Repository navigation
Expand file tree
/
Copy pathgenerative_continuation_geometry.tex
More file actions
2537 lines (2271 loc) · 120 KB
/
Copy pathgenerative_continuation_geometry.tex
File metadata and controls
2537 lines (2271 loc) · 120 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{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{conjecture}{Conjecture}[chapter]
\newtheorem{openproblem}{Open Problem}[chapter]
\title{\textbf{Generative Continuation Geometry}}
\author{Flyxion}
\date{July 2026}
\begin{document}
\maketitle
\tableofcontents
\chapter{Continuation Attractors}
A fern frond is not drawn leaf by leaf. Look closely at one pinna and
it is, near enough, a smaller fern; look closer still at one pinnule
of that pinna and it is, near enough, a smaller fern again. Nobody
specified the shape of the whole plant directly. A small number of
growth rules, applied to their own output over and over, produced
something with more detail than any of the rules themselves contain.
The shape is not stored anywhere. It is generated.
\emph{Continuation Geometry} spent its length measuring futures:
how much room a state has, how far apart two states' futures sit, when
a distinction between futures collapses to nothing. All of that took
the space of futures as given, something to be measured once it
already existed. This book asks a different question. Can a space of
futures itself be grown, the way the fern is grown, from a small
number of simple moves applied to their own output? If continuation
spaces can be generated rather than merely measured, the natural unit
of study stops being the space and becomes the small set of
operations that built it.
This opening chapter answers that question in the one case where the
answer is already, in a precise sense, classical. It is not a new
theorem. It is an old and beautiful one, due to Hutchinson, applied
here to the specific metric geometry the first book already built. The
payoff is that self-similar continuation structures, the kind
suggested by a duck, a crown, and a leaf all arising from the same
underlying rule, turn out to rest on exactly the Hausdorff-distance
machinery already in hand, with nothing new invented and nothing
smuggled in from outside.
\section*{Preliminaries: expansion is not generation}
Two facts from \emph{Continuation Geometry} are worth restating here,
briefly and as what they actually are: lemmas, not theorems, and
useful precisely because they show what generation is \emph{not}.
\begin{lemma}[Reachable expansion]
For $S_t \subseteq X$ and $\mathrm{Adj}(S_t) := \bigcup_{x \in S_t} \Phi(x)$
under a one-step reachability map $\Phi$, the sequence
$S_{t+1} := S_t \cup \mathrm{Adj}(S_t)$ satisfies $S_t \subseteq S_{t+1}$
for every $t$.
\end{lemma}
\begin{proof}
Immediate: $S_t \subseteq S_t \cup \mathrm{Adj}(S_t) = S_{t+1}$ by
definition of union.
\end{proof}
\begin{remark}
This, together with the reachable-set recursion proved by induction
in \emph{Continuation Geometry} (Kauffman's adjacent possible as an
instance of finite-horizon reachability), is the full content of what
that book called generation, and it is worth being honest that it
earns the name only loosely. The sequence $S_0, S_1, S_2, \dots$ grows
monotonically and without bound unless the underlying space is
finite; nothing about it converges, and nothing about it is
self-similar. It is accumulation, not generation in the sense the
fern illustrates. What follows is different in kind: rather than a
sequence of ever-larger unions, it is a fixed operator applied
repeatedly to itself, and the interesting behavior is not growth but
convergence to a shape that reproduces its own generating rule.
\end{remark}
\section*{The space of shapes}
Fix $X = \mathbb{R}^n$ (or any complete metric space) and let
$\mathcal{K}(X)$ denote the set of nonempty compact subsets of $X$,
equipped with the Hausdorff distance $d_H$ of Chapter 6 of
\emph{Continuation Geometry}.
\begin{theorem}[Completeness of the space of shapes]
\label{thm:complete}
If $(X,d)$ is complete, then $(\mathcal{K}(X), d_H)$ is complete.
\end{theorem}
\begin{proof}[Proof sketch]
Given a Cauchy sequence $\{A_n\}$ in $(\mathcal{K}(X), d_H)$, define
$A := \bigcap_{N \geq 1} \overline{\bigcup_{n \geq N} A_n}$. One shows,
using completeness of $X$ and the Cauchy property of $\{A_n\}$ in
$d_H$, that $A$ is nonempty and compact, and that $d_H(A_n, A) \to 0$.
This is a standard but genuinely technical result in its full form,
requiring care with total boundedness and a diagonal argument to
extract convergent selections from each $A_n$; the complete argument
is given in \cite{falconer2003} and \cite{barnsley1988} and is not
reproduced in full here.
\end{proof}
\begin{remark}
This theorem is doing real work that is easy to take for granted.
Without it, nothing built on top of $(\mathcal{K}(X), d_H)$ would be
guaranteed to have a limit at all; a sequence of shapes could get
closer and closer together, in the sense of Hausdorff distance,
without converging to any actual shape. That this cannot happen, that
the space of compact reachable sets is itself complete whenever the
underlying state space is, is the entire reason the construction
below is guaranteed to produce something rather than merely getting
closer to nothing in particular.
\end{remark}
\section*{Contractions and the Hutchinson operator}
\begin{definition}[Contraction]
$T : X \to X$ is a contraction with constant $c \in [0,1)$ if
$d(Tx, Ty) \leq c\, d(x,y)$ for all $x, y \in X$.
\end{definition}
\begin{lemma}[Contractions act contractively on shapes]
\label{lem:setcontract}
If $T$ is a contraction on $X$ with constant $c$, then for
$A, B \in \mathcal{K}(X)$, $d_H(T(A), T(B)) \leq c\, d_H(A,B)$, where
$T(A) := \{Ta : a \in A\}$.
\end{lemma}
\begin{proof}
Fix $a \in A$. Since $B$ is compact, there is $b \in B$ with
$d(a,b) = \operatorname{dist}(a,B) \leq d_H(A,B)$. Then
$d(Ta,Tb) \leq c\, d(a,b) \leq c\, d_H(A,B)$, and since
$Tb \in T(B)$, $\operatorname{dist}(Ta, T(B)) \leq c\, d_H(A,B)$.
Taking the supremum over $a \in A$ gives
$\sup_{a' \in T(A)} \operatorname{dist}(a', T(B)) \leq c\, d_H(A,B)$.
