WorksAgricultural and Biological Sciences
Rule-table structure of life-like cellular automata: activity, asymmetry, sensitivity, and the propagation storage factorisation
Abstract
What structure on the rule space of a cellular-automaton family exposes which rules can support which behaviours? For the 262,144 life-like rules – the two-dimensional outer-totalistic family containing Conway's Game of Life – the standard coordinate, Langton's activity parameter, is long known to be too coarse. I organise the family by three symmetry-derived rule-table coordinates: an unweighted activity, a birth-survival asymmetry (odd under the swap of birth and survival), and the Vichniac-Binder Boolean-derivative sensitivity in closed form, whose centre-bit component vanishes exactly on the centre-blind (B=S) submanifold. The cardinality coordinates lambda, delta are necessary but never sufficient to certify behaviour; the centre-bit sensitivity labels the B=S submanifold on which the paper's one exact exclusion holds. Combined with an exhaustive enumeration of every rule's compact gliders, still lifes and oscillators (three-by-three seed, period at most sixteen), they expose a hierarchical factorisation: the birth condition is a narrow gate deciding whether anything can propagate, the survival condition tunes storage, and every compact-glider-supporting rule is birth-light and survival-rich. One exclusion is exact: no B=S rule supports a compact glider at that footprint, proven by exhaustive enumeration and closing only for compact-footprint patterns – 36 B=S rules ship at four-by-four seed and 40 at five-by-five. Two dynamical results frame the whole: glider-capable rules that random sampling never expresses show that capability is certifiable from the table while its expression is only estimable from dynamics; and every uncatalogued rule producing a richly Life-like ecology turns out to be a Conway or HighLife variant on the neighbour counts its trajectories actually visit.
Introduction
A cellular automaton (CA) updates a lattice of cells in parallel by a rule that depends only on each cell and a small neighbourhood, yet even tiny rule families produce propagating structures, long transients and universal computation [21, 22, 4]. This paper concerns a basic organisational question for one such family: given the 262,144 two-dimensional outer-totalistic “life-like” rules, what coordinates on the rule table expose the structure of behavioural capability—which regions support propagation, which support storage, and which exclude a behaviour outright? A life-like rule is fixed by two subsets : a dead cell is born if its live-neighbour count lies in , and a live cell survives if its count lies in . The family contains Conway’s Game of Life and its named variants—HighLife [3], Day & Night, Seeds, 2×2, Replicator, Maze [10, 1, 2] (Eppstein 2010).
The standard coordinate is Langton’s , the fraction of the rule table that yields a live successor [13]. Sweeping moves through frozen, periodic, complex and chaotic regimes, and the intermediate “edge of chaos” was conjectured to host maximal computational capability [16, 13]. But is a single number summarising an 18-bit table: two rules that differ in which neighbour counts fire, not just how many, collapse onto the same point, and evolved-CA studies found no sharp critical [15]. Prior work has organised the life-like family along other single axes—Wuensche’s -parameter sensitivity ordering Wuensche [25], P.-M. Binder’s plane Binder [6], and more recent classifications of the family’s rule space by dynamical-descriptor statistics Pena & Sayama [17]—and evolutionary searches have located individual complex rules directly Sapin et al. [18]. The present paper is complementary: rather than a single ordering axis or an evolutionary search, it uses a small basis of symmetry-organised coordinates paired with an exhaustive combinatorial enumeration of what each rule can host at a fixed compact footprint.
Necessary, not sufficient. The collapse of rules that differ in which neighbour counts fire onto the same coordinate value is not a defect to be engineered away but a structural fact worth stating plainly, because it sets the ceiling on what any cardinality-based coordinate—one built only from how many elements are in each set—can do. Both and the asymmetry I introduce below are functions of the cardinalities alone; there are only such pairs, so the entire plane has at most 100 occupied points, each shared by thousands of behaviourally distinct rules. A coordinate this coarse can never certify that a rule does something—any value it takes is shared by a crowd. What it can do is the opposite: place a rule in a region, and—in the limiting case—exclude a behaviour entirely. My results are organised around this asymmetry. The asymmetry coordinate locates where capability concentrates (a necessary condition); the centre-bit sensitivity supplies the one exact exclusion, ruling out compact gliders on the centre-blind submanifold outright.
The coordinates. I add two complementary rule-table coordinates. The first is a birth–survival asymmetry: Langton’s counts how much of the table fires, and splitting that count into its birth and survival halves and taking their difference measures whether a rule births more than it preserves or the reverse. Built directly from Langton’s parameter this asymmetry is , and it inherits Langton’s multiplicity weighting—each neighbour count enters weighted by how many neighbourhoods realise it. I also use unweighted siblings and that count the birth and survival conditions themselves, without multiplicity. The structural results below are read on this unweighted pair, with the Langton-consistent kept as reference; the specific empirical case for over is developed alongside those results (§2.2). Either way the asymmetry is defined only for outer-totalistic families whose rule table separates by cell state (dead-cell births vs. live-cell survivals ); it is a coordinate for this family, not a re-parameterisation of that generalises to CA whose transitions do not factor this way. The second coordinate is the average Boolean-derivative sensitivity —the probability that flipping one input bit of the neighbourhood flips the output—introduced for CA by Vichniac Vichniac [20] and used as an a-priori ordering parameter by Binder Binder [6]. For life-like rules admits a closed form whose centre-bit part vanishes exactly on the diagonal . The two new coordinates are independent, not interchangeable: under the birth/survival swap , is odd while is even, so cannot tell a rule from its mirror and can. With and the bit-flip complement these three coordinates populate three of the four parity sectors of ; the fourth is not used analytically here (§ 2.4).
Findings. Pairing this coordinate system with an exhaustive combinatorial enumeration of every rule’s compact-spaceship, still-life and oscillator support, I establish three results. (i) A hierarchical factorisation of capacity, read off the birth–survival asymmetry : the birth set is a narrow gate that decides whether anything can propagate, while the survival set tunes how much structure is stored within that gate. The gate structure is robust across seed footprints (the propagation-enabling birth sets are a subset of the ones, both dominated by small birth sets); at the compact footprint this shows up as a clean fingerprint—the rules supporting compact gliders are systematically birth-light and survival-rich ()—which is the paper’s central structural finding, and which I show is specific to the compact footprint and dilutes as the window widens. (ii) Three coordinate projections of the rule table—the Binder plane and its and companions—that organise the family non-trivially, in which the 512 centre-blind rules occupy a sharply identifiable locus. (iii) A total exclusion of rules from compact (, period ) gliders—certified by the exhaustive enumeration (§4.1). Two empirical regularities of the enumeration organise the diagonal into a small candidate set that shipping rules must inhabit: every rule satisfies and , so shipping candidates collapse to with —32 rules. This is an organising observation, not itself a proof reduction, because is read from the same exhaustive enumeration as the exclusion itself; a proof of the exclusion from a mechanism independent of the enumeration remains open. The oscillator enumeration localises the exclusion to translation rather than in-place recurrence, and it softens into a widening handicap rather than vanishing at larger footprints. Two dynamical results frame these. First, a population of rules that support a glider combinatorially yet rarely or never express one under random sampling shows that capability is a rule-table property, distinct from what random dynamics reveal (auxiliary dynamical observables are correspondingly reframed as regime descriptors, not capability detectors). Second, and conversely, a table-blind screen of the uncatalogued build-and-hold rules for Game-of-Life–like dynamics returns a richly Life-like ecology only for Conway and HighLife variants—rules whose tables differ from an anchor’s solely on neighbour counts their trajectories never visit—so that table difference need not imply behavioural difference. The exact bit-flip duality (§ 2.4) is used only to gauge-fix that auxiliary dynamical sweep (§ 2.6).
