The Geometry of Endless Discovery

1.Traversal Space Configuration

Let $K$ be a state alphabet. A traversal $L$ over $K$ is an infinite sequence of states:

Infinite sequences over a discrete alphabet form the primitive object of symbolic dynamics.12 In that classical setting, the alphabet is fixed and the dynamics are shifts acting on sequences. Here we instead fix one realized traversal and ask what its asymptotic structure forces.

1.1 Finite Dwelling vs. Infinite Halting

A traversal may remain in a single state for a strictly finite number of consecutive steps. This is a finite dwell of length $m < \infty$, written $k^m$. Halting is defined as an infinite dwell, where the traversal enters a single state and resides there permanently. Formally, $L$ halts at index $N$ if

This is equivalent to an infinite extension of a single state, denoted $k^\omega$.

1.2 Infinite Traversal

An infinite traversal consists of an infinite succession of finite state-residence intervals. The sequence is represented as runs:

where every residence time $r_j \in \mathbb{N}$ is finite ($r_j < \infty$) and the number of runs is infinite. The adjective infinite modifies the continuation of transitions, not the duration of occupancy. Formally, an infinite traversal satisfies

This says: however long a finite dwell becomes, there is eventually another transition.

1.3 Finite Alphabets Force Recurrence

If $K$ is finite, an infinite traversal forces at least one state to recur infinitely often, by the pigeonhole principle.1 For the binary case $K = \{S, A\}$, the decomposition into alternating runs is

Because neither $S$ nor $A$ can become an infinite final residence, both necessarily recur infinitely often:

This uses only the non-halting condition and $|K|=2$; no further hypothesis on the run lengths is needed.

2.Infinite Alphabets and the Separation of Recurrence and Novelty

If $K$ is countably infinite, $K = \{k_1, k_2, k_3, \dots\}$, the implication that infinite traversal forces recurrence no longer holds. An infinite traversal may satisfy the non-halting condition by continually escaping through novel states rather than revisiting old ones.

This is the first conceptual fork in the document. Over a finite alphabet, continued motion eventually reuses available symbols because there are only finitely many. Over an infinite alphabet, continued motion has another option: spend a new symbol at every step. The question changes from “must something return?” to “what kind of structure does continual discovery produce?”

2.1 Zero-Recurrence Traversals

Consider the traversal $L = (k_1, k_2, k_3, k_4, \dots)$. Every residence is finite (specifically, $r_j = 1$). The traversal never halts. Yet every state occurs exactly once:

Define the exact point recurrent support as

For this traversal, $R_{\text{pt}}(L) = \emptyset$. Two concepts that coincided in the finite-alphabet case now separate: unbounded traversal and recurrence. The traversal can remain infinite by consuming novel states.

2.2 Cardinality of Zero-Recurrence Traversals

The phenomenon is structurally large. For a countably infinite alphabet, the space of all infinite traversals is $K^\omega$. Its cardinality is2

There are continuum-many distinct infinite zero-recurrence traversals over a countably infinite alphabet. This is established Cantor theory, not novel construction.7

3.Measure and Topological Extensions of Recurrence

Cardinality alone does not describe the geometric novelty of a traversal. A countably infinite alphabet can be represented as a discrete escaping set $\{1, 2, 3, \dots\}$, an accumulating set $\{1, \tfrac{1}{2}, \tfrac{1}{3}, \dots\}$, or a dense set $\mathbb{Q} \cap [0,1]$. All have cardinality $\aleph_0$. To analyze novelty properly, $K$ requires additional structure: a metric $d$, a topology $\tau$, or a measure $\mu$.

3.1 Nonatomic Measure Spaces

Let $(K, \Sigma, \mu)$ be a nonatomic measure space, such as $([0,1], \mathcal{B}, \lambda)$ with Lebesgue measure. In such a space, individual states satisfy $\mu(\{x\}) = 0$. This is the natural setting in which the classical recurrence theorem of Poincaré and Birkhoff applies.1314

A traversal can therefore encounter a different exact state at every step ($x_i \neq x_j$ for $i \neq j$) while remaining inside the same finite-measure region. For independent continuous sampling, exact repeats occur with probability zero. However, positive-measure regions can be revisited repeatedly.

For independent sampling, the regional-return statement follows from repeated positive-probability trials. Poincaré recurrence is the dynamical analogue: in a finite-measure system whose evolution preserves measure, almost every point that begins in a measurable region returns to that region infinitely often. The theorem does not say that one countable trajectory hits every positive-measure measurable set; indeed, the complement of its countable range can still have full measure. The modern dynamical treatments are due to Furstenberg,4 Walters, and Katok–Hasselblatt.15 For the topological recurrence viewpoint used in this section, see Alongi and Nelson.61

The beginner's takeaway is simple: “this exact point appeared again” is a microscopic statement, while “the trajectory returned to this region” is a finite-resolution statement. Nonatomic spaces let the second happen repeatedly without the first ever happening.

3.2 Limit Sets and Accumulation

Define the topological limit (or $\omega$-limit) set of a traversal as

Consider three traversals over $K = [0,1]$:

• $x_n = 1, 2, 1, 2, \dots$ on a discrete $\{1,2\}$: $R_{\text{pt}}(L) = \{1,2\}$, $\omega(L) = \{1,2\}$ — exact recurrence.
• $x_n = 1/(n+1)$: $R_{\text{pt}}(L) = \emptyset$, but $\omega(L) = \{0\}$ — interior accumulation.
• A dense non-repeating traversal of $[0,1]$: $R_{\text{pt}}(L) = \emptyset$, but $\omega(L) = [0,1]$.
• $x_n = n$ in $\mathbb{R}$: $R_{\text{pt}}(L) = \emptyset$ and $\omega(L) = \emptyset$ — escape from every compact subset.

