Larger Than Life: Expanding the Neighborhood
When Kellie Evans examined the patterns produced by her range-5 cellular automaton rules in the mid-1990s, she found spots. Not symbolic spots, not metaphorical spots — patterns of stable, rounded, pigmented regions separated by lighter background, arranged across the field with the irregular spacing of pigment cells in animal skin. The patterns matched, visually, the dorsal markings of leopards and the lateral stripes of zebrafish. She was not trying to model biology. She was studying what happens when you make the neighborhood bigger.
This was the first indication that the neighborhood size in Life-like rules is not merely a computational parameter. It is the dial that controls how biological the system looks.
The Extended Neighborhood
In Conway’s Life, every cell consults its 8 immediate neighbors — the Moore neighborhood, a 3×3 square minus the center cell. The rule is expressed in B/S notation: Born if 3 neighbors, Survive if 2 or 3. This is called a range-1 rule, because the neighborhood extends 1 cell in every direction.
Larger Than Life (LtL) rules generalize this by introducing a range parameter R. A range-R rule uses a (2R+1)×(2R+1) square neighborhood — all cells within R steps of the center in the Chebyshev metric. The neighbor count grows quadratically:
| Range R | Neighborhood size | Neighbor count |
|---|---|---|
| 1 | 3×3 | 8 |
| 2 | 5×5 | 24 |
| 3 | 7×7 | 48 |
| 5 | 11×11 | 120 |
| 10 | 21×21 | 440 |
At range 5, a single cell’s birth or survival is determined by up to 120 neighbors. At range 10, by up to 440.
The critical design insight is that birth and survival thresholds in LtL rules are specified as fractions of the total neighbor count, not as absolute numbers. In Conway’s Life, “Born if exactly 3 neighbors” is 3/8 ≈ 37.5% of the maximum neighborhood. “Survive if 2 or 3 neighbors” is roughly 25–37.5% of maximum. These fractions are the invariants — the mathematical core of the rule — that can be scaled up to any radius.
A range-R LtL rule is written as B[b_lo, b_hi]/S[s_lo, s_hi]/R, where the bracket values are neighbor counts (not fractions), chosen proportionally to (2R+1)² - 1. The analog of Conway’s Life at range 5 would have birth thresholds near 0.375 × 120 ≈ 45 and survival thresholds between 0.25 × 120 ≈ 30 and 0.375 × 120 ≈ 45. These are not the same rule as Conway’s Life — they live in a different part of a much larger parameter space — but they are motivated by the same underlying fractions.
This notation and the systematic study of LtL rules is the contribution of Kellie M. Evans, who developed the framework in her 1996 PhD thesis at the California Institute of Technology: Larger Than Life: It’s So Nonlinear.
Kellie Evans and the Original Work
Evans’s thesis is methodical where it could have been speculative. She defines the LtL framework formally — the range parameter, the fractional threshold notation, the class of rules it generates — and then runs systematic experiments across the parameter space.
The central finding is that as R increases, the qualitative character of LtL dynamics shifts away from the crisp, binary world of Conway’s Life and toward something that resembles continuous-field systems. At R = 1 or 2, rules in the LtL family look like Life-like rules: they produce gliders, oscillators, and still lifes with hard edges and grid-aligned motion. At R = 5 or above, the same structural class of rules produces patterns with rounded forms, smooth gradients at boundaries, and — in many parameter regimes — stable spatial patterns with a spatial frequency set by the neighborhood radius.
That spatial frequency is the key to the biological resemblance. Turing’s 1952 paper “The Chemical Basis of Morphogenesis” showed that two-chemical reaction-diffusion systems spontaneously produce periodic spatial patterns — stripes, spots, or labyrinths — with a characteristic wavelength determined by the ratio of the chemicals’ diffusion rates. The pattern scale in LtL is set analogously by R: the neighborhood size determines the distance over which local averaging occurs, and this distance selects the spatial frequency of any stable patterns. At R = 5, the selected wavelength is on the order of the neighborhood diameter — roughly 11 cells — and the resulting spotted or striped patterns have precisely this spacing.
