The Garden of Eden: Patterns With No Past
Here is a paradox embedded in the rules of Conway’s Game of Life.
Every pattern has a future: apply the four rules, and the next generation appears. The system is deterministic — the future is uniquely determined by the present. But the rules do not work the same way in reverse. Multiple different configurations can evolve into the same successor, which means the function “one generation forward” is not one-to-one. It destroys information: you cannot, in general, run Life backwards.
The consequence is startling. If the map from generation N to generation N+1 is not a bijection — if it loses information — then some configurations in generation N+1 have no pre-image. There exist valid Life configurations that cannot have arisen from any preceding state. They can only exist as initial conditions. They have no history.
These patterns are called Gardens of Eden.
The Meaning of “No Predecessor”
To be precise: a Garden of Eden is a finite configuration P such that there is no configuration Q — across the entire infinite grid — satisfying the property that applying Life’s rules to Q produces P. Not merely “no simple predecessor” or “no small predecessor.” No predecessor at all. If P is a Garden of Eden, then P has never been the output of any computation involving Life’s rules.
This is a statement about the global dynamics of an infinite grid. Verifying it for any specific pattern P requires examining, in principle, every possible configuration Q on the infinite grid. No finite algorithm can do this directly — which is part of what makes finding specific Gardens of Eden technically challenging. Verification requires the claim that for every finite region surrounding P, any configuration of that region has no predecessor, and that argument must be made systematically.
The alternate name for a Garden of Eden is orphan pattern — a pattern with no parent in the family tree of Life configurations.
The Existence Proof: Moore and Myhill
The existence of Gardens of Eden was established mathematically long before Conway invented Life, and before anyone found a specific example.
In 1962, the mathematician Edward F. Moore proved a theorem about a broad class of cellular automata: if a cellular automaton has what he called “twins” — two different configurations that produce the same successor — then it must also have “orphans” — configurations with no predecessor. Moore proved one direction: twins imply orphans.
In 1963, John Myhill proved the converse: orphans imply twins. Together, the two results give the Garden of Eden theorem (also called the Moore-Myhill theorem): a cellular automaton has a Garden of Eden if and only if it has twins — two configurations that are mapped to the same output.
In Conway’s Life, twins are easy to find. Consider any 5×5 region of all dead cells: it maps to a specific output regardless of what surrounds it. Now consider a 5×5 region with a single live cell in the center and all other cells dead: if the single live cell has zero live neighbors, it dies, and this region also maps to all-dead in the next generation (within the interior cells). These two 5×5 configurations produce the same interior output — they are twins. Therefore, by Moore’s theorem, Life must have Gardens of Eden.
The proof gives us existence without construction. We know Gardens of Eden exist in Life — the logic is airtight — but the proof tells us nothing about what they look like or how large they must be.
The First Explicit Garden of Eden
The gap between the existence proof (1962–63) and an actual example (1971) is eight years. Finding a specific Garden of Eden required a computational strategy, not just mathematical argument.
In 1971, Roger Banks and a group of researchers at the MIT AI Lab — the same institution where Gosper had found the Glider Gun the year before — conducted a systematic computer search. They were looking for a finite pattern with the property that no extension of it to a larger pattern could be a valid predecessor of the configuration under Life’s rules.
The strategy was to work outward from a candidate configuration. If the candidate is to have no predecessor, then for every possible state of the cells surrounding it, the extended configuration must fail the “predecessor” condition — it must produce a different output when Life’s rules are applied forward. Checking this systematically for finite candidates was achievable on the computing hardware of 1971, though barely.
Banks’s group found the first Garden of Eden: a pattern fitting in a 9×33 bounding box, containing 226 live cells. The verification was done by exhaustive backtracking search over possible predecessors.
The pattern itself is not beautiful — it looks like a random scattering of live cells within a narrow rectangle. Beauty was not the point. The point was existence: here, specifically, is a configuration with no predecessor. Life’s forward dynamics cannot reach this pattern. It stands outside the history of any Life universe; it can only be an initial condition, imposed from outside.
The Drive Toward Smaller: Record Holders
After Banks’s 226-cell, 9×33 example, researchers began asking: how small can a Garden of Eden be?
The answer has gotten progressively smaller over five decades.
Achim Flammenkamp became the dominant figure in this search during the 1990s and 2000s. Working at the University of Bielefeld, Flammenkamp built exhaustive computational searches for small orphan patterns. In 1991, he found Garden of Eden 2: 143 cells in a 14×14 bounding box. In 2004, he found two more records in quick succession: Garden of Eden 3 (81 cells, 13×12 box) and Garden of Eden 4 (72 cells, 12×11 box).
Flammenkamp’s records fell in September 2009, when Nicolay Beluchenko found a Garden of Eden with 69 cells in an 11×11 bounding box.
In December 2011, another record fell: Garden of Eden 6, with 56 cells in a 10×10 bounding box.
The current records (as of the mid-2020s) continue to be pushed lower. The search has also established what is impossible: no Garden of Eden exists within a 7×7 bounding box. Any Garden of Eden must occupy at least an 8×8 region, and likely larger.
Why Gardens of Eden Are Hard to Find
Intuitively, you might expect Gardens of Eden to be common — after all, the proof that they exist uses only a rough counting argument about information loss. But in practice, they are surprisingly rare and difficult to locate by random search.
The reason is that most configurations do have predecessors. Life’s rules are chaotic and expansive: given almost any configuration, you can work backwards and find many possible predecessor configurations for it. The set of configurations with no predecessor — the actual Gardens of Eden — is sparse in the space of all configurations, even though it is infinite.
Finding a specific Garden of Eden requires showing, for a specific candidate, that none of the astronomically large set of possible predecessors works. This is inherently a search problem, and the search space grows exponentially with the size of the candidate. Verifying a small candidate (56 cells in a 10×10 box) is computationally intense; verifying a large candidate is essentially infeasible for the current generation of hardware.
The techniques used in modern Garden of Eden searches are sophisticated constraint-propagation algorithms — similar in spirit to SAT solvers — that efficiently prune the search space by propagating local constraints outward from the candidate pattern.
Gardens of Eden and the Arrow of Time
The deeper significance of Gardens of Eden is not about interesting patterns or computational records. It is about the structure of time in Life.
Life’s rules implement a deterministic map from each generation to the next. That map is irreversible: it is surjective (every configuration has at least one successor — you can always compute the next generation) but not injective (multiple configurations can produce the same successor). The failure of injectivity is the failure of reversibility.
In physics, this is the analog of entropy and the second law of thermodynamics. A physical system with reversible dynamics — like a classical mechanical system with no friction — can, in principle, be run backwards. A system whose dynamics lose information — like a system with friction, or a cellular automaton that is not injective — cannot. The future is determined by the present, but the present is not determined by the future.
Gardens of Eden are the sharp edge of this irreversibility. They are the configurations that fall outside the image of Life’s forward rule — the configurations that the dynamics can never reach, no matter how many generations you run from any starting point. They are the thermodynamic dead ends of Life’s universe.
This is a deep fact about what Life’s four rules actually do. They do not simply generate patterns. They impose an arrow of time on the grid. The past can be erased; the future cannot be undone. Gardens of Eden are not pathological cases — they are the most direct evidence that Life’s universe is, at its mathematical core, a one-way street.
The name chosen for these patterns — Garden of Eden — was more precise than its biblical origin suggests. In the story, the Garden of Eden was a beginning place, a state before history. In Life, a Garden of Eden is exactly that: a configuration that exists before the rules have had any say in the matter, a state that cannot have arisen from history. It stands at the boundary between the mathematical and the initial, the lawful and the given.