The symmetric argument, exchanging the roles of $A$ and $B$, gives the
same bound the other way, so $d_H(T(A),T(B)) \leq c\, d_H(A,B)$.
\end{proof}
\begin{lemma}[Unions are 1-Lipschitz]
\label{lem:unionlip}
For finite collections $\{A_i\}_{i=1}^n$, $\{B_i\}_{i=1}^n$ of nonempty
compact sets, $d_H\left(\bigcup_i A_i, \bigcup_i B_i\right) \leq
\max_i d_H(A_i, B_i)$.
\end{lemma}
\begin{proof}
Let $m := \max_i d_H(A_i,B_i)$. For $x \in \bigcup_i A_i$, $x \in A_j$
for some $j$, so $\operatorname{dist}(x, B_j) \leq d_H(A_j,B_j) \leq m$,
and since $B_j \subseteq \bigcup_i B_i$,
$\operatorname{dist}(x, \bigcup_i B_i) \leq \operatorname{dist}(x,B_j) \leq m$.
Taking the supremum over $x \in \bigcup_i A_i$ gives
$\sup_{x} \operatorname{dist}(x, \bigcup_i B_i) \leq m$. The symmetric
argument, exchanging the roles of the two collections, gives the same
bound the other way, so
$d_H\left(\bigcup_i A_i, \bigcup_i B_i\right) \leq m$.
\end{proof}
\begin{definition}[Hutchinson operator]
Given contractions $T_1, \dots, T_n$ on $X$ with constants
$c_1, \dots, c_n$, define $\mathcal{H} : \mathcal{K}(X) \to \mathcal{K}(X)$
by $\mathcal{H}(A) := \bigcup_{i=1}^n T_i(A)$.
\end{definition}
\begin{proposition}[$\mathcal{H}$ is a contraction on shapes]
\label{prop:hcontract}
$d_H(\mathcal{H}(A), \mathcal{H}(B)) \leq c\, d_H(A,B)$ for
$c := \max_i c_i < 1$.
\end{proposition}
\begin{proof}
By Lemma~\ref{lem:unionlip},
$d_H(\mathcal{H}(A), \mathcal{H}(B)) = d_H\left(\bigcup_i T_i(A), \bigcup_i T_i(B)\right)
\leq \max_i d_H(T_i(A), T_i(B))$. By Lemma~\ref{lem:setcontract},
$d_H(T_i(A), T_i(B)) \leq c_i\, d_H(A,B)$ for each $i$. Hence
$d_H(\mathcal{H}(A), \mathcal{H}(B)) \leq \max_i c_i\, d_H(A,B) = c\, d_H(A,B)$.
\end{proof}
\section*{The Continuation Attractor Theorem}
\begin{theorem}[Continuation Attractor Theorem, after Hutchinson]
\label{thm:attractor}
Let $T_1, \dots, T_n$ be contractions on a complete metric space $X$
with constants $c_1, \dots, c_n$, and let $c := \max_i c_i < 1$. Then
there exists a unique $A^\ast \in \mathcal{K}(X)$ with
$\mathcal{H}(A^\ast) = A^\ast$, and for any $A_0 \in \mathcal{K}(X)$,
\[
d_H\big(\mathcal{H}^k(A_0), A^\ast\big) \leq \frac{c^k}{1-c}\, d_H(A_0, \mathcal{H}(A_0)) \xrightarrow{k \to \infty} 0.
\]
\end{theorem}
\begin{proof}
By Theorem~\ref{thm:complete}, $(\mathcal{K}(X), d_H)$ is complete. By
Proposition~\ref{prop:hcontract}, $\mathcal{H}$ is a contraction on
this complete metric space. The Banach Fixed Point Theorem
\cite{banach1922} then
supplies a unique fixed point $A^\ast$ and the stated geometric
convergence rate, applied directly to $\mathcal{H}$ on
$(\mathcal{K}(X), d_H)$.
\end{proof}
\begin{remark}
No part of this proof required inventing anything: completeness of
$\mathcal{K}(X)$ is Theorem~\ref{thm:complete}, the contraction
property of $\mathcal{H}$ is Proposition~\ref{prop:hcontract} built
from two short lemmas, and the conclusion is the Banach Fixed Point
Theorem, unmodified. This is the sense in which this chapter is
mathematically conservative even though its subject, self-similar
generated shape, sounds exotic: every piece was already sitting in
the metric structure Chapter 6 of \emph{Continuation Geometry}
built for an entirely different purpose.
\end{remark}
\section*{A worked attractor}
\begin{example}[Two refusals]
Let $X = [0,1]$, and let $T_1(x) := x/3$, $T_2(x) := x/3 + 2/3$, both
contractions with constant $c = 1/3$. The attractor $A^\ast$ satisfying
$A^\ast = T_1(A^\ast) \cup T_2(A^\ast)$ is the standard middle-thirds
Cantor set.
\end{example}
\begin{remark}
Read in the vocabulary of continuation, $T_1$ and $T_2$ are the two
survivors of a repeated refusal: at every stage, whatever occupies the
middle third of the current set of admissible outcomes is discarded,
and the remaining two-thirds are each rescaled to stand in for the
whole. What survives in the limit is neither of the two pieces nor
their simple union at any finite stage, but the unique shape that
reproduces itself exactly under the same rule that built it,
uncountable, nowhere dense, and entirely determined by two contraction
maps. Whether operators named Collapse or Refuse elsewhere in this
corpus are literally instances of $T_i$ in this technical sense, maps
satisfying the contraction inequality on some genuine underlying
metric, is a claim about those specific constructions and is not
asserted here; what is established is only that \emph{if} a family of
continuation-narrowing operators can be shown to be contractions on
$(\mathcal{K}(X), d_H)$, this theorem applies to them without
modification.
\end{remark}
\section*{What this chapter does not yet claim}
This chapter establishes that self-similar continuation structures are
mathematically available, not that any particular structure this
corpus has described informally is an instance of one. Two gaps in
particular are worth naming rather than leaving implicit. First,
nothing here says which operators arising from an actual continuation
problem, as opposed to a hand-picked pair of affine maps on $[0,1]$,
are genuinely contractions on a genuine metric; that has to be checked
case by case, the way it would for any application of a fixed-point
theorem. Second, and more speculatively, the question of whether a
small, fixed family of primitive operators could generate \emph{every}
admissible continuation geometry to arbitrary Hausdorff-distance
accuracy, a continuation-geometric analogue of universal approximation,
is not addressed by this theorem and is not claimed here. It is a
genuine open question, and answering it is a natural place for this
book to go next, once the machinery in the chapters between here and
there, quotient maps and operator composition among them, has been
built on equally solid ground.