Evans notes explicitly in her thesis that the spotted patterns produced by several LtL rules at range 5 are visually indistinguishable from biological pigmentation patterns. This is not coincidence. The mathematical mechanism is the same: local self-activation, lateral inhibition, and a spatial scale set by the inhibition range. In biology, that range is set by diffusion rates. In LtL, it is set by R.
The thesis introduced a zoology of LtL rules that Evans named and characterized by behavior class. Three of them have become the standard examples: Bosco, Bugs, and Waffle.
Notable Rules
Bosco
Bosco is Evans’s flagship rule: B34..58/S34..45/R5 (range 5, with birth requiring 34–58 neighbors and survival requiring 34–45 out of the 120-cell neighborhood). Bosco produces stable gliders that travel at angles impossible in standard Life, along with oscillators and the spotted stationary patterns Evans compared to animal markings. It is LtL’s answer to the question “does this framework produce anything interesting?” — the answer is yes, unambiguously.
The gliders in Bosco look different from Conway’s glider. They are rounded, roughly circular objects — blobs rather than the angular five-cell diagonal figure of the original. They travel across the background in non-axis-aligned directions. Their existence demonstrates that glider-like objects are not a specific property of Conway’s rule or its 3×3 neighborhood; they are a feature of this entire class of local-rule systems, surviving at any scale where the rule supports both local activation and lateral inhibition.
For context on why the existence of traveling objects matters computationally, see Spaceships →.
Bugs
Bugs uses a smaller range — typically R = 5 — with birth and survival thresholds tuned so that isolated clusters of live cells behave like coordinated agents. The patterns produce dense, moving groups that split, merge, and reorganize over time. Individual clusters are persistent objects with irregular, organic shapes. They do not resolve into a single stable pattern; instead, the field maintains a population of interacting objects in a kind of turbulent equilibrium.
Bugs is notable because it sits in a region of parameter space where the dynamics are neither fully ordered (stable patterns, regular gliders) nor fully disordered (random noise). The system sustains complex, long-lived structures without either crystallizing or dissolving. Langton identified this regime — the boundary between order and chaos — as the region where complex computation tends to emerge in CA systems.
Waffle
Waffle produces a qualitatively different behavior: a quasi-regular grid of structural regions separated by stable boundaries. The field settles into a pattern resembling a physical waffle — square or rectangular cells of high density separated by lower-density walls — and this structure persists indefinitely. The boundary network is not perfectly regular; defects propagate and anneal over time, but long-range order is stable.
Waffle demonstrates that LtL rules can produce patterns with two distinct length scales: the scale of the individual structural units (set by R) and the scale of the overall grid (set by the simulation domain). This two-scale behavior has analogs in physical systems: crystal grain boundaries, lipid domains in cell membranes, and the columnar structures in biological cortex.
Behavior at Large Scale
Three properties of LtL dynamics change systematically with R:
Pattern smoothness. At range 1, patterns have hard, pixelated edges because the smallest resolvable feature is one cell. At range R, the smallest feature that the rule can “see” is one cell, but the smallest feature that can be stable has a diameter on the order of R — smaller features are averaged away by the large neighborhood. This is spatial low-pass filtering, and it produces smooth, rounded boundaries as a structural consequence.
Direction of motion. Conway’s glider moves diagonally at c/4 because the grid constrains it. LtL gliders at large R can move in approximately any direction. The grid is still present — these are still discrete cells — but the large neighborhood acts as a spatial average that partially suppresses grid anisotropy. Motion at non-axis-aligned angles becomes possible and natural. This is the same phenomenon that Rafler’s SmoothLife demonstrates by dissolving the grid entirely: the directional quantization of motion in Conway’s Life is an artifact of the minimal neighborhood, not a property of the rule structure.
Speed of light. In Conway’s Life, the maximum propagation speed is 1 cell per generation — the “speed of light” c, by analogy with physics. A pattern cannot affect cells more than N cells away in fewer than N generations. In a range-R LtL rule, the maximum propagation speed is R cells per generation. This follows directly from the definition: a cell at position (x, y) can only affect a cell at position (x + d, y) after at least ⌈d/R⌉ generations, because the causal influence hops at most R cells per generation. The LtL speed of light is c_R = R.