Zero exact recurrence does not determine asymptotic geometry. Novelty requires asymptotic structure, not merely the counting of new symbols.

4.The Geometry of Novelty

To formalize novelty in a metric space $(K, d)$, define the distance of the current state from its previous history at step $n$:

(Here $\inf \varnothing := +\infty$, equivalently $d(x, \varnothing) := +\infty$, so $\delta_0 = +\infty$: the opening state is novel at every resolution.)

Exact novelty requires $x_n \notin \{x_0, \dots, x_{n-1}\}$, i.e., $\delta_n > 0$. Geometric novelty asks how large $\delta_n$ is. A state is novel at resolution $\varepsilon > 0$ when $\delta_n > \varepsilon$.

4.1 Resolution-Dependent Novelty

Let $P(K, \varepsilon)$ denote the packing number of $K$, the maximum cardinality of an $\varepsilon$-separated subset:

Picture placing non-overlapping safety bubbles around discovered points. If every pair of centers must stay more than $\varepsilon$ apart, then bubbles of radius roughly $\varepsilon/2$ cannot overlap. The packing number asks for the largest number of such mutually separated centers that fit inside $K$. On $[0,1]$, the answer grows on the order of $1/\varepsilon$: coarse resolution admits only a few distinguishable locations, while fine resolution admits many.

Compactness forces arbitrarily close approximation but not exact repetition. This is the standard finite-resolution counterpart of the pigeonhole argument. The quantity $P(K,\varepsilon)$ is the input to the entropy functionals of Kolmogorov–Sinai (measure-theoretic) and Adler–Konheim–McAndrew / Dinaburg–Bowen (topological), which quantify information-production rates of dynamics at scale $\varepsilon$.5161718 Standard dimension texts treat the covering, packing, and separated-set estimates behind this resolution picture.495051

4.2 The Three Regime Classification

An infinite traversal over a metric alphabet $K$ falls into one of three mutually exclusive classes:

1. Exact recurrence: $R_{\mathrm{pt}}(L) \neq \varnothing$. At least one state is revisited infinitely often.
2. Novelty with accumulation: $R_{\mathrm{pt}}(L) = \varnothing$ but $\omega(L) \neq \varnothing$. Nothing repeats exactly, yet some subsequence converges. This class includes accumulation at a point ($x_n = 1/(n+1)$, $\omega(L) = \{0\}$), on a proper subset, and on the whole space (a dense non-repeating traversal of $[0,1]$, $\omega(L) = [0,1]$).
3. Escape: $R_{\mathrm{pt}}(L) = \varnothing$ and $\omega(L) = \varnothing$. No subsequence converges; the trajectory eventually leaves every compact subset of $K$ ($x_n = n$ in $\mathbb{R}$).

These are qualitatively different asymptotic behaviors. The first forces replication; the second can occur within a compact region; the third requires noncompact room to leave.

Two calibrations. On compact $K$, escape is impossible — sequential compactness gives $\omega(L) \neq \varnothing$ for every traversal — and Theorem 4.1 sharpens regime 2: a zero-recurrence traversal on compact $K$ satisfies $\liminf_{n \to \infty} \delta_n = 0$. Uniform separation is sufficient but not necessary for escape: $\delta_n \ge c > 0$ makes the sequence $c$-separated, so no subsequence converges; but $x_n = \log(n+1)$ escapes with $\delta_n = \log(1 + 1/n) \to 0$. Escape is the absence of accumulation, not a lower bound on the novelty gap.

5.Occurrence Space and Replication

For traversals that do exhibit recurrence ($R_{\text{pt}}(L) \neq \emptyset$), state identity alone is insufficient to identify an occurrence. The system requires an occurrence space $\mathcal{O}$ to track individual instances of recurrent states.

For a recurrent state $k \in R_{\text{pt}}(L)$, the occurrence space contains an infinite set of indexed instances:

The total occurrence space is the union over all recurrent states:

5.1 Projection and Distinguishability

Define a projection map $\pi : \mathcal{O} \to K$ with $\pi(k_i) = k$. Elements in $\mathcal{O}$ share state identity but maintain distinct occurrence identities. For any two occurrences $k_i, k_j \in \mathcal{O}$ with $i \neq j$:

The traversal $L$ induces a strict total order $<_L$ on $\mathcal{O}$. Representation must therefore preserve three logically distinct pieces of information:

6.Canonical Prefix Geometry

The traversal should first generate its own geometry. Hyperbolic realization is a later question, not an input.

Let $K$ be a discrete alphabet and define the prefix tree

The empty word $\varnothing$ is the root. A vertex $u=(x_0,\ldots,x_{n-1})$ is joined by an edge of length one to each one-symbol extension $u^\frown x_n$. Thus a traversal $L=(x_0,x_1,\ldots)$ determines the ray

At depth $n+1$, the vertex is exactly the accessible prefix $I_n(L)=(x_0,\ldots,x_n)$. Order is encoded by depth, and distinct finite histories are distinct vertices. No auxiliary state labels or arbitrarily chosen geodesics are needed.

Why a tripod? Choose any three vertices $u,v,w$. The unique paths between them share a central branching vertex $b$. Traveling from $u$ to $v$, from $v$ to $w$, and from $w$ back to $u$ merely traverses the three arms meeting at $b$. There is no open triangular interior. In the language of hyperbolic geometry, the triangle is as thin as possible: it is $0$-thin.