Methods
Notation and computed quantities
Table 1 collects every rule-table and dynamical quantity used in the paper. The rule-table coordinates are defined in §2.2–§ 2.4, the combinatorial indices in § 2.5, and the single dynamical sweep in § 2.6. All rule-table quantities are functions of alone and are independent of any simulation protocol.
| Symbol | Definition | Measures |
| Langton’s parameter: multiplicity-weighted fraction of the full table that fires | ||
| Langton-consistent birth–survival asymmetry | ||
| unweighted (condition-counting) activity; working coordinate | ||
| unweighted birth–survival imbalance; the paper’s factorisation axis | ||
| avg. over inputs/bit-positions of | sensitivity: how often flipping one input cell flips the rule’s output | |
| how much a cell’s own state matters to its fate; zero exactly when | ||
| # distinct -orbit spaceships, | how many compact gliders the rule can support at all | |
| # distinct -orbit still lifes | how many small static patterns the rule can hold fixed | |
| # distinct -orbit oscillators, | how many small patterns the rule can hold as in-place blinkers | |
| birth sets with at least one companion giving | which birth rules “open the gate” for propagation / oscillation / storage | |
| peak off-axis residual spatiotemporal autocorrelation | whether random dynamics actually show a coherent moving structure |
Table 1. Notation and computed quantities. The first six rows are rule-table coordinates (functions of alone); , , count the small moving / static / blinking patterns a rule supports; is the paper’s single dynamical observable. is the set of counts where birth and survival disagree; and are the swap and complement symmetries of §2.4; is the eight-fold symmetry group of the square.
Rule-space coordinates:
Each rule splits its look-up table into a birth component on dead cells and a survival component on live cells. For a set , denotes its cardinality—the number of elements it contains—so is the number of neighbour counts that trigger birth and the number that trigger survival, each an integer in ; for Conway and , so and . Langton’s parameter is the fraction of the full -entry transition table that fires; on an outer-totalistic rule this weights each neighbour count by its multiplicity , and splits by cell state into birth and survival halves,
with antisymmetric companion
This is the birth–survival asymmetry built directly from Langton’s .
Alongside these I use unweighted siblings that count firing conditions rather than multiplicity-weighted table entries—each of the nine neighbour counts contributing equally:
with antisymmetric companion
The two coordinates share their form and differ only in the summand: the weighted counts each firing count with the multiplicity of neighbourhoods that realise it—the fraction of the full 512-entry table that fires—while gives every count weight one, the fraction of the 18 reduced outer-totalistic conditions that fire. The weighted and unweighted forms agree only at the extremes and differ in between—for Conway, against , because the multiplicity weights concentrate on the middle counts that a bare cardinality treats equally. The choice is not cosmetic: and order rules differently from and , and the unweighted pair discriminates compact-glider capability better (the birth-light propagation family disagrees at low counts, which the binomial weighting underemphasises; separates from at rank-biserial against 's ; the same ordering shows up under AUC (0.71 vs. 0.63) and holds direction at the footprint too, though weakly (0.56 vs. 0.54), consistent with the compact-fingerprint story fading with footprint (§ 3.2).) No monotone rescaling relates the two, since the difference is a reordering rather than a change of scale. I therefore read the structural results on the unweighted , and retain the weighted as the Langton-consistent reference to which the family connects. The population remains a single connected region in the weighted plane as well (10-vertex convex hull, enclosing of the bulk against under the unweighted coordinates); the weighted convex hull is thus tighter as a bounding envelope for . On each 1D axis the weighted coordinate takes 479 distinct values across the 262,144 rules against 19 for the unweighted (weighted counts every subset sum of ; unweighted counts only), so a scatter view of the weighted plane spreads the population across many more discrete coordinate positions and looks more diffuse pointwise, even while its convex hull is tighter overall.
versus . These loci are easily conflated but combinatorially distinct, and the distinction is exactly the necessary/sufficient one. is the equal-cardinality submanifold , containing rules; the strictly diagonal (the centre-blind rules) is far smaller, only 512. This set is not new: it is the Life-like family’s centre-blind subclass—rules whose update depends only on the neighbour count, not on the centre cell’s own state. It has sometimes been called “totalistic” in the Life literature, but this differs from Wolfram’s classical totalistic convention (which sums over all 9 cells including the centre); I use “centre-blind” throughout to avoid the ambiguity. What is new is the characterisation of this subclass as the exact zero set of the centre-cell sensitivity below, and the compact-glider exclusion I prove on it. Equivalently, writing the unweighted set distance (the number of counts in exactly one of ), is while is only , so . The coordinate that isolates exactly is the centre-bit sensitivity introduced next, of which is the unweighted shadow: as follows from the closed form in §2.3 (point 1), is precisely the locus; §3.1 confirms this empirically over the full family.
Sensitivity coordinate (Boolean derivative)
The coordinates summarise activity but say nothing about how strongly a rule responds to single-bit perturbations of its input. The standard summary is the average Boolean derivative [20, 6]: for a rule with input bits and binary output ,
where runs over all binary configurations of the -bit neighbourhood (for life-like rules : the centre cell and its eight Moore neighbours), is the string with a single 1 in position , so is with exactly bit flipped ( is bitwise exclusive-or), and the brackets are the Iverson convention—1 when the enclosed condition holds and 0 otherwise. The prefactor averages the resulting flip-count over all (configuration, bit-position) pairs, so is the probability that a single-bit flip flips the output; it is the nearest sensitivity-style ordering to Wuensche’s parameter [25], which similarly captures how much the update depends on individual input bits but is defined on the full lookup table rather than as the closed-form Boolean derivative for the outer-totalistic reduction I use below. For an outer-totalistic life-like rule on the Moore neighbourhood the inputs split as one centre bit and 8 neighbour bits, and the table depends on inputs only through with the centre value and the neighbour count. With , the centre-bit contribution has the closed form
Only the centre-bit part is used analytically below; the neighbour-bit contribution has the analogous closed form
where picks out the counts at which set-membership switches between and (the “edges” of as a subset of ). Then , a weight-/weight- average. Giving the two sides of the disagreement opposite signs turns into a -odd quantity, ; the shared counts cancel, so this equals the weighted asymmetry of §2.2 identically. Signing a sensitivity thus reproduces the Langton-consistent , not the unweighted —a magnitude cannot be made odd without becoming a cardinality difference (§ 2.4). The notation needs a word. is the symmetric difference of the two sets: the neighbour counts at which birth and survival disagree—a count belongs to when it triggers exactly one of the two, birth or survival, but not both and not neither. For Conway (, ) the two sets share the count 3 and differ only at 2, so : a cell’s own state matters only when it has exactly two live neighbours. Two consequences follow.
is the submanifold. By the formula iff , i.e. . So is exactly the binomially-weighted set distance from centre-blind, and the 512 centre-blind rules are characterised a priori: they cannot respond differently to the centre cell’s own state at any count.
is independent of : one measures the fraction of table entries that fire, the other the fraction of input-bit flips that change the output. Together they form a Binder plane (§3.1).
A symmetry-organised coordinate system
The rule table carries two natural involutions,
where and the bar is the complement in . swaps birth and survival; is the bit-flip duality, in the sense that if then for every configuration. Both square to the identity, and the rule space falls into the four parity sectors of . Of these four sectors, three are populated by coordinates the paper uses: occupies , occupies , and (with ) occupies . The remaining sector, , is not empty in principle—a natural candidate is an edge-contiguity coordinate that compares how bunched the two thresholds are as bands of neighbour counts—but the one such coordinate I constructed and tested proved subordinate to : conditional on it retained only a small fraction of ’s effect on , so I do not carry it through the analysis (§ 4.4).