\chapter{Continuation Quotients}
Two rivers can rise on opposite sides of a mountain range, cross
entirely different terrain, pass through valleys that share nothing
in common, and still arrive at the same confluence and become, from
that point on, one river. Everything about their separate histories,
which rocks each one wore smooth, which tributaries fed it, how long
each took to descend, becomes irrelevant downstream of the merge. A
boat launched below the confluence cannot tell which branch its water
came from. Not because the information was hidden. Because it no
longer makes a difference to anything the boat will experience from
here on.
The previous chapter showed that a continuation space can be grown
from a small number of generating operators. The natural next
question is what happens when two different generating histories,
different sequences of operators, different starting points, produce
the same continuation space at the end. When does that happen, and
what, if anything, does it mean? The temptation is to reach
immediately for language like information loss or memory collapse,
language already used loosely elsewhere in this corpus. This chapter
resists that temptation on purpose. It defines the relevant object
first, proves what can be proved about it without reference to any
interpretation, and only afterward asks what, if anything, has been
lost.
\section*{The continuation quotient}
\begin{definition}[Continuation quotient]
Let $H$ be a set of histories and let $\Gamma : H \to \mathcal{K}(X)$
assign to each history its continuation space (or, following Chapter 6
of \emph{Continuation Geometry}, its finite-horizon reachable set
$R_T$ for a fixed horizon $T$). Define $h_1 \sim_\Gamma h_2$ iff
$\Gamma(h_1) = \Gamma(h_2)$. This is an equivalence relation
(reflexivity, symmetry, and transitivity all following immediately
from equality of sets), and the continuation quotient is
$H/{\sim_\Gamma}$, with quotient map $q : H \to H/{\sim_\Gamma}$
sending $h$ to its equivalence class $[h]$.
\end{definition}
Nothing has been said yet about distinctions, memory, or loss. This is
deliberate: the object exists, and has structure, independently of any
of those interpretations.
\begin{proposition}[Well-defined continuation projection]
\label{prop:universal}
There exists a unique map $\bar\Gamma : H/{\sim_\Gamma} \to \mathcal{K}(X)$
such that $\Gamma = \bar\Gamma \circ q$.
\end{proposition}
\begin{proof}
Define $\bar\Gamma([h]) := \Gamma(h)$. This is well-defined: if
$[h_1] = [h_2]$, then $h_1 \sim_\Gamma h_2$, so
$\Gamma(h_1) = \Gamma(h_2)$ by definition of $\sim_\Gamma$, so the
value assigned to $[h_1] = [h_2]$ does not depend on which
representative was chosen. By construction,
$\bar\Gamma(q(h)) = \bar\Gamma([h]) = \Gamma(h)$ for every $h \in H$,
so $\Gamma = \bar\Gamma \circ q$. For uniqueness, suppose
$\bar\Gamma'$ also satisfies $\Gamma = \bar\Gamma' \circ q$. Since $q$
is surjective, every element of $H/{\sim_\Gamma}$ is $q(h)$ for some
$h$, and $\bar\Gamma'(q(h)) = \Gamma(h) = \bar\Gamma(q(h))$ for every
such $h$, so $\bar\Gamma' = \bar\Gamma$.
\end{proof}
\begin{remark}
This is routine, and it is worth saying plainly that it is routine:
it is the universal property of a quotient set, applied to nothing
more exotic than the map $\Gamma$ itself. What makes it worth stating
is not its difficulty but what it licenses: from here on,
$\bar\Gamma$ is a genuine function on the quotient, not merely a
suggestive notation, and anything proved about $\bar\Gamma$ is a fact
about the quotient set itself, not about any particular history used
to reach a point in it.
\end{remark}
\section*{A structural lemma}
The worked example below relies on the following fact about
time-homogeneous, deterministic-image dynamics, worth isolating here.
\begin{lemma}[One-step agreement propagates]
\label{lem:propagate}
Suppose $R_1(x) = R_1(y)$ for states $x, y$ under a time-homogeneous
reachability map. Then $R_T(x) = R_T(y)$ for every $T \geq 1$.
\end{lemma}
\begin{proof}
By induction on $T$. The case $T=1$ is the hypothesis. Suppose
$R_T(x) = R_T(y)$. Since the dynamics are time-homogeneous,
$R_{T+1}(x) = \bigcup_{z \in R_T(x)} R_1(z)$ and likewise for $y$; since
$R_T(x) = R_T(y)$, these unions range over the same set of states $z$,
so $R_{T+1}(x) = R_{T+1}(y)$.
\end{proof}
\begin{remark}
This is the precise sense in which a collision, once it happens, is
permanent under time-homogeneous dynamics: two states whose immediate
futures coincide exactly will have every subsequent future coincide
as well, since there is nothing left in either state's future to
distinguish it from the other's.
\end{remark}
\section*{Quotient invariants}
\begin{proposition}[Quotient invariance]
\label{prop:invariance}
Let $F : \mathcal{K}(X) \to Y$ be any function. Then
$F \circ \Gamma : H \to Y$ factors through $q$: there is a unique
$\bar{F} : H/{\sim_\Gamma} \to Y$ with $F \circ \Gamma = \bar{F} \circ q$.
\end{proposition}
\begin{proof}
$F \circ \Gamma : H \to Y$ satisfies $(F \circ \Gamma)(h_1) = (F \circ \Gamma)(h_2)$
whenever $h_1 \sim_\Gamma h_2$, since $\Gamma(h_1) = \Gamma(h_2)$ by
definition of $\sim_\Gamma$ and $F$ is a function, hence single-valued.
Apply Proposition~\ref{prop:universal} with $F \circ \Gamma$ in place
of $\Gamma$ to obtain the factorization.
\end{proof}
\begin{corollary}
Continuation volume $\operatorname{Vol}_T$, the distinction margin
$A_{\mathcal D}^T$ relative to any fixed reference distinction, and
boundary proximity $\delta$, all defined in \emph{Continuation
Geometry} as functions of a reachable set or of a state's position
relative to one, descend to well-defined functions on
$H/{\sim_\Gamma}$.