6.1 Rays and the Canonical Boundary

Every infinite word selects exactly one successor at each depth, and every rooted ray determines exactly one infinite word. Hence the endpoint map

is a canonical bijection. This repairs the central defect of the earlier direct embedding into $\mathbb H^2$: the tree geometry itself remembers the complete traversal.

The word boundary can be misleading on first encounter. The endpoint $\xi_L$ is not an additional vertex sitting one step after every finite prefix. There is no “last edge” from a very distant vertex to $\xi_L$. Instead, $\xi_L$ is the equivalence class of the entire ray: it records the direction in which the rooted path continues forever. Boundary points describe infinite behavior rather than locations reached in finite time.

6.2 Cylinders and the Prefix Metric

For a finite word $u\in K^{<\omega}$, the boundary points compatible with $u$ form the cylinder

If $u_N=I_N(L)$, then finite information identifies $[u_N]$, not a single endpoint. The cylinders refine exactly:

Let $\ell(L,L')$ be the number of symbols in the longest common prefix, with $\ell(L,L)=\infty$. For any $a>1$, define

Convention. All sequences and all resolution levels in this document are indexed from $0$, and agreement length always means the first disagreement index: $\ell(x,y) = \min\{j : x_j \neq y_j\}$, with $\min \varnothing := \infty$. Agreement through index $N$ is therefore $\ell \ge N+1$ and distance at most $a^{-(N+1)}$.

This nested-ball geometry is also the basic language of ultrametric analysis; the $p$-adic identification in §7.1 is one arithmetic instance of the same pattern.52

6.3 Halting as a Boundary Property

The boundary $\partial T_K$ contains every infinite sequence, including eventually constant ones such as $(S,A,A,A,\ldots)$. The non-halting traversal space is therefore the boundary subset

Recurrence does not create the boundary; it classifies particular rays inside it. For $K=\{S,A\}$, a non-halting ray has infinitely many finite runs and therefore alternates between $S$-runs and $A$-runs indefinitely.

7.Boundary Topology and Alphabet Size

The topology of $\partial T_K\cong K^\omega$ changes sharply with the branching capacity of the alphabet. That transition is the first rigorous connection between symbolic novelty and boundary geometry.

Compactness is best read here as a statement about escaping descriptions. In a compact boundary, no sequence of candidate completions can keep fleeing every possible limiting pattern; some subsequence must stabilize prefix by prefix. Noncompactness means that even the first symbol can keep changing so that no such stabilization begins.

7.1 Finite Alphabets: Cantor Space

Suppose $K$ is finite and $2\le |K|=q<\infty$. Then $T_K$ is the rooted $q$-ary tree, with $q^n$ vertices at depth $n$. The product space $K^\omega$, equivalently the prefix-ultrametric boundary, is compact, totally disconnected, and perfect. Up to homeomorphism it is Cantor space.25

This is a diagonal form of compactness: finite branching permits the repeated extraction of an infinite subsequence through one common successor. It does not say that individual symbols cannot be novel for long stretches; it says the space of all infinite completions is topologically bounded.

Exact anchor: the $p$-adic integers

The prefix-boundary construction already contains a standard number-theoretic object. Fix a prime $p$ and take the digit alphabet

For a finite word $w=(x_0,\ldots,x_{n-1})\in K_p^n$, define

This is a bijection. More importantly, it respects the parent maps: deleting the newest tree digit $x_n$ corresponds exactly to reducing a residue modulo $p^{n+1}$ to one modulo $p^n$. Thus the diagram

commutes. Passing to inverse limits gives the exact chain

The connection between the full $p$-ary tree, infinite words, nested $p$-adic balls, and $\mathbb Z_p$ is classical.31 It is an exact example of the raw prefix boundary developed in §6, not an application of the ordered-novelty construction in §10.

When the digit stream is generated by a finite automaton, this address-space picture also meets the theory of automatic sequences: finite-state generation can constrain the infinite word without changing the underlying prefix boundary.53

If two digit rays first disagree at index $r$, their images agree modulo $p^r$ but not modulo $p^{r+1}$. Therefore

and Proposition 6.2 with base $a=p$ becomes the $p$-adic metric exactly:

The profinite integers provide a broader companion example:

where the index set ranges over all positive moduli ordered by divisibility and the product decomposition follows from the Chinese remainder theorem. Unlike a single $\mathbb Z_p$, this is not literally the boundary of one fixed finite-alphabet prefix tree; it assembles all prime-power completions into a profinite product. Its role here is to demonstrate the broader inverse-limit pattern, while $N_P(K)$ remains the traversal-specific construction whose coordinates are ordered finite-resolution discoveries.

7.2 Countable Alphabets: Baire Space

For $K=\mathbb N$ with the discrete topology, every vertex has countably many children and

the classical Baire space. It is completely metrizable, separable, and totally disconnected, but not compact.25 Indeed, the first-symbol cylinders $[n]$ form an open cover with no finite subcover:

8.Novelty Quotients

Prefix completion and novelty identification are distinct operations. First, the full traversal determines its canonical endpoint $\xi_L\in\partial T_K$. Only then does a chosen invariant identify endpoints that carry the same asymptotic novelty information:

Let $\mathcal N_\infty$ be an asymptotic invariant and define

The resulting space is

8.1 Closed Equivalence Relations

For finite $K$, the raw boundary $K^\omega$ is already compact metric. The quotient is well behaved precisely when novelty equivalence does not acquire new identifications only in the limit. Write

Closedness prevents a limiting-behavior trap. Suppose pairs $(L_m,L'_m)$ are equivalent for every $m$ and converge to $(L,L')$. If the relation is closed, the limiting pair remains equivalent. If it is not closed, classes that looked consistently glued together can tear apart at the limit, and the quotient may fail to distinguish points topologically.