Parity table. I verified the action of and on every coordinate by direct evaluation on all 262,144 rules (max residue ): is -even, -odd; ); is -odd, -even; invariant under ); are -even, -even). The key observation is that and are both -even, so neither can distinguish a rule from its mirror; is the unique -odd coordinate among the activity/sensitivity axes, and is therefore the natural birth-versus-survival scalar. Since acts as , self-dual rules need not have —consistent with Day & Night () being self-dual. Pearson correlations across the full rule space are exactly zero between every pair of , as required by the orthogonality of parity sectors under the uniform measure.
The table is forced, not chosen. That the sensitivity coordinates all occupy one sector is not an artefact of which observables I happened to compute. A Boolean derivative asks only whether a single-bit flip changes the output; it is a magnitude, carrying no sign and no direction, whereas an odd coordinate must carry one. Concretely, if then for every configuration, so flipping input bit changes ’s output at exactly when it changes ’s at ; since preserves the uniform measure, . Likewise leaves fixed and symmetrises the neighbour term. Every sensitivity of this kind is therefore -even, -even), and the remaining three sectors can be populated only by counting functionals of the rule table; the two families are not interchangeable options for a cell. A second constraint fixes which coordinate can carry the exclusion. Any -odd obeys , so it vanishes on the fixed-point set of —but also wherever its two sides merely balance, a strictly larger set ( on 48,620 rules). Only a -even functional that is non-separable, a function of jointly rather than of and apart, can have as its exact zero set. is such a functional, which is why the exclusion of §3.2 is stated with it and could not have been stated with ; its binomial weighting is inherited from the configuration measure and plays no part, since any strictly positive weighting of has the same zero set. The exclusion depends on the disagreement set being empty, not on how the disagreement is weighted.
Roles of the coordinates. Each answers a distinct question the others structurally cannot, and the necessary/sufficient distinction runs through all three. places a rule on the frozenchaotic spectrum but collapses rules of equal total firing fraction, so it is a necessary entry point, never sufficient. is the swap-odd factorisation axis: rules are birth-light and survival-rich (), the rule-table fingerprint of the -gate / -knob structure (§3.4); without the claim cannot even be stated. is the one coordinate whose zero set coincides exactly with the diagonal, so it labels the exclusion in a single coordinate value; this is a re-parameterisation of the exclusion, not a mechanism for it (§3.2, §4.1).
Combinatorial enumeration: , ,
The following are functions of the rule table alone and independent of any simulation protocol. For each rule I enumerate every binary pattern (reduced to its orbit representative) on a quiescent zero-padded background, evolve it to , and classify the orbit by its return behaviour: a nonzero-offset translated return is a spaceship ( counts distinct orbit-representative spaceships, with period and velocity set); a same-position return at period is an oscillator (); a fixed pattern () is a still life (). “3×3” refers to the seed footprint: the pattern may expand during the cycle and must contract back at ; therefore counts compact-seed spaceships of period , not spaceships whose bounding box stays inside throughout. The period cut-off is an empirical truncation but is close to saturating: the min-period distribution of the 424 rules peaks sharply at (312) with a rapidly decaying tail (:48, :2, :48, :8, :4, :1, :1, none at or ), so the cutoff censors negligibly. Directed spaceship searches over life-like rules exist [8]; my enumeration differs in being exhaustive over all -orbit seeds of a fixed footprint and all 262,144 rules simultaneously. asks “can the rule support a compact glider at all,” independent of whether a random trajectory ever exhibits one.
Protocol specifics. Each seed is placed centred on a padded grid of size (in cells, where is the seed footprint, cell/step is the maximum life-like propagation speed, and the extra padding guarantees no live cell reaches the boundary within steps). Evolution halts if the pattern extincts, if the bounding box exceeds (a growth escape criterion; the seed is then not classified), or at . Return classification checks whether the current live set is a rigid translate of the seed’s live set (with any integer offset ): I do not accept glide-reflected returns as spaceships at this footprint, which is a convention choice, not a mathematical necessity—glide-reflected returns at compact footprint are possible in principle for some rules but were not observed under this protocol. The pre-screen relies on a necessary condition: if a phase at is a translate of the seed, both configurations have the same live-cell count. I evaluate the count at every step and short-circuit the full shape-comparison only when the count matches the seed’s count; this is a check applied at the potential return step, not a mid-cycle rejection (mid-cycle populations can differ from the seed’s, and the pre-screen never prunes them). No false negatives: any true translated return is caught because the count equality is a necessary condition, and the shape comparison at return steps is exhaustive.
Dynamical probe and coherent-propagation index
To identify rules whose combinatorial propagation is or is not expressed by random dynamics I run a single fixed sweep. For each rule I draw 5 shared Bernoulli ICs at density on a torus, together with their bit-flipped partners at density 0.70 (the symmetric-IC ensemble described in the next paragraph), evolve 600 steps, and measure on the 200-frame tail a residual spatiotemporal autocorrelation : a coherent structure of period and velocity produces an off-axis peak at , validated by an injection test in which an isolated Conway glider peaks at . I use only to flag rules with no measurable coherent propagation; the remaining dynamical observables of the sweep are regime descriptors, reported as such.
Bit-flip duality and the symmetric-IC ensemble. The bit-flip duality (§2.4) is exact at the level of trajectories ( for ) but is broken by sparse Bernoulli ICs: a rule and its dual visit different attractors from the same IC. Pairing every sample with its bit-flipped twin at density restores the duality at the level of ensemble means; I verify a residual twin-gap at the seed-noise floor. The 512 -fixed points (e.g. Day & Night) are the self-dual rules. Beyond the sweep records per-frame compressibility, a basin-asymmetry index and the cell-flip rate; gauge-fixed by symmetric-IC averaging these co-locate -twins and separate canonical anchors into recognisable regimes. They are reported primarily as regime descriptors, deferred to the repository; capability detection through them is a partial secondary signal rather than the primary route.
Results
The rule-table phase diagrams (, ), (, ), (, )
The sensitivity coordinate places every rule on a second axis complementing (Figure 1). The cloud of all 262,144 rules forms a structured band, and the submanifold () traces a sharply identifiable locus on which all 512 centre-blind rules lie. Replicator sits at the top with —centre-insensitive but maximally neighbour-sensitive—showing is not synonymous with low sensitivity, only with centre-flip insensitivity. The 424 rules with cluster off that locus at moderate and moderate : full range [0.32, 0.94], interquartile range [0.44, 0.71], median 0.55. Conway () sits at the low- edge of the cluster and HighLife (0.46) inside it, while the storage-heavy Class IV anchor Day & Night lies at yet is itself (Figure 5)—a first illustration that inhabiting the region on (, ) is necessary for compact propagation but does not certify it (Wolfram 1984). (2×2, another Class IV anchor with , sits at , inside the IQR; both rules are complex through engineered rather than compact-glider dynamics, a point I return to in §4.2.) The middle panel shows the same rules in (, ), where the cluster sits at (the birth-light/survival-rich offset of §3.4) and the rules collapse to the line ; the right panel shows the (, ) projection, where the rules form a diamond near (0.4, −0.2). I do not claim the planes as a sharp classifier: the band is contained in, but does not exhaust, the moderate region—necessary, not sufficient.

Figure 1. Coordinate projections of all 262,144 life-like rules. Grey: subsample of the full rule space. Royal-blue X: the 512 centre-blind () rules (). Crimson: the 424 rules with . Anchors are outlined markers coloured by Wolfram class. (a) (, ): rules trace the locus. (b) (, ): rules collapse to while rules cluster at ; HighLife lies on by equal cardinality without being centre-blind (§2.2). (c) (, ): three nested convex-hull envelopes show the propagation-storage-oscillation factorisation of § 3.4 in coordinate space—crimson dashed for (, 6 vertices), teal dotted for (), amber dashed for (). Envelopes are geometric necessary regions but not coordinate-level exclusions: the crimson envelope alone contains rules ( of the bulk), and the amber envelope covers most of the attainable plane.