\end{corollary}
\begin{proof}
Each is of the form $F(\Gamma(h))$ or $F(R_T(x(h)))$ for a function
$F$ depending only on the reachable set itself (Lebesgue measure of
the set, Hausdorff distance to another fixed set, ambient distance
from a fixed reference set), so Proposition~\ref{prop:invariance}
applies directly to each.
\end{proof}
\begin{remark}
This is a broader statement than it might first appear. It says that
any quantity this book or its predecessor has defined, or will define,
purely in terms of a reachable set, without reference to which
history produced that set, is automatically constant on
$\sim_\Gamma$-equivalence classes, with no further argument required
each time a new such quantity is introduced. The work of checking
invariance, done once here, does not need to be redone for every
future functional built the same way.
\end{remark}
\section*{Finite checkability}
Lemma~\ref{lem:propagate} shows that equal one-step images propagate
forward forever. It says nothing about the converse: two states whose
one-step images differ can still coincide at some later horizon.
Nothing in this chapter has claimed otherwise, and the general
question, whether continuation equivalence at every horizon can still
be checked in finitely many steps despite this, deserves a correct
answer rather than an implied one.
\begin{proposition}[Eventual periodicity bounds the check]
\label{prop:periodicity}
If $X$ is finite, then for any $x, y \in X$ the paired sequence
$\big(R_1(x),R_1(y)\big), \big(R_2(x),R_2(y)\big), \dots$ is eventually
periodic, with both the pre-period and the period bounded by
$4^{|X|}$. Consequently, whether $R_T(x) = R_T(y)$ holds for all
sufficiently large $T$ can be determined by comparing $R_T(x)$ and
$R_T(y)$ for $T$ ranging only up to $2 \cdot 4^{|X|}$.
\end{proposition}
\begin{proof}
Each pair $\big(R_T(x), R_T(y)\big)$ is an element of $2^X \times 2^X$,
which has $4^{|X|}$ elements, so the paired sequence takes at most
$4^{|X|}$ distinct values; among its first $4^{|X|}+1$ terms, two must
coincide by pigeonhole, $\big(R_p(x),R_p(y)\big) =
\big(R_q(x),R_q(y)\big)$ for some $1 \leq p < q \leq 4^{|X|}+1$.
Time-homogeneity means $R_{T+1} = \Psi(R_T)$ for the fixed map
$\Psi(S) := \bigcup_{z \in S} \Phi(z)$, applied identically to the $x$
and $y$ components, so equality of the pair at $p$ and $q$ propagates:
$\big(R_{p+j}(x), R_{p+j}(y)\big) = \big(\Psi^j(R_p(x)), \Psi^j(R_p(y))\big)
= \big(\Psi^j(R_q(x)), \Psi^j(R_q(y))\big) = \big(R_{q+j}(x), R_{q+j}(y)\big)$
for every $j \geq 0$: the \emph{paired} sequence is periodic with
period dividing $q - p \leq 4^{|X|}$ from index $p \leq 4^{|X|}$
onward. Comparing the pair over one further full period, at most
$4^{|X|}$ more steps beyond the pre-period bound, settles whether $x$
and $y$ agree from that point on.
\end{proof}
\begin{remark}
Bounding each sequence's own period separately and then comparing
them over a single shared window is not sufficient on its own: two
sequences with different individual periods $p_x, p_y$ need not
realign within $\max(p_x,p_y)$ steps, only within
$\operatorname{lcm}(p_x,p_y)$ steps, which can exceed either period
substantially. Working with the pair from the start avoids this,
since the pair's own period already accounts for however the two
component periods interact.
\end{remark}
\begin{remark}
This is the correct version of the claim that continuation
equivalence is finitely checkable. It is a coarser bound than a sharp
analysis of any particular system would give, since $2^{|X|}$ is a
worst case over all possible dynamics on a state space of that size,
but it is unconditional: no structural assumption on $\Phi$ beyond
finiteness of $X$ and time-homogeneity is needed. The traffic light
below reaches its own eventual behavior in four steps, far inside this
bound, and it is worth seeing exactly how.
\end{remark}
\section*{The traffic light, revisited}
Recall the traffic light of \emph{Continuation Geometry}, Chapter 10:
states $(k, r) \in \{0,1,2,3\} \times \{0,1\}$, phase $k$ and pending
request $r$. To compute $\Gamma$ properly requires treating the
system's branching explicitly rather than leaving it implicit. Let
$e \in \{0,1\}$ record whether a new pedestrian request arrives during
a given step (exogenous, not controlled by the phase). The transition
is
\[
\Phi((k,r), e) = \left( (k+1) \bmod 4,\ r' \right), \qquad
r' = \begin{cases}
0 & k+1 \equiv 0 \pmod 4 \text{ (clear on arrival at red)} \\
1 & k+1 \not\equiv 0 \pmod 4 \text{ and } (r = 1 \text{ or } e = 1) \\
0 & k+1 \not\equiv 0 \pmod 4 \text{ and } r = 0 \text{ and } e = 0.
\end{cases}
\]
Let $H$ be the set of finite paths of the system, and
$\Gamma(h) := R_1(x(h))$, the one-step reachable set from the
endpoint of $h$, ranging over $e \in \{0,1\}$.
\begin{proposition}[Exact computation of the traffic-light quotient at $T=1$]
\label{prop:tlquotient}
Under $\sim_\Gamma$, the eight states fall into exactly seven classes:
the single nontrivial class $\{(3,0),(3,1)\}$, and six singleton
classes, one for each remaining state.
\end{proposition}
\begin{proof}
Compute $R_1$ for all eight states directly from $\Phi$. For
$k \neq 3$, arrival is at a nonzero phase and the ordinary branching
rule applies, $r' = 1$ iff $r=1$ or $e=1$:
\begin{align*}
R_1(0,0) &= \{(1,0),(1,1)\}, & R_1(0,1) &= \{(1,1)\}, \\
R_1(1,0) &= \{(2,0),(2,1)\}, & R_1(1,1) &= \{(2,1)\}, \\
R_1(2,0) &= \{(3,0),(3,1)\}, & R_1(2,1) &= \{(3,1)\}.