8.2 Novelty Compactification Criterion

For a countably infinite alphabet, quotienting can change the decisive topological property. Begin with the noncompact Baire boundary $K^\omega$. Suppose there is a compact metric target $Z$ and a continuous surjective quotient map

In everyday terms, a quotient map guarantees that the topology on $Z$ is exactly the topology seen after collapsing the fibers of $\mathcal N_\infty$. It rules out giving the target an unrelated topology after the fact. The theorem is therefore not merely saying “there is a surjection onto a compact set”; it says the compact set is genuinely the topological space of equivalence classes.

9.Proper Hyperbolic Realization

Only after the canonical boundary and the novelty quotient have been defined should one ask whether the result occurs as the Gromov boundary of another space.

9.1 The Compactness Obstruction

Therefore a noncompact novelty quotient cannot be the boundary of a proper geodesic hyperbolic space. In particular,

cannot itself be such a boundary. There is no contradiction: $T_{\mathbb N}$ is hyperbolic but not proper, exactly as Proposition 7.2 records.

Properness is the local finiteness condition needed here: every closed bounded region must be compact. A countably branching root already violates it, because the radius-one ball contains infinitely many children separated from one another. The noncompact Baire boundary is the asymptotic shadow of that local explosion. A proper hyperbolic space cannot have such a noncompact Gromov boundary.

9.2 Realization After Quotienting

Combining Theorems 8.1 or 8.2 with Theorem 9.2 gives the rigorous pipeline

Hyperbolicity does not solve infinite novelty. The invariant first compresses the raw, possibly noncompact completion space into a compact asymptotic object; hyperbolic geometry can then realize that object.

The hyperbolic-approximation theorem can be read constructively: start from a compact metric space $Z$, place vertices representing sets visible at successively finer scales, and connect fine sets to nearby coarse parents. The resulting graph expands resolution into depth. Its large-scale geometry is hyperbolic, and its ideal boundary reconstructs $Z$. In Section 10, the ordered novelty construction becomes even more explicit because its compatible finite profiles already form the required tree.

9.3 What the Poincaré Disk Represents

For finite $K$, the regular tree $T_K$ admits a quasi-isometric realization by a bounded-valence tree in $\mathbb H^2$. Its boundary is Cantor space, whereas $\partial\mathbb H^2\cong S^1$. Consequently the traversal boundary occupies a Cantor-like subset of the circle at infinity, not the entire circle.9 Beardon is a standard reference for the Möbius, Fuchsian, and Poincaré-disk geometry behind this picture.57

10.Candidate Novelty Invariants

The remaining mathematical choice is substantive: what asymptotic information should survive quotienting? Different invariants retain different geometries.

10.1 Empirical Measures

For a traversal in a compact metric alphabet $K$, define the empirical measures

($\delta_x$ is the Dirac mass at $x$; the novelty gap of §4 never appears with a point subscript.)

On the domain $\mathcal T_\mu\subseteq K^\omega$ where $\mu_N^L$ converges weakly, let $\mathcal N_\infty(L)=\mu_L$. Equality of $\mu_L$ forgets order and exact path while retaining asymptotic occupancy frequencies. The target $\mathcal P(K)$ of Borel probability measures is compact metrizable in the weak topology when $K$ is compact metric.27

Read $\mu_N^L$ as a histogram made from the first $N$ observations. It assigns mass $1/N$ to each observation, counting repeats with multiplicity. Weak convergence means that, when tested against continuous measurements, these histograms settle toward a stable probability distribution. Shuffling a long block barely changes the histogram, which is precisely why temporal order disappears.

10.2 Accumulation Sets

Another invariant retains where the path accumulates:

If $K$ is compact metric, then $\omega(L)$ is a nonempty compact subset of $K$. The hyperspace $\mathcal K(K)$ of nonempty compact subsets, equipped with the Hausdorff metric, is compact metric.28 This quotient forgets visitation frequencies but retains the topological locus of recurrent approximation.

For $L=(1,1/2,1/3,\ldots)$, the accumulation set is only $\{0\}$. It remembers the destination of the shrinking tail but forgets the entire ordered ladder $1,1/2,1/3,\ldots$. A different sequence converging to $0$ receives the same accumulation set. That is useful when only limiting location matters and unacceptable when discovery order is the object of study.

10.3 The Ordered Novelty Process

Empirical measures and accumulation sets destroy the temporal order of discovery. To preserve it, define the first-discovery indices recursively. Put $\tau_0=0$ and, whenever another unseen state exists, let

The ordinary clock advances at every $n$. The discovery clock advances only at the indices $\tau_m$. If

then the discovery clock reads $\tau_0=0$, $\tau_1=3$, $\tau_2=6$, $\tau_3=8$, and the corresponding spine is $(a,b,c,d,\ldots)$. Repeated dwell and revisits remain part of the original traversal but vanish from this particular summary.

The ordered discovery spine is the sequence of states seen for the first time:

For metric $K$, attach the geometric scale of each discovery: for $m \ge 1$,

and $\delta_0 = d(x_{\tau_0}, \varnothing) = +\infty$ by the convention of §4. Every state occurring before $\tau_m$ is a repeat of an earlier discovery, so $\{x_j : j < \tau_m\} = \{x_{\tau_0}, \ldots, x_{\tau_{m-1}}\}$ as sets, and therefore $\delta_m = \delta_{\tau_m}$: the discovery-scale sequence is the nearest-history gap sequence of §4 sampled at the discovery clock.