Figure 5. Joint propagation–storage plane in coordinates. Conway and HighLife sit in the upper-right corner (both gliders and rich storage). The column supports no compact glider and contains the Class IV anchors 2×2, Day & Night and Move—whose complexity rests on engineered, large-footprint propagation this enumeration does not capture (§ 4.2)—while Seeds and Replicator sit at the origin. The silent population from the main sweep is shown split into 33 robustly silent rules (crosses) and 5 rare-basin rules (open circles), per § 3.3; the two together total 38 points.
Combinatorial spaceship support and the exclusion
Of the 262,144 rules, holds for 424 (0.16%); these cluster in the (, ) diamond centred near (0.4, −0.2) (Figure 1c, innermost crimson envelope), with Conway at its low- edge. They support 876 distinct orbit-representative spaceships, of which 560 are diagonal and 316 orthogonal, with no 3×3 knight-ship (an empirical finding of the exhaustive enumeration, not a geometric impossibility: a knight-vector translation is compatible with a 3×3 seed and in principle, but no rule realises one). The period distribution peaks at . Direction is a clean per-rule property: of the 424, 300 support only diagonal ships and 124 only orthogonal, none both at the 3×3 footprint. Only Conway and HighLife appear among the canonical anchors, leaving many rules that support a compact glider yet remain uncatalogued in prior searches ([8, 11])—though, as §4.2 shows, those with genuinely Conway-like bulk dynamics turn out to be anchor variants rather than new rules.
The exclusion and its footprint dependence. The 512 rules with (equivalently ) contain zero rules with at any , against the bulk rate of 0.16%. This is the sharp form of “exclude, not certify”: the diagonal (equivalently the level set) is a coordinate-level label for the empirical exclusion. It is specific to the compact footprint. Enumerating larger seeds exhaustively over all rules, 36 of the 256 vacuum-stable ones support a 4×4 spaceship and 40 a 5×5 one (against 0 at 3×3), so the exclusion holds at the compact footprint alone, not at all scales. ( enumerates exactly that compact channel, not ordinary activity spread, which any non-frozen rule supports.) The recovery is bounded on the two footprints examined: the count sits near 15% of vacuum-stable rules at both 4 × 4 and 5 × 5 (a two-point observation, not extrapolable), and build-and-hold—a glider-together with static storage—remains rare, holding for 4 rules at 4 × 4 and 8 at 5 × 5 (counting still lifes; every glider-bearing rule oscillates, so oscillation is not an informative storage test here). Moreover the gap to Conway widens with footprint: the still-life count of the richest such rule trails Conway’s by 6.3× at 4×4 and 9.2× at 5×5. A rule forced onto the diagonal tunes propagation and storage with a single set and cannot exploit a larger window on both axes at once, as a rule with independent and can; the strict 3×3 exclusion therefore softens into a widening handicap rather than vanishing. The combinatorial recovery does not, however, bring behavioural complexity with it. Running the richest build-and-hold candidates—the rules that combine a larger-footprint glider with the most still lifes—from random initial conditions across a range of densities, and inspecting the trajectories directly, I find chaotic (Class III-like) bulk dynamics: the field boils to a disordered steady state without the enumerated glider visibly emerging or persistent objects surviving. Within the footprints examined (≤ 5×5), no centre-blind rule was found to express its enumerated glider from a random start: the diagonal gains combinatorial glider capability at larger scales without expressing it. A broader behavioural screen, applied without reference to , flags a small number of centre-blind rules as candidates for structured dynamics; whether these constitute a genuine small-object ecology or densely-packed frozen tessellations is not settled by the sampling used here, and I leave it open. Discussion of possible partial mechanisms and their limits is in §4.1. (The 4×4 enumeration is exhaustive over all 262,144 rules—8,206 support a spaceship, and the 424 compact-glider rules are contained in this set exactly; the 5×5 enumeration is exhaustive over but sampled elsewhere. Enumeration at 6×6 and larger footprints is beyond my computational means, so I cannot rule out that the recovery continues to widen or eventually saturates further out.)

Figure 2. Combinatorial characterisation of the 424 rules with . (a) Period distribution, mode ; the tail decays rapidly with a single rule at and a single rule at , none at or , so the enumeration’s cutoff censors negligibly. (b) Direction in the plane, points coloured by whether the rule supports only diagonal ships (blue circles) or only orthogonal ships (red squares); no rule at the compact footprint supports both. Orthogonal motion concentrates in high and large jointly—the joint (upper-right) corner is 86% orthogonal, either lever alone 21–34%, and the unfavoured (lower-left) corner 3%; diagonal is the default elsewhere (§4.1).
Glider-capable but dynamically silent rules
Cross-referencing against isolates a sharp population that makes the capability/expression gap concrete. I find 33 rules that combinatorially support a glider yet produce no coherent propagation from random initial conditions across a deliberately wide search—13 densities , 25 seeds each, runs of steps (325 trials per rule).
The asymmetry between the two claims is the point. That the glider exists is a construction, not a sample: enumerated from the table and verified by injecting it onto a quiescent grid, where it propagates steadily. That random dynamics do not produce it is, by contrast, only ever an estimate over sampled initial conditions; I state it as a measure, not an absolute—the glider’s basin is negligible under random sampling, undetected in those 325 trials. Capability is certifiable from the table; expression is only estimable from dynamics. (A further five rules silent at express intermittently at other densities; they are rare-basin, not silent, and are excluded from the 33.)
The obstruction is not a missing background. Every rule has a stable dead vacuum by construction—a glider is enumerated on a quiescent background, so a rule whose vacuum ignited could not register —and indeed all 33 have . The glider genuinely lives on empty space; the random-IC trajectory simply never assembles it there. Measuring the trajectories shows two opposite failures: 22 of the 33 freeze (the field settles to near-vacuum, observed tail densities –) and 11 flood (the field saturates, –), with none in the moderate band that expressing rules—Conway among them—occupy. On the silent population the split is read by a single survival bit: every flood rule has (saturated regions persist; 11/11), while the freeze rules have restrictive survival (mean across the freeze rules, two-thirds with loneliness death , ). This is descriptive of the 33, not a general law: on a uniform 3,000-rule sample from the full 262,144, randomising the density alone recovers the flood-versus-freeze (dense-versus-sparse settling) split in the large majority of pairwise comparisons—about 86%—while no simple birth/survival set relation predicts the split out of sample. The silent rules are those whose glider-bearing birth set pairs with a survival rule that drives the trajectory out of the viable band, too dense or too sparse for the glider to persist on clean background. Which regime a rule falls into is determined by the table but not reducible to a closed-form predictor of it: a mean-field density map in the style of Schulman and Seiden [19] reproduces the gross density ordering across rules (rank correlation 0.77 with measured , against 0.34 for the best single-bit feature) yet mis-places most silent rules at a moderate interior fixed point () the spatially-fragmenting dynamics never reach. The regime must be simulated—the dynamical counterpart of “certifiable versus estimable.”
A controlled witness fixes capability and isolates the survival rule. The family contains rules that share a common supported diagonal glider, Conway among them. Holding the glider fixed and varying only , some express it from random ICs and others do not: Conway’s does, while —the same birth set, the same supported glider—does not. Expression is thus not a function of capability alone. Propagation capacity is set by ; whether random dynamics express it depends on (§ 4.2).
Hierarchical factorisation of propagation and storage
Define the propagation-enabling birth sets . Only 26 of 512 birth sets (5.1%) qualify, and by the reduction of § 4.1 every one of them contains 3 (and none contains 2); outside propagation capacity is absent by construction. Within the enabled slice (13,312 rules) the Spearman correlation between still-life count and is 0.551; outside the slice it drops to 0.135 (Figure 4). The factorisation is hierarchical, not symmetric: gates whether any structure is possible, tunes storage richness once the gate is open. One caveat on what is being discovered here: that a still life’s persistence involves (empty neighbourhoods of the still life must not birth) is partly definitional — a enumeration on a quiescent background necessarily reads any would-be still life against the birth condition. What is not definitional, and is what the enumeration actually finds, is the strict nesting with the specific ratios 26/128/176 (out of 512) and the four-fold Spearman contrast 0.551 vs. 0.135; both quantify how much -tuning happens conditional on the -gate.