\end{align*}
For $k=3$, arrival is at phase $0$, so clearing fires unconditionally
regardless of $r$ or $e$:
\[
R_1(3,0) = \{(0,0)\}, \qquad R_1(3,1) = \{(0,0)\}.
\]
Comparing all eight sets directly: the three size-two sets,
$\{(1,0),(1,1)\}$, $\{(2,0),(2,1)\}$, $\{(3,0),(3,1)\}$, are pairwise
distinct; the four remaining size-one sets, $\{(1,1)\}$, $\{(2,1)\}$,
$\{(0,0)\}$ (appearing twice, from $(3,0)$ and $(3,1)$), and
$\{(0,0)\}$ again, give exactly one coincidence, $R_1(3,0) = R_1(3,1)
= \{(0,0)\}$, and no other.
\end{proof}
Now define the mechanical quotient $\sim_M$ independently of
$\Gamma$, as the identification the design actually performs: the
clearing rule fires on the transition \emph{into} phase $0$, so it is
the two states about to make that transition, $(3,0)$ and $(3,1)$,
that the design treats as interchangeable, both being followed
unconditionally by $(0,0)$.
\begin{proposition}[The design is exact at $T=1$]
$\sim_M\ =\ \sim_\Gamma$.
\end{proposition}
\begin{proof}
Both relations have exactly one nontrivial class, $\{(3,0),(3,1)\}$,
and agree as singletons on every other state: $\sim_M$ by definition
of the clearing rule, $\sim_\Gamma$ by Proposition~\ref{prop:tlquotient}.
Two equivalence relations on the same finite set with identical
partitions are equal.
\end{proof}
\begin{remark}
It is worth being precise about which of two similar-sounding designs
this result actually vindicates. Clearing on arrival at phase $0$
(the rule above) is exact. A superficially similar alternative,
clearing whenever the system is currently \emph{at} phase $0$ rather
than arriving there, would instead identify $(0,0)$ and $(0,1)$. But
$R_1(0,0) = \{(1,0),(1,1)\} \neq \{(1,1)\} = R_1(0,1)$: a state at the
red light with no pending request can still branch toward either
outcome next, while one with a pending request cannot, precisely
because that request is about to be served. The alternative design
would merge two states with genuinely different one-step futures,
exactly the coarser, information-destroying case this chapter opened
by distinguishing from the exact one. The two designs differ only in
whether clearing is keyed to the transition or to the state, and only
one of them is correct.
\end{remark}
\section*{What one horizon hides}
Restricting Proposition~\ref{prop:tlquotient} to $T=1$ was not a
simplification; it is what $\Gamma$ was defined to measure. But
Proposition~\ref{prop:periodicity} guarantees the longer-horizon
question has a definite, finitely checkable answer, and for this
system that answer is worth computing directly rather than left as an
abstract guarantee.
\begin{proposition}[The request bit is eventually redundant everywhere]
\label{prop:eventual}
For each phase $k$, the pair $(k,0)$ and $(k,1)$ satisfies
$R_T(k,0) = R_T(k,1)$ for all $T \geq 4 - k$, and not before.
\end{proposition}
\begin{proof}[Proof by direct computation]
Compute $R_T(k,0)$ and $R_T(k,1)$ for $T = 0, \dots, 4-k$ at each
phase directly from $\Psi$, rather than inferring later horizons from
earlier ones: a size mismatch at one horizon does not, on its own,
preclude equality reappearing or vanishing at a different horizon, so
each phase's first point of equality is verified independently here
rather than propagated.
For $k=3$: $R_0(3,0) = \{(3,0)\} \neq \{(3,1)\} = R_0(3,1)$;
$R_1(3,0) = R_1(3,1) = \{(0,0)\}$ (Proposition~\ref{prop:tlquotient}).
First equality at $T=1=4-3$.
For $k=2$: $R_0(2,0) = \{(2,0)\} \neq \{(2,1)\} = R_0(2,1)$;
$R_1(2,0) = \{(3,0),(3,1)\} \neq \{(3,1)\} = R_1(2,1)$ (both sets
computed directly from $\Phi$, not equal since one has two elements
and the other one); $R_2(2,0) = \Psi(\{(3,0),(3,1)\}) = \{(0,0)\} =
\Psi(\{(3,1)\}) = R_2(2,1)$. First equality at $T=2=4-2$.
For $k=1$: $R_0, R_1$ inequality follows the same way (cardinality
$2$ versus $1$ at $T=1$, distinct singletons at $T=0$); $R_2(1,0) =
\{(3,0),(3,1)\} \neq \{(3,1)\} = R_2(1,1)$, still unequal;
$R_3(1,0) = \Psi(\{(3,0),(3,1)\}) = \{(0,0)\} = \Psi(\{(3,1)\}) =
R_3(1,1)$. First equality at $T=3=4-1$.
For $k=0$: $R_0(0,0) = \{(0,0)\} \neq \{(0,1)\} = R_0(0,1)$, distinct
singletons; $R_1, R_2, R_3$ each pit a two-element set against a
one-element set, as computed directly from $\Phi$ at each step, so
none of the three can coincide; $R_4(0,0) = \{(0,0)\} = R_4(0,1)$.
First equality at $T=4=4-0$.
In every case the inequality at each horizon strictly before $4-k$
was checked directly from the explicit sets, not inferred from the
inequality at a different horizon.
\end{proof}
\begin{remark}
This is the sharper and more interesting fact the single-horizon
result was hiding. The request bit is not redundant in general; it is
redundant \emph{by a deadline}, and that deadline is exactly the
number of steps remaining until the light is guaranteed to cycle
through its clearing transition, four steps for a state that has just
left it, one step for a state about to enter it. Every state's
request bit stops mattering to its continuation space at precisely
the moment continuation stops being able to tell whether that request
will still be pending when clearing finally arrives, and not one step
before. The design clears the bit at the earliest possible moment,
$T=1$ after arrival, which is also the latest moment at which failing
to clear it would still have been observably wrong; every other
phase's bit becomes redundant on its own, without any design
intervention, simply because the light's own periodicity eventually
answers the question the bit was tracking.