The resulting ordered novelty object is

This core object preserves which novelty appears, its order, and its geometric separation while discarding redundant revisits. If discovery pace must also survive, augment it with the gaps $g_m:=\tau_m-\tau_{m-1}$. Keeping $g_m$ distinguishes a rapid sequence of discoveries from the same discoveries separated by long dwell periods; omitting it deliberately quotients those dwell periods out.

This distinction prevents a subtle overclaim. The sequence $(\delta_m)$ measures geometric separation between successive discoveries; it does not by itself say how quickly discoveries occur in ordinary time. Pace belongs to $(g_m)$ or to the two-variable statistic $M_L(\varepsilon,N)$ introduced later. Geometry, order, and rate are three different coordinates of novelty.

The obstruction concerns the prefix topology, not ordered novelty itself. A compact encoding becomes possible once depth records spatial resolution rather than elapsed time.

10.4 Resolution-Depth Completion

Let $(K,d)$ be compact metric. Fix a refining partition tower $\mathcal P = (\mathcal P_j)_{j\ge 0}$: finite partitions with $\mathcal P_{j+1}$ refining $\mathcal P_j$, parent maps $p_j : \mathcal P_{j+1} \to \mathcal P_j$, and $\operatorname{mesh}(\mathcal P_j) \le \varepsilon_j \downarrow 0$. Everything constructed in §§10.4–10.6 depends on $\mathcal P$. We write $N_P(K)$ and $C^P_{\mathrm{nov}}$ when the dependence matters and suppress the superscript while a single tower is fixed.

At resolution $j$, scan $L$ chronologically and retain a cell only on its first occurrence. The resulting finite injective word

records the order in which new $\varepsilon_j$-regions are discovered. Dwell and revisits disappear, but discovery order remains. A fixed-$\varepsilon$ discovery tree cannot do this job: total boundedness bounds every $\varepsilon$-separated branch by the packing number, so that tree has finite depth and empty boundary. The scale must refine with depth.

Given a word over $\mathcal P_{j+1}$, apply $p_j$ to each letter and delete every repeated parent after its first appearance. Denote this parent-and-reduce operation by $q_{j+1,j}$. First entry into a coarse cell is necessarily first entry into some previously unseen fine cell inside it; hence

Let $X_j$ be the finite discrete set of nonempty injective words over $\mathcal P_j$. The parent-and-reduce maps form an inverse system. Define the associated ordered novelty space

Indexing. The convention of §6.2 applies: $r(\mathbf D,\mathbf E) = \min\{j : D_j \neq E_j\}$ is the first disagreement index, so agreement through level $J$ is $r \ge J+1$ and $d_{\mathrm{nov}}(\mathbf D,\mathbf E) \le 2^{-(J+1)}$.

Indeed, $\prod_jX_j$ is a countable product of finite discrete spaces, and the compatibility equations define a closed subset. This is the classical inverse-limit/Stone-space pattern for finite discrete approximants.5859 The ultrametric says that two profiles are close precisely when their ordered discoveries agree through many levels of spatial resolution. The accompanying zero-dimensional compact-space perspective belongs to the classical dimension theory of Hurewicz and Wallman.60

The map

is generally not continuous in the prefix topology. For $K=\{0,1\}$, let $A_T=0^T1^\omega$ and $B=010^\omega$. Once the partitions separate $0$ and $1$, every pair $(A_T,B)$ has the same ordered profile $(0,1)$, but $A_T\to0^\omega$, whose profile is only $(0)$. Thus the kernel of $C^P_{\mathrm{nov}}$ is not closed in the original prefix topology.

The topology has now changed its question. In prefix topology, two traversals are close when their first many moments agree. In discovery topology, they are close when their complete first-discovery orders agree through many resolutions. A late discovery can be invisible to every short prefix yet decisive to a coarse discovery word; conversely, enormous dwell-time differences disappear if they do not change which cells are first encountered.

This completion also has an explicit hyperbolic realization. Form a rooted tree $T_R$ whose level-$j$ vertices are the compatible blocks $(D_0,\ldots,D_j)$ occurring as initial blocks of profiles in $R$, and join each block to its one-coordinate extension. When the realizable blocks span the full compatibility tree — under finite-profile extension, §10.6 — we write $T_{\mathrm{nov}}(\mathcal P)$ for $T_R$. Every level is finite, so $T_R$ is locally finite; with unit edges it is proper, geodesic, and $0$-hyperbolic. A ray is exactly a coherent profile whose every finite block is realizable, which is exactly the product-topology closure of $R$. Therefore

10.5 The Reciprocal-Sequence Computation

Take $K=[0,1]$, dyadic partitions of width $2^{-j}$, and $L=(1,1/2,1/3,\ldots)$. Put $N=2^j$ and $M_j:=|D_j(L)|$. Apart from an endpoint convention affecting at most a constant number of cells, $M_j$ is the number of distinct integer values of $\lfloor N/n\rfloor$.

The exponent $1/2$ is real, but it must be named correctly. The scalar $M_j$ forgets discovery order and records the eventual dyadic covering complexity of the visited set $\{1/n:n\ge1\}$; it is a box-counting exponent, not a temporal discovery-rate exponent.29 It differs from that set's Hausdorff dimension $0$ and from the ambient dimension $1$, while the full words $D_j(L)$ still distinguish different enumerations of the same set.