Figure 4. Hierarchical factorisation. (a) Birth-set support of each capability (a birth set qualifies iff at least one of the 512 possible survival companions yields the structure, i.e. ): with 176 (still lifes), 128 (oscillators) and 26 (spaceships) birth sets respectively, nested strictly within the 512 possible—propagation is the narrowest gate, oscillation wider, storage widest. (b) #SL against survival-set cardinality , split by (orange) vs. outside (grey): inside the propagation gate storage tunes with (Spearman ), outside it does not ().
In-place recurrence forms one storage axis. Enumerating in-place oscillators by the same protocol, 57,270 rules (21.85%) have ; all 424 rules are among them, 56,846 oscillate without any spaceship, and none support a spaceship without also oscillating. Oscillation is birth-gated through a strictly wider gate: contains 128 birth sets against 26 in , and storage (any still life) through a wider one still, 176 in ; the three form a strict three-level nesting (26 128 176 of 512; Figure 4a). The same nesting appears in coordinate space as three nested convex-hull envelopes on the plane (Figure 1c). The coordinate-space view understates the sharpness of the factorisation: the amber storage envelope covers most of the attainable plane because holds for of rules, so “inside the storage envelope” is a very loose necessary condition on . The tight version of the factorisation— for storage, for oscillation, for propagation—is a birth-set statement (Figure 4a), not a cardinality one; the cardinality coordinates locate but do not certify, and the visible gap between the three envelopes on Figure 1c is a coarse silhouette of a much sharper birth-set structure. Storage is thus wider than oscillation—modestly, not vastly—and propagation is the narrowest gate. I find Spearman against ; the informative quantity is the contrast, not the magnitude (still lifes and oscillators are the same enumeration at period 1 and , so a strong correlation is expected). The relevant partition is not static-versus-moving but translating-versus-non-translating: in-place recurrence at any period for a single -tuned storage axis, and translation is empirically separated from it at : only rules with (i.e. ) show compact-glider support, while in-place recurrence occurs on both sides of the line. The submanifold confirms this directly: rules support oscillators at 96/512 = 18.75% (within reach of the 21.85% full-space rate, all at period ) while supporting 0/512 compact spaceships. closes the translation channel essentially intact (Figure 4; mechanism in § 4.1).
as the factorisation axis. Because the factorisation is intrinsically a birth-versus-survival statement, the -odd coordinate is its natural projection. Among the 424 rules the distribution is sharply skewed to (median −0.222, 340 of 424 below zero), whereas the bulk is symmetric about zero by the -symmetry of rule space (Figure 3). This plays, for the asymmetry axis, the role Langton’s critical value played for the activity axis—a single scalar where capability concentrates—though the comparison is qualitative: Langton’s is on his multi-activity-weighted scale, not directly comparable to my unweighted (§2.2). But the paper’s framing applies to it in full—both and locate without certifying. The same slice is shared by thousands of rules, so landing on it is necessary for the propagation–storage structure but never sufficient for a glider; the value marks a region, not a guarantee. This fingerprint is specific to the compact footprint: among the 8,206 rules that support a spaceship the skew washes out (median 0, rank-biserial −0.12 against −0.42 at ), because a wider window admits birth-heavier orthogonal shippers. What survives footprint widening is the gate structure, not the sign: the 26 birth sets of are contained in the 64 of exactly (both counted at the existence threshold ), and both remain dominated by small birth sets (mean = 3.5 at , against a uniform peak at = 5). The birth-light asymmetry is thus what compact propagation costs, not a universal law; the narrow gate of small birth sets is the footprint-robust core. is the only odd coordinate the results here require; auxiliary odd-sector coordinates were explored but do not carry additional capability information beyond what already provides (§ 4.4).

Figure 3. Cardinality asymmetry over all 262,144 rules. Grey: bulk, symmetric about zero (forced by -symmetry). Crimson: (), sharply skewed to with median ; 340 of the 424 compact-glider rules sit at .
Discussion
The mechanism of the exclusion
“Propagation” here means coherent translation: a pattern returning to its own configuration displaced by a fixed offset, keeping its shape as it moves. It is not activity spread—disturbances travelling outward at up to one cell per step through a churning background, which any non-frozen rule does, rules included. The exclusion below concerns coherent movers, not motion.
A compact spaceship must create state ahead of itself and clear state behind, both enacted by the same object in the same update. What changes with footprint is how many vantage points the object provides for carrying that difference.
A pattern is exactly the Moore neighbourhood of a single cell: one cell’s update depends on the whole object, and no other cell—inside or outside—sees all of it. Indeed 3×3 is the largest pattern contained in a single neighbourhood, which is what makes “compact” intrinsic rather than an arbitrary window. At that footprint the front/back distinction must pass through one centre, and makes the update a function of the neighbour count alone. This is the heuristic; the argument that follows makes it concrete by identifying the specific rule-table resource a compact glider needs and closes.
State-distinction channels. The mechanism is easiest to state in terms of channels. Each of the 9 possible neighbour counts carries a potential state-distinction channel: the rule’s ability to give a live cell and a dead cell with the same count different fates. The channel at count is open when (birth and survival disagree at ) and closed otherwise. The centre-bit sensitivity measures the (binomially-weighted) fraction of channels open in a rule; (equivalently , equivalently ) closes all 9 channels at once. The argument below is that any compact glider needs at least one specific channel open, and has closed every one.
An empirical reduction. Every one of the 424 rules that supports a 3×3-seed spaceship at , without exception, satisfies and . The half has a complete proof. Consider the live cell of the pattern maximising ; the cell diagonally beyond it (at position ) has exactly one live neighbour, so under it births. The same argument applies in all four diagonal directions, so the pattern’s diagonal extent grows in every generation and can never recur as a translate. The requirement then follows from the geometry of compact translation: to move, a pattern must at some step birth a cell just outside its bounding box, and such a cell sees at most three live neighbours; since and are excluded, only can supply that birth. The prohibition is empirical: I have no combinatorial derivation of it, and report it only as an exhaustive fact from the enumeration. Under these constraints transfer to survival, and the shipping candidates collapse to with —32 rules. This is an organising observation that shrinks the case space, not a proof reduction, since is read off the same enumeration that already contains the exclusion. Exhaustive exploration over those 32 confirms none ships.
Different glider families use different channels. The failure mode differs across compact-glider families, and the differences are precisely differences in which channels each family requires. The rule’s is the set of open channels; but the channels a specific glider actually requires is a smaller set—only those counts at which the glider’s cycle both queries the birth side and the survival side with opposite required fates. Conway () has , and its glider requires exactly the channel; the half of the reduction already kills any extension of it. HighLife () has —two open channels—but its glider still only requires the channel (the additional disagreement is not queried by the glider’s cycle), so again kills it. Other families require channels the reduction does not force closed. A period-4 ship at requires the channels at 4 and 5; a period-10 diagonal pair at requires the channel at 6 (the disagreement is present in the rule but not queried by the ship); a period-6 orthogonal at requires channels at 4, 7 and 8. Each family needs at least one of its required channels to be open under , but closes all 9 channels simultaneously, and the enumeration confirms no instantiation exists at . The exclusion is a genus of case-checks, one per compact-glider family, each turning on which channels that family requires—unified only by the reduction.
