differential
inequality characterization agree is Aubin's own result and is not
re-derived here \cite{aubin2001}. What matters for this chapter is only
that $\Gamma^\star(T,x)$ is defined as a subset of $\Gamma(x) = P(x)$ by
an extra condition beyond mere availability, so trivially
$\Gamma^\star(T,x) \subseteq \Gamma(x)$, and the question is whether
that inclusion is ever strict in a way that matters.
\end{remark}
\begin{proposition}
$\Gamma^\star(T,x) \subseteq \Gamma(x)$ for every $T, x$.
\end{proposition}
\begin{proof}
Immediate: $\Gamma^\star(T,x)$ is defined as $\{p \in P(x) : \dots\}$, a
subset of $P(x) = \Gamma(x)$ by construction.
\end{proof}
\section*{A worked example}
The inclusion above is trivial as a set-theoretic fact; what is not
trivial is that it is often strict, and that the gap it leaves out
represents controls that are viable in the sense of the previous
chapter yet indefensible in the sense of this one.
\begin{example}
Let $X = \mathbb{R}$, $P(x) = \{-1, +1\}$ for all $x$ (two available
constant controls), $Q(x) = [-\tfrac12, \tfrac12]$ (a bounded
perturbation), and $f(x,u,v) = u + v$. Take terminal payoff
$u(x) = x^2$ and horizon $T > 0$, with the system started at $x = 0$.
Consider two feedbacks: $u_e^{(1)}(x) \equiv +1$ for all $x$, and
$u_e^{(2)}(x) = -\operatorname{sign}(x)$ for $x \neq 0$ (with the
value at $x=0$ resolved by the Filippov sliding-mode convention
\cite{filippov1988,utkin1992}, as is standard for discontinuous
feedbacks of this type).
\end{example}
\begin{proposition}
Under $u_e^{(1)}$, $\sup_{(\gamma,v(\cdot))} \gamma(T)^2 = \tfrac{9}{4}T^2$.
Under $u_e^{(2)}$, $\sup_{(\gamma,v(\cdot))} \gamma(T)^2 = 0$. Hence