This means that large-R spaceships travel at speeds that would be “faster than light” relative to Conway’s Life but are well within the LtL speed limit. A Bosco glider moving at 2 cells per generation is traveling at speed 2 in a system with a speed limit of 5 — well within bounds. The speed limit has expanded with the neighborhood.
The practical consequence: in LtL, the relationship between physical distance and signal delay changes. Large-R systems can propagate information faster, which changes the geometry of any dynamic or computational structure built from these rules.
Connections to Biology
The biological motivation for LtL is explicit in Evans’s thesis and is not merely aesthetic. In real biological tissue, a cell’s chemical environment is shaped by molecules that diffuse from neighboring cells — and diffusion is not a nearest-neighbor process. A molecule released by a cell at position (x, y) reaches cells 10, 20, or 50 cell-diameters away within biologically relevant timescales. The effective “neighborhood” of a cell in tissue chemistry is large.
This means that the 8-neighbor Moore neighborhood of Conway’s Life is, as a model of biology, simply too small. The relevant neighborhood for chemical signaling in tissue is closer to what LtL provides at moderate to large R. Evans’s LtL framework can be read as the answer to: what happens to Life-like rules when you scale the neighborhood to biologically realistic sizes?
The answer is Turing patterns. The mathematical relationship is not coincidental. In the large-R limit, the LtL update rule approximates a reaction-diffusion equation. The birth-and-survival logic provides the reaction kinetics — local activation and lateral inhibition — and the large neighborhood provides the spatial averaging that acts like diffusion. The resulting patterns (spots, stripes, labyrinths) are exactly what Turing’s theory predicts.
This is not the same as saying LtL is literally a reaction-diffusion system. It is not. But it occupies the same region of dynamical behavior: where local activation and non-local inhibition interact to produce stable periodic spatial patterns. At range 1, Conway’s Life is too local to access this regime systematically. At range 5 or above, it is structurally in this regime, and the biological resemblance is not incidental.
The bridge to SmoothLife is direct. SmoothLife replaces the discrete grid with a continuous field and the square Moore neighborhood with a disk-and-ring kernel: an inner disk (the “cell”) and an outer annular ring (the “neighborhood”). The outer ring radius r_a is the continuous analog of R in LtL. Both frameworks produce Turing-like patterns when the neighborhood is large relative to the cell size. They approach the same continuous-field limit from different directions: LtL by expanding R while keeping the grid, SmoothLife by dissolving the grid entirely while keeping the neighborhood geometry.
The Dial Between Two Worlds
Conway’s Life is a single point in a large parameter space. The neighborhood radius R = 1 is not a principled choice; it is the smallest possible value, the choice that was convenient for a game played on graph paper in a Cambridge common room in 1969.
LtL is what you find when you turn R up. The immediate effect is more neighbors and larger absolute thresholds. The downstream effect is a systematic transformation of the character of the dynamics: harder edges become softer, axis-aligned motion becomes arbitrary-direction motion, the speed limit expands, and the patterns that emerge begin to resemble the outputs of biological developmental processes.
At R = 1, you have Conway’s Life: crisp, computational, grid-aligned, with integer neighbor counts and a fixed speed limit of 1. At large R, you have something that looks like a Turing reaction-diffusion system: smooth, biological, orientation-independent, with a speed limit that scales with the neighborhood. The transition is continuous.
Turning R is a controlled experiment. It reveals which of Life’s properties are accidents of its minimal neighborhood — grid-aligned gliders, the specific speed limit of c, the pixelated boundaries — and which survive at all scales: persistent objects, traveling objects, local activation and lateral inhibition, the capacity for complex self-organizing behavior. The properties that survive are the ones that belong to the rule structure. The ones that disappear belong to the grid.
Evans’s leopard-spot patterns were not a curiosity. They were the first clear signal that the neighborhood radius is a fundamental parameter — that what Life’s rules produce at small scales and what reaction-diffusion chemistry produces in biological tissue are connected by a single continuous family of rules, and the connection is the size of the neighborhood.