The nose-bottleneck mechanism. The reduction carries geometric content that suggests—but does not prove—the specific channel one class of families would require. Consider a phase three-wide transverse to the direction of motion, under and . The cells just ahead of it see only the front row; a birth at any of them needs the front row to contribute exactly three live neighbours, which pins the front to 111 and allows only the middle-ahead cell to be born (the flanking-ahead cells see only two neighbours and stay dead). Motion is therefore forced through a single-cell nose bottleneck at count 3. This is where an intuition about the exclusion enters: the nose cell is born at count 3, and any interior live cell in the pattern at count 3 shares its fate under (both governed by ), so the leading/trailing distinction the shipping motion appears to need cannot straightforwardly be built from a count-3 state distinction. I offer this as a suggestive mechanism, not a proof: I have not shown that no -consistent phase evolution realises coherent translation through the nose bottleneck, only that the state-distinguishing tool the natural mechanism would use is unavailable. Some enumerated shipping families visit wider phases mid-cycle, and their required channels live at other counts (the family-specific case-checks above); a covering set of arguments extending this intuition into a proof remains open (§4.4).
The claim above is compact-footprint specific. A object is not contained in any single Moore neighbourhood; cells with disjoint partial views can carry the front and back roles at distinct positions, and the requirement drops away. 36 rules ship at and 40 at , all with every state-distinction channel closed (they are ), so “translation needs an open channel” is a claim scoped to the compact footprint, not a general property of shipping under . At the state-blind machinery instead exploits count-4 births: all four rules that ship at while remaining inside the 32-rule set at the compact footprint have (, , , ), suggesting the count-4 birth (available at because a head-cell no longer sees a single-row front) is what the nose bottleneck denied. The oscillator results localise the exclusion within the compact footprint: rules oscillate at near-background rate and are excluded only from compact spaceships—in-place recurrence does not need the leading/trailing distinction that translation, at , appears to. This mechanism is heuristic and case-analytic, not yet a theorem (§4.4): the exclusion is proven by enumeration; a proof from the reduction plus a covering set of family-specific channel arguments remains open.
The same centre-bit coordinate also grades the direction of translation at the compact footprint (Figure 2b): among spaceships orthogonal motion is confined to rules simultaneously high in and large in (the joint corner is 86% orthogonal, either lever alone only 21–34%), with diagonal the default elsewhere. The two coordinates act roughly additively and neither is reducible to the other. This diagonal dominance is itself compact-specific and inverts with footprint: at orthogonal spaceships outnumber diagonal ones by roughly seven to one, the reverse of the ratio. The geometric reading—that axial motion needs a sharper, denser leading edge than diagonal motion, and that the extra row of a wider window supplies exactly the structure orthogonal motion requires—I offer only as a hypothesis.
For storage the binding constraint is a -condition—a still life persists only if its surrounding dead cells have counts outside —while which interior configurations survive is set by . This is why correlates with only once one conditions on , closing the factorisation: gates, tunes.
Conway’s specialness
Conway sits at the intersection of three conditions, the first two combinatorial and the third dynamical: admits the glider; supports rich still lifes; and the random-IC attractor at routes trajectories into a regime that visits the glider on clean background. Rules in the family share a common supported glider—the first two conditions hold for every one of them—but only a survival rule that keeps the trajectory in the moderate-density band, neither flooding nor freezing, lets random dynamics express it. Conway’s is privileged not because it enables the glider but because it places the dynamics where the glider can appear.
A minimal pair. The first of these three conditions is sharper than it looks, and the cleanest way to see it is to apply to Conway itself. The mirror rule has the same activity as Conway (), the same sensitivity (), the same centre-cell sensitivity (), and even the same disagreement set . The two rules differ in exactly one coordinate: flips sign, against . Note that is a symmetry of the coordinate system and not of the dynamics—unlike , which is exact at the level of trajectories (§ 2.6)—so the mirror is a genuinely different rule, and nothing forces it to behave like Conway.
What the coordinates do here is instructive. is the only one of them that distinguishes the pair at all; and the sensitivities are -even and see the two rules as identical. But no coordinate excludes the mirror: is nonzero, so the one exact exclusion does not apply; is Conway’s own value; and only places the rule on the birth-heavy side, atypical for compact propagation but no guarantee either way. The system sees a difference and cannot resolve it. The enumeration can, and does: the mirror supports no compact glider.
The mechanism is the orientation, not the magnitude. Both rules are centre-sensitive at exactly one count, , and to exactly the same degree; what differs is which side of the disagreement the birth condition falls on. Conway holds at if the cell is already alive and declines to create it if dead—the shape-preserving condition a coherent mover requires (§ 4.1). The mirror inverts both: it creates at and kills at , so every adjacent live pair births into the dead cells that see both, and the pattern bloats instead of translating. The enumeration bears this out in aggregate—no rule with supports a compact glider, 0 of the 424. Conway’s birth condition is thus privileged not by how much of the table it fires, which the mirror matches exactly, but by the direction in which its one centre-sensitive count is oriented; and the pair is the necessary/sufficient thesis in miniature, two rules separated by a single sign, distinguished by a coordinate that locates without certifying.
I can ask the converse question directly: among the rules that support both a glider and a still life but are absent from the standard catalogue, are any dynamically Conway-like—not merely capable, but producing a Game-of-Life–style ecology of blinkers, blocks and other small objects from random starts? The census is deliberately Conway-object-biased: it asks whether a rule reproduces Life’s small-object zoo, not whether it hosts a rich small-object zoo at all. A rule with a rich but non-Life-like object ecology would fail the census by construction, so the finding below is best read as “rules passing a Conway-object census are Conway-family on visited configurations,” not as a claim about all rich object ecologies in the family. Of the 424 rules with , the 2 named anchors and 2 further rules with no still life () are set aside, leaving 420 uncatalogued build-and-hold rules. I screen these on behavioural criteria alone—a stabilising tail density (met by 314 of the 420 at some density), followed by a direct census requiring that canonical Game-of-Life still lifes and oscillators dominate the small live structures (a coarse, per-rule analogue of the soup census of Catagoule [11]), plus a richness floor of at least one canonical still life (block, beehive, loaf, boat or tub) present in the tail, so that a rule producing only sparse blinkers—which are canonical oscillators and would trivially satisfy the raw dominance test—does not qualify. All criteria are applied without reference to the rule table, so that table similarity is a finding rather than a filter. Without the richness floor 88 rules pass; with it, 24 remain, and the 14 excluded are precisely the blinker-only degenerates whose census is Life-like in the letter but not in the spirit. Only then do I compare the 24 rich-ecology rules’ tables to the anchors: all 24 agree with Conway or HighLife on every visited (centre, count) pair (15 Conway, 9 HighLife), differing only at neighbour counts 7 and 8, which require locally saturated regions and do not fire at the tested density. So every uncatalogued rule that both stabilises and grows a rich Life-like object ecology from random starts is, on its visited configurations, a Conway or HighLife variant—the pool contains no rule whose table is distinct from both anchors yet reproduces their small-object zoo. This is the converse of the paper’s leading theme—capability resides in the table but need not be expressed—made concrete on the expression side: two rules can differ wherever their tables are defined yet be dynamically indistinguishable, because behaviour is fixed by the table only on the configurations a trajectory visits. Scope: the screen identifies rules Game-of-Life–like in their small-object ecology, not the complete set of dynamically interesting rules—a rule with a rich but non-Life-like object zoo would pass unrecognised—and the anchor identity is scoped to the tested densities, where the differing high count is rare rather than combinatorially impossible.
Both results are, at root, a genotype–phenotype distinction. The rule table is the genotype: it fixes what is possible and can be read directly, which is why capability is certifiable. What a random trajectory develops is the phenotype: it depends on initial conditions and dynamics, so it can only be estimated, and it need not realise everything the genotype permits—a rule may carry a glider it never expresses (§ 3.3), just as two distinct genotypes may yield the same phenotype when the differing genes are never transcribed (the anchor variants above). The birth–survival asymmetry adds a smaller morphological note: translation requires a rule that distinguishes “ahead” from “behind”, a broken front–back symmetry that in-place recurrence does not need—loosely, the asymmetry that a moving body plan requires and a radially symmetric one does not.