To measure pace without choosing a partition tower, retain time as a second variable. For example, let

The function $N\mapsto M_L(\varepsilon,N)$ measures how quickly a traversal realizes spatial complexity, and the inverse discovery time

records how long the $m$th scale-$\varepsilon$ discovery takes. Thus the architecture separates three kinds of information cleanly:

10.6 Density of Realizable Profiles

Let $\mathcal T\subseteq K^\omega$ be an admissible traversal class and write

For $\mathbf D=(D_0,D_1,\ldots)\in N_P(K)$, the level-$J$ cylinder is

Compatibility makes this equivalent to fixing the single coordinate $E_J=D_J$. These cylinders are the clopen balls of the novelty ultrametric.

For the first claim, given $w=(P_1,\ldots,P_m)\in X_J$, choose $x_i\in P_i$ and use

For non-halting traversals, alternate forever between two already discovered points; the one-cell word $(P)$ shows why every cell must contain two points. For injective traversals, continue through unused points of an already discovered cell; again the one-cell word shows why infinitude of every cell is necessary.

The quotient $\mathcal T/{\sim_0}$ embeds isometrically as $C^P_{\mathrm{nov}}(\mathcal T)$. Its metric completion is therefore the closure of that image inside the compact, hence complete, ultrametric space $N_P(K)$. Under finite-profile extension, every finite compatible block occurs, so the tree spanned by realizable blocks is the full compatibility tree $T_{\mathrm{nov}}(\mathcal P)$.

10.7 Memory as a Sufficient Statistic for Novelty

Novelty is not a property of a candidate state alone. It is a comparison between that candidate and information retained from the past. This observation turns memory from a cognitive metaphor into a precise factorization problem.

Let

be the space of finite histories, including the empty history. For a traversal $L$, write $H_n=(x_0,\ldots,x_n)$. A novelty predicate is a map

where $\mathsf{Nov}(x\mid H)=1$ means that $x$ is judged novel relative to $H$. Exact first-occurrence novelty and metric novelty at scale $\varepsilon>0$ are respectively

where $V(H)$ is the set of states appearing in $H$, with $V(\varnothing)=\varnothing$. We use the convention $d(x,\varnothing)=+\infty$ when metric novelty is evaluated before any state has been observed. Therefore

The terminology is a factorization-style analogue of a sufficient statistic: once $M(H)$ is known, the original history can be discarded for this particular decision problem. It is not the full Fisher–Neyman probabilistic notion unless a statistical model and parameter are also supplied.32

The canonical one-step memory is obtained by quotienting histories according to the decisions they induce. Define

and let

The quotient map $H\mapsto[H]$ is sufficient because

is well-defined.

Exact and metric memory

For exact novelty,

Equality of visited sets clearly gives the same decisions. Conversely, if the sets differ, any state in their symmetric difference is novel relative to exactly one history. The minimal exact memory is therefore

Dwell times and visitation order are unnecessary for the binary question “has this exact state appeared before?”

At a fixed metric scale, let

Then

so two histories induce the same decisions exactly when their closed $\varepsilon$-neighborhoods agree. Hence

is a canonical model of the one-step memory quotient at scale $\varepsilon$. The family $\{M_\varepsilon\}_{\varepsilon>0}$ compares what must be retained at different resolutions. For an arbitrary metric space these memories do not automatically form a refinement inverse system; the nested partition tower of §10.4 supplies the explicit bonding maps needed for that stronger structure.

From one decision to reusable online memory

To model every future decision under repeated updates, define the novelty-output word along a finite continuation $w=(y_1,\ldots,y_m)$ by

Now set

This is the novelty analogue of the future-indistinguishability quotient underlying minimal deterministic state representations. The Myhill–Nerode tradition supplies the corresponding finite-automaton perspective on indistinguishable continuations.4344 Exact and fixed-scale metric novelty already satisfy the update requirement:

Equipped with its transition and output maps, the quotient is a deterministic output machine

Memory is the state, the next observation is the input, novelty is the output, and the transition carries the compressed state forward. This is categorically stronger than a retrospective summary that answers only the present query.

Connection to recurrent neural memory

Modern linear-attention architectures provide a concrete engineering instance of the same compression problem, without constituting a realization theorem for the quotient above. A conventional transformer KV cache retains separate key–value representations of earlier tokens, so its stored context grows with sequence length. Recurrent linear attention instead compresses the processed context into a fixed-dimensional state matrix $S_t$, reads from that state, and updates it online:

DeltaNet-style updates first query the association currently stored at the incoming key and then write a correction. In one common orientation,

The residual term changes an association in light of what the memory already predicts instead of blindly adding another outer product. Gated DeltaNet adds broad learned decay, while Kimi Delta Attention uses finer-grained gating to control the finite recurrent state. The linked architectural essay traces this progression from growing KV history through linear attention, delta updates, gated forgetting, and Kimi-style attention; the Kimi Linear paper describes KDA as an extension of Gated DeltaNet designed to use limited finite-state RNN memory more effectively.33 The primary technical lineage includes linear attention as recurrent computation,45 fast-weight programming and delta updates,46 Gated DeltaNet,47 and selective state spaces in Mamba.48

The parallel is structural, not identificatory. A learned matrix state is finite-dimensional, approximate, task-dependent, and optimized from data. Nothing in the architecture proves that $S_t$ equals the canonical class $[H_t]_{\mathrm{fut}}$, that it is minimal, or that the model’s output task is a novelty predicate. Rather, recurrent attention confronts the engineering version of the theorem’s question:

In this sense, modern recurrent attention attempts to learn a useful finite online memory, while Theorem 10.12 characterizes exact task-sufficient online memory abstractly once the required output behavior has been specified. Forgetting then becomes controlled loss of distinctions: targeted replacement, decay, or gating is useful only insofar as the discarded differences are irrelevant to later outputs.