\end{remark}
\section*{What quotients do not yet explain}
This chapter has shown when two histories are interchangeable from
the standpoint of every quantity built purely from continuation
spaces, and it has shown, in one worked case, that a designed
identification can be checked against that standard exactly rather
than merely defended by intuition. It has not yet said anything about
what is lost when a quotient is coarser than $\sim_\Gamma$, only that
the traffic light's particular quotient happens not to be. The next
chapter takes up the coarser case directly: not whether information is
destroyed, which this chapter's machinery can already detect, but how
much, and by what measure.
\chapter{Distinction Loss Under Quotients}
A library catalog entry is a quotient of a book: title, author, and a
handful of subject headings standing in for everything between the
covers. A good catalog loses nothing a patron actually needs in order
to decide whether to pull the book from the shelf. A bad one can lose
quite a lot; a catalog that records only ``fiction'' or
``nonfiction'' will happily file a memoir next to a novel and a field
guide next to a textbook, and a patron relying on that heading alone
has no way to tell them apart until the book is already in hand. The
catalog entry is not wrong, exactly; it simply was not built to
preserve the distinctions this particular patron cares about.
The previous chapter showed how to check whether a given
identification of histories is exact: whether it merges only
histories whose continuation spaces already coincide, or whether it
merges some that do not. That check was binary, exact or not. This
chapter asks the sharper question a librarian actually has to answer.
When an identification is not exact, how much is lost, and where
exactly does the loss occur? The traffic light's badly designed
alternative from the previous chapter turns out to have a precise
answer, not just a qualitative one.
\section*{Fibers and their content}
\begin{definition}[Fiber, $\Gamma$-content, loss]
For a quotient $q : H \to O$ and $o \in O$, the fiber over $o$ is
$q^{-1}(o) \subseteq H$. Its $\Gamma$-content is
$C_\Gamma(o) := \{\Gamma(h) : h \in q^{-1}(o)\}$, the set of distinct
continuation spaces appearing in the fiber. The loss at $o$ is
\[
\mathcal{L}(o) := \log |C_\Gamma(o)|,
\]
so $\mathcal{L}(o) = 0$ exactly when every history mapping to $o$
under $q$ shares a single continuation space.
\end{definition}
\begin{proposition}[Zero loss characterizes $\Gamma$-respecting quotients]
\label{prop:zeroloss}
$\mathcal{L}(o) = 0$ for every $o \in O$ if and only if $q$ refines
$\sim_\Gamma$: $h_1 \sim_q h_2 \implies h_1 \sim_\Gamma h_2$.
\end{proposition}
\begin{proof}
If $q$ refines $\sim_\Gamma$, then $h_1 \sim_q h_2$ implies
$\Gamma(h_1) = \Gamma(h_2)$, so every fiber of $q$ lies inside a
single $\sim_\Gamma$-class and has $|C_\Gamma(o)| = 1$, giving
$\mathcal{L}(o) = 0$. Conversely, if $\mathcal{L}(o) = 0$ for every
$o$, then every fiber has $|C_\Gamma(o)| = 1$, so any two histories in
the same fiber share the same $\Gamma$-value: $h_1 \sim_q h_2 \implies
\Gamma(h_1) = \Gamma(h_2) \implies h_1 \sim_\Gamma h_2$, so $q$
refines $\sim_\Gamma$.
\end{proof}
\section*{The coarsest lossless quotient}
\begin{theorem}[$\sim_\Gamma$ is the coarsest zero-loss quotient]
\label{thm:coarsest}
Every zero-loss quotient $q : H \to O$ refines $\sim_\Gamma$, and
$\sim_\Gamma$ itself has zero loss everywhere. Consequently
$\sim_\Gamma$ is, up to relabeling of $O$, the unique coarsest
quotient of $H$ with zero loss.
\end{theorem}
\begin{proof}
The first claim is Proposition~\ref{prop:zeroloss}. That
$\sim_\Gamma$ itself has zero loss is immediate: by definition,
$h_1 \sim_\Gamma h_2$ already means $\Gamma(h_1) = \Gamma(h_2)$, so
every $\sim_\Gamma$-class has exactly one $\Gamma$-value. Since every
zero-loss $q$ refines $\sim_\Gamma$, no zero-loss quotient can be
strictly coarser than $\sim_\Gamma$, and $\sim_\Gamma$ achieves zero
loss, so it is the coarsest one.
\end{proof}
\begin{remark}
This is the quotient-theoretic form of a fact Chapter 5 of
\emph{Continuation Geometry} established for sufficient statistics
without this vocabulary. A sufficient statistic $\varphi$ there was
required to satisfy $\varphi(H_t) = \varphi(H_t') \implies
\Gamma(x_t \mid H_t) = \Gamma(x_t \mid H_t')$, which is exactly the
condition that the quotient induced by $\varphi$ refines
$\sim_\Gamma$. Theorem~\ref{thm:coarsest} says $\sim_\Gamma$ itself is
the coarsest possible sufficient statistic: the most compressed
summary of a history that still loses nothing continuation-relevant.
Any coarser summary is not merely less detailed; by
Proposition~\ref{prop:zeroloss}, it has strictly positive loss
somewhere.
\end{remark}
\begin{proposition}[Coarsening never decreases loss]
\label{prop:monotone-ch3}
If $q'$ refines $q$ (every $q'$-fiber is contained in some $q$-fiber),
then for every $q$-fiber $q^{-1}(o)$, expressed as a union of
$q'$-fibers $q'^{-1}(o'_1), \dots, q'^{-1}(o'_k)$,
\[
\mathcal{L}_q(o) \geq \max_i \mathcal{L}_{q'}(o'_i).
\]
\end{proposition}
\begin{proof}
$C_\Gamma(o) = \bigcup_i C_{\Gamma}(o'_i)$, since the $\Gamma$-values
appearing in $q^{-1}(o)$ are exactly those appearing in its
constituent $q'$-fibers. A union of sets is at least as large as any
one of them, so $|C_\Gamma(o)| \geq |C_\Gamma(o'_i)|$ for each $i$,
and $\log$ is monotone increasing.
\end{proof}
\begin{remark}
Merging two fibers can only add to the set of continuation spaces an
observer sees when looking at the merged fiber; it can never remove
one. This is the precise sense in which coarsening a quotient is a
one-way trip toward loss: refining a lossy quotient can repair the
loss, by splitting a mixed fiber back into pieces that each see only
one $\Gamma$-value, but no further coarsening of an already-lossy
quotient can undo the damage, only add to it or leave it unchanged.