External corroboration, and the scope of the claim. Comparing against the literature’s catalogue of complex life-like rules sharpens both what these results establish and what they do not. Of twenty rules recorded as Class IV or computationally universal in the standard references [1] (Eppstein 2010), exactly one is centre-blind (Replicator), and it is Class III rather than Class IV. This is consistent with the paper’s exclusion at the level of behavioural classification, though “Class IV” and “supports a compact glider” are different properties, and one rule in a twenty-rule list is thin evidence—call it consistent with the exclusion, not corroboration of it. Second, my behavioural screen, run without reference to the literature list, independently recovered four of these named rules—DryLife, EightLife, Pedestrian Life and LowDeath—and the subsequent table comparison correctly identified each as a Conway- or HighLife-family rule differing only at unvisited high neighbour counts. Third, and against these, the scoping must be stated plainly. Eleven of the twenty have and so could never enter my screen: Day & Night (, ), 2×2, Move and Life without Death are complex through engineered constructions and large-footprint propagation that a compact enumeration cannot see. A vanishing compact-glider count therefore does not imply dynamical simplicity, and the column of Figure 5 mixes inert rules with rules whose propagation is merely non-compact. The six spontaneous Class IV rules—the Conway family—do cluster where my coordinates predict (, , ). My claim is accordingly the narrow one: among rules supporting a compact glider, spontaneous Game-of-Life–like ecology concentrates on Conway, HighLife and their behaviourally redundant variants. It is not the claim that these are the family’s only complex rules—they demonstrably are not.
Why compact tracks spontaneous. A naive combinatorial reading of expression would run through the seed itself: a specific -cell pattern appears at a given site of a Bernoulli- field with probability for an -cell window, so the chance of a particular seed being drawn falls exponentially in its cell count as the window grows from 9 to 16 to 25 sites, predicting that larger gliders should essentially never appear. That reading is wrong, and this paper’s own data refutes it: the thirty-three silent rules of § 3.3 support a compact glider whose -cell seed occurs constantly in any random field, yet produce none in 325 trials. Gliders in random dynamics do not arrive by being drawn in the initial condition; they are assembled by local collisions, and the governing quantity is the measure of the object’s basin under the dynamics, not the probability of its seed—exactly the quantity I argue is estimable only by simulation and never certifiable from the table. Basins do shrink with footprint—more cells must be simultaneously right—but far more gently than the seed picture would predict: empirically the silent fraction rises from 7.8% at to about 35% at , a factor of about 4.5, not the exponential collapse the seed-probability picture implies. Expression degrades with footprint; it does not vanish. Compact objects have basins large enough to be reached from a random start, which is why Conway’s gliders appear unbidden and why my random-IC screen can see them at all. Non-compact structures—Day & Night’s spaceships and guns, the machinery underlying 2×2 and Move—were found by construction rather than by observing soup, and their basins under random ICs are correspondingly negligible. The compact/non-compact boundary therefore very nearly coincides with the spontaneous/engineered one: a strong tendency across the known catalogue rather than a law: Day & Night does emit small still lifes and short-period oscillators from soup (documented on Catagolue) even though its glider does not spontaneously appear, so “no soup structure” is not the right claim; the right one is that the compact glider in particular has a negligible random-IC basin. A tendency with a mechanism (basin size). This is precisely why my census, which measures spontaneously-reachable compact propagation, recovers Conway and HighLife yet cannot see Day & Night: not because Day & Night lacks propagation, but because its propagation lives at a scale random initial conditions do not assemble.
Rule-table coordinates and the phase-transition analogy. Langton’s sweep is a temperature analogy: tune the table, watch the dynamics pass from frozen through complex to chaotic (citep)(Langton 1990). The family’s exact swap symmetry lets the analogy be completed rather than merely invoked. In the resulting dictionary the coordinates each have a place. The temperature-like axis is the sensitivity , not the activity —damage-spreading treatments of the order/chaos transition locate criticality in average sensitivity (citep)(Derrida & Pomeau 1986)—with a temperature proxy. The -odd is the analogue of symmetry-breaking field: a direction in parameter space, vanishing on the symmetric locus, explicitly breaking off it. Its field magnitude is , yet zero exactly at —and the identity of §2.3 is then the statement that signing the magnitude recovers the signed field. A -odd order parameter would have to be a -odd trajectory observable, and none of the swept quantities qualifies.
The dictionary is algebra; whether any transition accompanies it is a dynamical question, and the answer is no. Sweeping at fixed over uniformly sampled rules at each attainable cardinality pair, and measuring on a torus the density susceptibility , the correlation length , and the Binder cumulant of (citep)(K. Binder 1981), unrelated to the P.-M. Binder plane of §2.3, I find none of the signatures: monotone with no interior peak, no cumulant crossing, accounting for under 1% of the variance in —the rule-to-rule spread within a fixed exceeds the variation across by two orders of magnitude. Nor could finite-size scaling settle the matter: at fixed takes between 1 and 10 values (depending on the parity of ; at up to ten values are attainable), spaced 0.222 apart, and criticality requires a continuous control parameter. The family’s one live transition is instead the absorbing-state freeze boundary of § 4.4, whose order parameter—the stationary activity density—is -even. Transition structure lives in the even sector; the odd sector, where lives, carries capability without a transition. Consistently, the locus where compact gliders concentrate shows the smallest density fluctuations of the sweep: capability sits in the calm, far from any critical point. In this language the exclusion of §3.2 reads: compact-glider-supporting rules occupy only the “field-on” side of the phase diagram, and the diagonal is labelled by —a coincidence of level sets, not a causal implication.
The coordinate logic beyond life-like rules
The construction has a structural template independent of the birth/survival distinction. Any CA family whose rule table carries an involutive symmetry admits the same question: which coordinate on the table is odd under that involution, so that it vanishes on the symmetric locus and can serve as a signed asymmetry. Langton’s counts firings without regard to which side they occur on, so it is even under the involution and cannot be that coordinate; the natural candidate is the odd asymmetry. For life-like rules the involution is the birth↔survival swap , the odd coordinate is (equivalently in weighted form), and the symmetric locus is the diagonal . The exclusion of §3.2 is then structural in this template: it is the statement that one small-object channel closes exactly on the -fixed set.
The same logic transfers to elementary one-dimensional cellular automata under the left↔right reflection , a canonical ECA symmetry that structures the classical classifications [22, 24, 14]. Langton’s is -even, so the natural analogue of under the parity logic is a left–right asymmetry that vanishes on amphichiral rules (one form: ). I do not develop this coordinate here; reflection symmetry in ECA is a mature subject and my family is 2D outer-totalistic for reasons unrelated to it. The transferable point is what state-factorisation does to the form of the odd coordinate: a life-like rule table splits by centre state, so the -odd asymmetry admits a cardinality expression distinct in form from the signed centre sensitivity even though the two coincide with the weighted once binomial multiplicities are restored. An ECA table does not factor by left/right, so the analogous -odd coordinate can only be built as a signed reflection-defect—structurally the analogue of , not of . The two objects that coincide in the outer-totalistic family would be a single object in a family without state factorisation; the split is a feature of the family, not of the parity logic itself.