Connection to ordered discovery

At partition level $j$, let $P_j(x)$ denote the cell containing $x$ and let $S_j(H)$ be the set of cells already visited. The next-step decision

needs only the unordered set $S_j(H)$. The full discovery word $D_j(H)$ retains strictly more: it remembers the order in which those cells first appeared. That extra information is unnecessary for the binary next-step novelty judgment but essential to the ordered-novelty invariant studied in this section.

The two constructions therefore form a deliberate pair:

Symmetry and forgetting

Let a group $G$ act on $K$ and hence componentwise on histories. If

then $H\sim_{\mathrm{mem}}H'$ implies $gH\sim_{\mathrm{mem}}gH'$, and the action descends equivariantly to the memory quotient:

For metric novelty, every isometry has this property because $d(gx,gV)=d(x,V)$. In this precise sense, novelty is a symmetry-invariant comparison against memory.

10.8 Algorithmic Realization: Fenwick-Indexed Novelty

The quotient theorem identifies which distinctions are semantically necessary. It does not prescribe a data structure. Once a particular sufficient statistic and a family of online queries have been chosen, a separate algorithmic question begins: how should that statistic be represented and updated efficiently? This is the same broad resource question studied by bounded-memory streaming algorithms, where exactness, approximation, and state size must be traded explicitly.6263

Finite ordered alphabets

Let

with its given order. For exact first-occurrence novelty, Theorem 10.12 reduces the minimal online state to the visited set. Since every subset of $[m]$ is realizable by a finite history,

Represent $V(H_t)$ by its occupancy vector

Then

A bitset or direct Boolean array answers this membership query in $O(1)$ time. If membership is the only required operation, a Fenwick tree is unnecessary. It becomes useful when the order on $K$ gives meaning to cumulative questions.

Define the discovered prefix mass

A Fenwick tree—also called a binary indexed tree—stores an implicit decomposition of this cumulative-frequency array and supports point updates and prefix-sum queries in logarithmic time.34 On the first occurrence of $x$, perform

Before that update, the spatial rank of the new state among discovered states is

This is a rank in the underlying state order, not the time at which $x$ was discovered.

The Fenwick representation is lossless for the occupancy vector because

It is not a smaller semantic quotient than $V(H_t)$; it is a dynamic index over one representative of that quotient. Its advantage is the additional family of ordered cumulative queries.

Multiresolution occupancy

Now suppose the finite partition at resolution $j$ is ordered:

Define

If $r_j(x)$ is the index of the cell containing $x$, then the resolution-$j$ novelty output is

A candidate can therefore be old at a coarse scale and novel at a finer one. For a computational cutoff $J$, one observation produces the multiresolution signature

The infinite signature is a mathematical idealization; an implementation must truncate, evaluate levels lazily, or exploit the shared path of $x$ through the refinement tree.

Let $\mathrm{BIT}_j$ be the Fenwick index over $b_j$. Its prefix query

measures cumulative discovered mass in the chosen spatial order. The time-dependent number of cells discovered at scale $\varepsilon_j$ is

This is the temporal statistic deliberately absent from the eventual quantity $|D_j(L)|$. The curve

records how novelty saturates over time at resolution $j$. Its current total can be maintained by an $O(1)$ scalar counter; Fenwick is needed only when prefix mass, spatial rank, or selection queries are also required.

Occupancy is not discovery order

Fenwick does not by itself retain the ordered novelty word $D_j$. The following four objects answer different questions:

A complete online implementation at active level $j$ may therefore maintain

The bitset answers cell membership, the Fenwick index answers spatial cumulative queries, and the append-only log records $D_j$. Whenever a cell is first encountered, all three states update: mark the bit, add one to the index, and append the cell identifier to the log.

Across levels $0,\ldots,J$, a direct implementation uses

storage and at most

time per observation for membership/index work, before accounting for the cost of locating each cell. Sparse Fenwick maps, compressed bitsets, or the shared refinement tree can be preferable when most cells remain undiscovered.

Where the bridge stops

Fenwick requires an ordered, indexable domain and an additive prefix statistic. It is not canonical for a general metric space, graph, or arbitrary future-equivalence quotient. Other regimes naturally lead to other representations: a bitset for pure finite membership, a segment tree for richer interval aggregates, a trie for symbolic prefixes, a spatial or cover tree for metric queries, the refinement tree for nested partitions, or an approximate sketch when exact memory is too large.

The hierarchy is therefore

The quotient determines what information must survive. Geometry determines which multiscale distinctions are meaningful. The data structure determines how selected statistics are maintained without replaying history.

10.9 Streaming Systems Interpretation: Apache Flink

The quotient theorem has a direct systems interpretation in unbounded stream processing. A runtime such as Apache Flink does not recompute an operator from the complete event history after every arrival. It exposes updateable operator and keyed state—including primitives such as ValueState and MapState—so that each event can read, update, and checkpoint a state associated with the current stream key.35 The streaming literature supplies the broader one-pass state-complexity context.6263 For a key $k$, the computational shape is

This resembles online novelty memory, but the roles must remain distinct. Flink provides a fault-tolerant state substrate; it does not derive the minimal quotient. For exact novelty, a keyed map or set can represent the visited states $V(H_t^{(k)})$. Inside that state, a bitset, Fenwick tree, trie, refinement tree, or sketch may support the chosen observables. The architecture is therefore

The quotient supplies a correctness criterion. If two histories produce the same stored state, they must be future-equivalent for the intended output process:

A representation may distinguish more histories than necessary and remain correct, although it stores redundant information. If it merges histories that are not future-equivalent, some continuation must eventually expose an error.