\end{remark}
\section*{Distinction margin as a local loss detector}
\begin{proposition}[Vanishing margin detects a redundant distinction]
For a distinction $\mathcal{D}$ and $x \in K$, $A_{\mathcal D}^T(x) = 0$
(Chapter 7 of \emph{Continuation Geometry}) if and only if some $y$
outside $x$'s $\mathcal{D}$-cell satisfies $R_T(x) = R_T(y)$: that is,
$\mathcal{D}$ places $x$ and $y$ in different cells even though their
continuation spaces coincide at horizon $T$. Vanishing margin is a
diagnostic for $\mathcal{D}$ being \emph{finer} than $\sim_\Gamma$ at
$x$, drawing a distinction continuation geometry does not support, not
for $\mathcal{D}$ being coarser.
\end{proposition}
\begin{proof}
Recall $A_{\mathcal D}^T(x) := \inf_{y \notin D(x)} d_H(R_T(x), R_T(y))$.
This infimum is $0$ exactly when some $y$ outside $x$'s
$\mathcal{D}$-cell has $R_T(x) = R_T(y)$ (attained, under the
compactness hypothesis already used in Chapter 7), which is exactly
the statement that $x$ and that $y$ share a $\Gamma$-value at horizon
$T$ despite $\mathcal D$ having placed them in different cells: they
are $\sim_\Gamma$-equivalent even though $\mathcal D$ distinguishes
them. This is over-refinement, not collapse.
\end{proof}
\begin{definition}[Within-cell spread]
For a $\mathcal{D}$-cell $D$, $\operatorname{diam}_\Gamma(D) :=
\sup_{x,y \in D} d_H(R_T(x), R_T(y))$.
\end{definition}
\begin{proposition}[Zero spread is the actual collapse-safety detector]
\label{prop:diamzero}
$\operatorname{diam}_\Gamma(D) = 0$ if and only if $D$ is contained in
a single $\sim_\Gamma$-class at horizon $T$, i.e.\ $\mathcal{D}$
introduces no loss in the sense of the fiber-loss definition earlier
in this chapter, applied to $D$ as a fiber.
\end{proposition}
\begin{proof}
$\operatorname{diam}_\Gamma(D) = 0$ means $d_H(R_T(x),R_T(y)) = 0$ for
every $x,y \in D$, i.e.\ $R_T(x) = R_T(y)$ for every pair in $D$, i.e.\
$D$ lies inside one $\sim_\Gamma$-class; this is exactly
$|C_\Gamma(D)| = 1$ from the definition of fiber loss, i.e.\
$\mathcal{L}(D) = 0$.
\end{proof}
\begin{remark}
This gives the distinction margin of the first book a second reading,
and a corrective one. It was introduced there as a measure of how
close a state sits to losing a distinction, language that invites
exactly the confusion the propositions above are meant to head off.
$A_{\mathcal D}^T(x) = 0$ is not a report of quotient loss; it is a
report that $\mathcal{D}$ is drawing a line continuation geometry
does not, distinguishing two states with identical futures. Quotient
loss, in the sense the rest of this chapter measures it, is instead
the province of $\operatorname{diam}_\Gamma$: a cell $D$ with
$\operatorname{diam}_\Gamma(D) > 0$ is the one actually merging states
with different futures, the failure mode Chapter 3 of this book was
built to quantify. The two diagnostics point in opposite directions,
over-refinement and under-refinement, and a single distinction
$\mathcal{D}$ can exhibit either, both, or neither, independently at
different points of $K$.
\end{remark}
\section*{The traffic light, quantified}
Return to the two designs compared in the previous chapter: the
correct one, clearing on arrival at phase $0$, and the alternative,
clearing whenever currently at phase $0$.
\begin{example}[Zero loss, and one bit]
The correct design's quotient is $\sim_\Gamma$ itself (Proposition~4
of the previous chapter), so by Theorem~\ref{thm:coarsest} its loss is
$\mathcal{L}(o) = 0$ at every fiber $o$. The alternative design merges
$(0,0)$ and $(0,1)$ into a single fiber $o_{\mathrm{bad}}$. Their
$\Gamma$-values at $T=1$ are $R_1(0,0) = \{(1,0),(1,1)\}$ and
$R_1(0,1) = \{(1,1)\}$, which are distinct, so
$C_\Gamma(o_{\mathrm{bad}}) = \{\{(1,0),(1,1)\},\ \{(1,1)\}\}$, a set
of size $2$, giving
\[
\mathcal{L}(o_{\mathrm{bad}}) = \log 2,
\]
exactly one bit, and $\mathcal{L}(o) = 0$ at every other fiber, since
the alternative design agrees with the correct one everywhere except
at this single merge.
\end{example}
\begin{remark}
This is what the previous chapter's qualitative verdict, that the
alternative design was coarser and information-destroying, actually
amounts to once measured: not a diffuse degradation spread across the
whole system, but a single, exactly located bit of loss, sitting at
precisely the fiber where the alternative design merged two states
that a patron of this particular library, a controller trying to
plan four steps ahead, would still have wanted to tell apart. Every
other fiber of the alternative design is, by coincidence or by the
limited scope of its one design error, exactly as good as the correct
one.
\end{remark}
\section*{What loss does not yet explain}
This chapter can say, precisely, how much is lost by a given
quotient and exactly where. It has nothing to say about why a
particular coarser quotient might be chosen anyway, in full knowledge
of the loss it incurs. A designer might prefer a one-bit-lossy
traffic light if the saved bit of memory matters more than the
distinction it would have preserved; a librarian might prefer the
cheap two-category catalog if most patrons never need the finer one.
Measuring loss and deciding whether a loss is worth paying for are
different questions, and this chapter has only ever answered the
first of them.
\chapter{Continuation Operators}
The first three chapters of this book have all, in their own way, been
about nouns: a shape a system of contractions converges to, a
partition of histories that share a future, a number measuring what a
partition throws away. None of them named the verb. A contraction
does something to a shape. A design decision, clearing a bit on
arrival at a red light rather than while sitting at one, does
something to a history. This chapter is about the doing: what counts
as a legitimate move in this framework, and what survives when two
different moves are composed into one.