Limitations and open directions
The main enumeration is at footprint, , exhaustive over all 262,144 rules. The enumeration is also exhaustive over all rules (8,206 support a spaceship, 3.13%); the enumeration is exhaustive over the 512 rules and sampled elsewhere. Three footprint observations follow. First, containment is perfect—all 424 rules retain , with 7,782 further rules gliding only at —so the compact enumeration is a strict lower bound on propagation capability. Second, the storage axis is scale-invariant (within-slice Spearman at and 0.556 at ). Third, the exclusion is compact-specific but does not simply dissolve with scale: 36 of the 256 vacuum-stable rules admit a spaceship and 40 a , yet those that combine a glider with static storage stay rare (4, then 8) and fall progressively further from Conway (still-life gap 6.3× at , 9.2× at ), so the strict exclusion softens into a widening handicap. and beyond are presently out of reach in both time and memory (the number of orbit-representative seeds grows from at to , , and at -, -, ). The factorisation and the exclusion are established empirically and argued heuristically; a proof that no rule admits a compact spaceship would convert the exclusion into a theorem and remains open. The partial mechanisms of §4.1 (the argument; the nose-bottleneck sketch at three-wide transverse phases) go partway; a covering set of family-specific arguments would complete it.
A stronger negative bounds what this coordinate system can do. Every additional rule-table coordinate I constructed and tested collapsed, on inspection, into a proxy for the birth gate : a binomially-reweighted -odd sibling of discriminates worse and reverses sign conditional on ; its uniformly-weighted -even counterpart shares s zero set and adds nothing; and apparent fingerprints in the disagreement set (for instance that 91% of rules have ) reduce to the gate and the Conway survival pattern in other variables, since all rules have and none has . A -odd, -odd edge-contiguity coordinate—the imbalance between the on/off transition counts of and as a signed count—was also tested on , retained only a small fraction of 's effect on ; it is subordinate to and is not used in this paper’s analysis. The coordinates locate the gate— because small correlates with , because an empty disagreement set forbids the centre-bit discrimination the gate requires—but none transcends it. Whether a cardinality-based coordinate can carry information about beyond the birth set’s identity is, on this evidence, doubtful.
The silent population was characterised across a 13-density, 25-seed sweep, but the underlying claim remains a measure, not an absolute: a basin of vanishing but nonzero weight cannot be excluded by sampling. A follow-up seeded with known spaceships rather than random ICs would map the basin structure directly, and the flood/freeze regime that gates expression—determined by the table yet not captured by a mean-field predictor—invites a sharper theory of which rules settle into the moderate-density band that propagation requires. That regime boundary is also where a genuine criticality question lives: every rule with has an absorbing vacuum, and the freeze transition is an absorbing-state transition of the kind studied in probabilistic cellular automata [12]. Testing it requires a continuous control parameter, which no cardinality coordinate supplies; mixing two rules stochastically would furnish one, at the cost of leaving the deterministic family this paper concerns. Whether the freeze boundary falls in the directed-percolation universality class is a well-posed question I do not address.
Conclusion
The paper’s core contribution is combinatorial: exhaustive -seed enumerations of compact spaceships (), still lifes () and oscillators () across the entire family of 262,144 life-like rules, and the spaceship enumeration on all rules. From those enumerations three results emerge cleanly: (i) a hierarchical factorisation—the birth condition gates whether small objects exist at all (, sizes 26/128/176 of 512), and the survival condition tunes storage richness once the gate is open; (ii) the exact exclusion of centre-blind () rules from compact-glider support, proven by enumeration; two empirical regularities of that enumeration ( and across every rule) collapse the shipping candidates to 32 rules on the diagonal as an organising observation, though not yet as an independent proof reduction; (iii) a silent population of 33 rules that support a compact glider combinatorially yet never express one under random sampling, sharpening the distinction between capability (certifiable from the table) and expression (only estimable from dynamics). The rule-table coordinates , and locate where each capability concentrates in coordinate space but do not certify it; the sharp facts of the paper live at the birth-set level, and the coordinates are useful primarily as ordering axes that expose those facts. Two open threads remain: converting the exclusion from an enumeration result into a proof via a covering set of family-specific channel arguments (§ 4.1), and testing whether the freeze transition at the boundary of the bulk is directed-percolation in universality class (§ 4.4).
Code and data availability
Both the code and the derived data supporting this paper are archived at https://github.com/vairchn/ca_lifelike_rulespace under the MIT licence. The repository contains the coordinate implementation (a self-contained NumPy module that computes , , , and over all 262,144 rules; additional auxiliary coordinates and trajectory-sensitivity traces are also included but are not used in the paper), the per-rule enumerations of compact spaceships, still lifes and oscillators at (and the exhaustive spaceship enumeration used in § 3.4), the composite scores table that supplies the dynamical probe (§ 2.6), the 33–5 silent-population split of Figure 5 (§ 3.3), and one figure-regeneration script per manuscript figure. The parity self-check for every coordinate reported in § 2.4 is included and runs on the full rule set.
References
- [1]A. Adamatzky, editor. Game of Life Cellular Automata. Springer, 2010.DOI
- [2]C. Bays. Introduction to cellular automata and Conway’s Game of Life. In A. Adamatzky, editor, Game of Life Cellular Automata, pages 1–7. Springer, 2010.DOI
- [3]D. I. Bell. HighLife — an interesting variant of Life. Technical report, 1994. Available at the LifeWiki HighLife page (https://conwaylife.com/wiki/OCA:HighLife).
- [4]E. R. Berlekamp, J. H. Conway, and R. K. Guy. Winning Ways for Your Mathematical Plays, Vol. 2. Academic Press, 1982.
- [6]P.-M. Binder. A phase diagram for elementary cellular automata. Complex Systems, 7:241–247, 1993.
- [8]D. Eppstein. Searching for spaceships. In More Games of No Chance, MSRI Publications 42, pages 433–453. Cambridge University Press, 2002. (arXiv:cs/0004003.)DOI
- [10]M. Gardner. Mathematical games — the fantastic combinations of John Conway’s new solitaire game “life”. Scientific American, 223(4):120–123, 1970.
- [11]A. P. Goucher. Catalogue: a distributed census of natural pattern occurrence in cellular automata, 2026. https://catalogue.hatsya.com. Accessed July 2026.
- [12]H. Hinrichsen. Non-equilibrium critical phenomena and phase transitions into absorbing states. Advances in Physics, 49(7):815–958, 2000.
- [13]C. G. Langton. Computation at the edge of chaos: phase transitions and emergent computation. Physica D, 42(1–3):12–37, 1990.
- [14]W. Li and N. H. Packard. The structure of the elementary cellular automata rule space. Complex Systems, 4(3):281–297, 1990.
- [15]M. Mitchell, P. T. Hraber, and J. P. Crutchfield. Revisiting the edge of chaos: evolving cellular automata to perform computations. Complex Systems, 7:89–130, 1993.arxiv.org/abs/adap-org/9303003
- [16]N. H. Packard. Adaptation toward the edge of chaos. In Dynamic Patterns in Complex Systems, pages 293–301. World Scientific, 1988.
- [17]E. Peña and H. Sayama. Life Worth Mentioning: Complexity in Life-Like Cellular Automata. Artificial Life, 27(2):105–112, 2021.DOI
- [18]E. Sapin, O. Bailleux, and J.-J. Chabrier. Research of a cellular automaton simulating logic gates by evolutionary algorithms. In Genetic Programming (EuroGP 2003), LNCS 2610, pages 414–423. Springer, 2003.
- [19]L. S. Schulman and P. E. Seiden. Statistical mechanics of a dynamical system based on Conway’s Game of Life. Journal of Statistical Physics, 19(3):293–314, 1978.DOI
- [20]G. Y. Vichniac. Boolean derivatives on cellular automata. Physica D, 45(1–3):63–74, 1990.DOI
- [21]J. von Neumann. Theory of Self-Reproducing Automata. Edited and completed by A. W. Burks. University of Illinois Press, 1966.DOI
- [22]S. Wolfram. Statistical mechanics of cellular automata. Reviews of Modern Physics, 55(3):601–644, 1983.
- [24]S. Wolfram, editor. Theory and Applications of Cellular Automata. World Scientific, Singapore, 1986.
- [25]A. Wuensche. Classifying cellular automata automatically. Complexity, 4(3):47–66, 1999.