TTL is a semantic choice, not merely cleanup

Flink can assign time-to-live policies to keyed state, including per-entry expiration for collection state. With a never-return-expired visibility policy, an expired value behaves as absent even before physical cleanup finishes. Mathematically, expiration applies an additional identification to histories. It is safe only when that identification preserves the future behavior required by the application:

For global exact novelty, a finite TTL is generally unsafe: after an old state expires, its next occurrence may be reported as novel even though it appeared earlier. TTL can be correct for a bounded-retention or window-relative predicate. A privacy policy may also require expiration, but then it intentionally overrides global novelty semantics rather than preserving them.

Windows change the predicate

A sliding, tumbling, or session window does not merely optimize global novelty. It defines a different question, such as window-relative novelty

The minimal sufficient state must therefore support both insertion and expiration. A plain visited set may no longer suffice: if the same state occurs twice inside a sliding window, expiring the older occurrence must not erase the newer one. Counts, last-seen times, queues, or timestamped maps may be required depending on event-time ordering and the window definition.

10.10 Boundary-Object Machines

This is a candidate boundary semantics, not a new theorem about computation. Domain theory records nontermination as $\bot$ in the denotational tradition of Scott–Strachey and its modern domain-theoretic treatments;3738 the present construction retains more structure whenever two runs have different discovery profiles. But distinct runs may still collapse to the same point because dwell time and repeated visits are deliberately forgotten. Any claim that the resulting object is computable from a temporal stream belongs to represented-space computable analysis, where the representation matters and computability entails continuity under the represented topology.39 The source note for this extension is boundary-machine.md.

When configurations are encoded over a finite alphabet, the run $\gamma(M,x)$ is also an infinite word. That places the construction beside $\omega$-automata and the theory of infinite-word languages, although §10.10 introduces no acceptance condition and claims no automata-theoretic equivalence.404142

11.RBDS Companion Lecture

The companion lecture Why AI Hallucinates (It’s the Same Math as Gravity) develops the broader framework of Resource-Bounded Distinction Systems (RBDS): order requires distinguishable states, observations consume resources, and a finite observer acts on an operational quotient rather than on objective state space directly.30 The present paper supplies a canonical traversal geometry for that thesis. This section states the connection precisely and separates exact algebra from cross-domain interpretation.

11.1 Budgeted Operational Quotients

Let $\mathcal O$ be a family of observations on an objective state space $X$ (not the occurrence space of §5; the two objects never co-occur), and let $c(o)$ be the cost of observation $o$. At budget $B$, the admissible observations and the resulting observation signature are

Define budgeted indistinguishability by equality of every accessible observation:

and define the operational state space

This is the RBDS form of the partition tower in Section 10.4. Each partition $\mathcal P_j$ represents the equivalence classes visible at a resolution budget $B_j$; the word $D_j(L)$ records the order in which those operational classes are first encountered. The compatibility map $q_{j+1,j}$ is therefore not merely a spatial convenience—it is the quotient map from a more discriminating observer to a less discriminating one.

Budget monotonicity is automatic once the admissible observation families are nested. A larger budget never removes a test that was previously affordable, so it can split an old equivalence class but cannot merge two classes that an earlier observer already distinguished. The arrow therefore points from fine quotients to coarse quotients:

When the budgets increase and the partitions separate points, the ordered novelty completion

is a coherent infinite-budget limit of finite operational descriptions. Exact identity appears only in the limit; every finite level retains resource-bounded equivalence classes.

11.2 The Conceptual Bridge

This gives a precise non-identifiability criterion for model claims. Given accessible evidence $E_{N,B}$—a finite prefix together with budget-$B$ observations—let $\mathcal C(E_{N,B})$ be the set of compatible complete traversals. A property $h:K^\omega\to H$ is identifiable from that evidence exactly when

If $h$ varies across $\mathcal C(E_{N,B})$, then choosing one value of $h$ requires information not contained in the accessible evidence. In this formal sense, hallucination is not merely “being wrong”; it is presenting a non-identifiable representative as though the operational quotient had selected it uniquely. This finite-evidence limitation sits next to the broader algorithmic-information treatment of description, inference, and asymptotic uncertainty.64

11.3 The Attention-Drift Identity

The lecture's physics–transformer bridge begins with an exact transport identity. Let $A\in\mathbb R^{n\times n}$ be row-stochastic, as a softmax attention matrix is, let $V\in\mathbb R^{n\times d}$ contain value vectors, and let

Define column imbalance relative to unit receipt by

Row normalization conserves the amount allocated by each source, but it does not force equal total receipt by each destination. The vector $\Delta$ measures that column imbalance, and $\Delta^\mathsf T V$ is the resulting aggregate displacement. The same algebra can describe directed redistribution in attention, active-matter interactions, or other asymmetric transport systems.

In sentence form: every token distributes one unit of attention, but some tokens can receive more total attention than others. Multiplying that receipt imbalance by the value vectors tells us the net shift in the aggregate representation. The formula is short because the accounting identity is simple; the difficult scientific question is what that shift means in a particular model.

11.4 Causality as Adaptive Refinement

The construction in Section 10 uses a fixed partition tower. RBDS suggests a stronger, observer-inclusive model in which current discoveries determine which distinctions become affordable or relevant next:

The causal architecture is then a loop,

The fixed inverse limit developed earlier is the nonadaptive baseline. An adaptive RBDS would replace predetermined bonding maps by history-dependent maps and ask when the resulting path-dependent completion remains compact, metrizable, and realizable by a proper hyperbolic boundary. That is the natural next extension connecting ordered novelty to learning systems whose observations change their future state space.