The two kinds of move already used in this book look, on the surface,
like opposites. Chapter 1's contractions built detail up, each
application adding a smaller, self-similar copy of the whole. Chapters
2 and 3's quotients tore detail down, collapsing histories that had
become, for every purpose this book has formalized, the same. It
would be a missed opportunity to leave them looking like two
unrelated tools. They are both instances of one idea, and this
chapter's job is to say what that idea is.
\section*{Operators and their composition}
\begin{definition}[Continuation operator]
A continuation operator on $X$ is a Lipschitz map
$T : \mathcal{K}(X) \to \mathcal{K}(X)$: there is a constant
$L \geq 0$ with $d_H(T(A), T(B)) \leq L\, d_H(A,B)$ for all
$A, B \in \mathcal{K}(X)$. Write $L(T)$ for the smallest such constant.
\end{definition}
\begin{proposition}[Closure under composition]
\label{prop:closure}
If $T_1, T_2$ are continuation operators, so is $T_2 \circ T_1$, with
$L(T_2 \circ T_1) \leq L(T_2)\, L(T_1)$.
\end{proposition}
\begin{proof}
$d_H(T_2(T_1(A)), T_2(T_1(B))) \leq L(T_2)\, d_H(T_1(A), T_1(B)) \leq
L(T_2)\, L(T_1)\, d_H(A,B)$.
\end{proof}
\begin{remark}
This proof is one line because it has to be; the definition was
chosen precisely so that composition would cost nothing further to
verify. What makes the definition worth having is not this
proposition but what it is a common home for. The contractions of
Chapter 1 are continuation operators with $L(T) < 1$. The map
$\Psi(A) := \bigcup_{z \in A} \Phi(z)$ used throughout Chapters 2 and
3 to advance a reachable set by one step is a continuation operator
too, generally with $L(\Psi) \geq 1$: it need not contract, but it is
still Lipschitz whenever the underlying one-step map $\Phi$ is well
behaved. Both belong to the same category. What distinguishes a
generative operator from a reductive one is not the definition above;
it is what happens under repeated composition, contraction toward a
single attractor in one case, and, as the next section makes precise,
collapse of distinctions in the other.
\end{remark}
\section*{Operators on histories, and what they respect}
The operators above act on shapes directly. A different and equally
natural kind of operator acts on histories: appending one more real
transition, editing a codebase, applying a repair. The interesting
question for an operator of this kind is not its Lipschitz constant
but whether it interacts predictably with $\sim_\Gamma$.
\begin{definition}[$\Gamma$-compatible operator]
$T : H \to H$ is $\Gamma$-compatible if there is a map
$\hat{T} : \mathcal{K}(X) \to \mathcal{K}(X)$ with
$\Gamma(T(h)) = \hat{T}(\Gamma(h))$ for every $h \in H$: the
continuation space of the transformed history depends on the original
history only through its continuation space, not otherwise.
\end{definition}
\begin{proposition}[$\Gamma$-compatible operators descend to the quotient]
\label{prop:descend}
If $T$ is $\Gamma$-compatible, then $h_1 \sim_\Gamma h_2 \implies
T(h_1) \sim_\Gamma T(h_2)$: $T$ induces a well-defined map
$\bar{T} : H/{\sim_\Gamma} \to H/{\sim_\Gamma}$.
\end{proposition}
\begin{proof}
This is Proposition~4 of the previous chapter (quotient invariance),
applied to $F := \hat{T} \circ \Gamma$ in place of $\Gamma$ itself:
$h_1 \sim_\Gamma h_2$ gives $\Gamma(h_1) = \Gamma(h_2)$, hence
$\Gamma(T(h_1)) = \hat{T}(\Gamma(h_1)) = \hat{T}(\Gamma(h_2)) =
\Gamma(T(h_2))$, so $T(h_1) \sim_\Gamma T(h_2)$.
\end{proof}
\begin{remark}
This is the sense in which $\Gamma$-compatibility is the right
condition to ask for: it says an operator cannot secretly depend on
information that continuation equivalence has already declared
irrelevant. An operator failing this condition would be able to treat
two continuation-equivalent histories differently for reasons that
have nothing to do with what either history can still become, which
is exactly the kind of dependence Chapter 5 of \emph{Continuation
Geometry} worried about under the name of history-dependence in the
first place.
\end{remark}
\begin{example}[The traffic light's step operators]
For $e \in \{0,1\}$, let $T_e : H \to H$ append the transition
$\Phi(\cdot, e)$ to a history. Whether $T_e$ is $\Gamma$-compatible
(at horizon $T=1$) needs checking only on the one nontrivial
$\sim_\Gamma$-class found in Chapter 2, $\{(3,0),(3,1)\}$, since every
other class is a singleton and compatibility there is automatic.
Direct computation gives $T_0(3,0) = T_0(3,1) = (0,0)$ and
$T_1(3,0) = T_1(3,1) = (0,0)$: both step operators send the two
equivalent histories to the identical resulting state, not merely to
states with equal $R_1$, so compatibility holds here in the strongest
possible form.
\end{example}
\begin{remark}
This is a narrow check, confined to the single class where there was
anything to check, and it should not be read as a general fact about
$T_0$ or $T_1$ on this system, only as a confirmation that neither
step operator introduces a new dependence on which of $(3,0)$ or
$(3,1)$ a history happened to pass through, at exactly the point
where such a dependence would have mattered.
\end{remark}
\section*{The family $\Gamma_T$ and the operator $\Psi$}
A different relationship, easy to conflate with $\Gamma$-compatibility
but distinct from it, has been used without a name since Chapter 2. It
concerns not an operator on histories but the relationship between the
finite-horizon continuation spaces at consecutive horizons for one and
the same history.
\begin{remark}[$\Psi$ relates horizons, not histories]
By construction, $R_{T+1}(x) = \bigcup_{z \in R_T(x)} \Phi(z) =
\Psi(R_T(x))$, i.e.\ $\Gamma_{T+1} = \Psi \circ \Gamma_T$ as an
identity between functions $H \to \mathcal{K}(X)$, where
$\Gamma_T(h) := R_T(x(h))$. This is not an instance of
Proposition~\ref{prop:descend}; no operator on $H$ is involved, only
the fixed continuation operator $\Psi$ acting on the shape side, once
for each increment of the horizon. It is this identity, not any
property of an operator on histories, that did the real work in
Lemma~2 of the previous chapter (one-step agreement propagates) and in