L-Systems: Parallel Rewriting and the Geometry of Growth

Aristid Lindenmayer was a biologist studying filamentous algae --- chains of cells that divide and differentiate in predictable sequences --- and in 1968 he wanted a mathematical notation for describing how cell colonies develop over time. He was not thinking about computer graphics or fractals. He published the idea in the Journal of Theoretical Biology: a parallel rewriting system where a set of rules simultaneously replaces every symbol in a string at each generation, representing the simultaneous division of all cells in a growing colony.

The word “simultaneous” is what sets Lindenmayer’s system apart from Chomsky’s sequential grammars, where rules apply to one symbol at a time. Real organisms do not develop one cell at a time. Every cell is dividing, differentiating, or dying at every moment. Parallel rewriting models this correctly. The difference in behavior between sequential and parallel application is the difference between a linguistics tool and a developmental model.

Setup

An L-system has three components:

An alphabet V. A finite set of symbols. In the graphical interpretation introduced by Przemyslaw Prusinkiewicz, symbols are turtle graphics commands: F (move forward drawing a line), + (turn left by angle delta), - (turn right by delta), [ (push current position and heading onto a stack), ] (pop position and heading from the stack). The bracket pair enables branching: [ starts a branch, ] returns to the parent stem.

An axiom omega. The starting string --- the initial state of the system. Typically a single symbol such as F.

Production rules P. For each symbol in the alphabet, a rule specifying what string replaces it at the next generation. A symbol not mentioned in any rule is assumed to map to itself (identity production).

The system carries no spatial state beyond what the string encodes. There is no grid, no neighborhood topology, no boundary condition. The geometry emerges only when the string is interpreted through a rendering scheme.

The Rule

At each discrete time step, every symbol in the current string is simultaneously replaced by the string specified in its production rule. This is a single, deterministic, parallel rewriting operation.

Example: Axiom F, rule F -> F[+F]F[-F]F, angle delta = 25.7 degrees.

After one step: F[+F]F[-F]F. After two steps: every F in that string expands again. After five steps, the string has 3^5 = 243 terminal segments. Render it with turtle graphics and the result is unmistakably a plant --- a herbaceous stem with irregular branching.

Update order: Synchronous. All symbols rewrite simultaneously. This is not a computational convenience --- it is the biological point. Sequential application produces fundamentally different strings and different geometry.

Tunable parameters:

Branching angle delta: controls the geometry of the branching structure. Small changes in delta produce visually distinct plant architectures.
Production rules: the specific replacement strings determine the branching pattern, growth rate, and self-similar structure.
Number of iterations n: controls the depth and complexity of the output.

Extensions:

Stochastic L-systems: Productions are chosen with specified probabilities, introducing variation between instances generated by the same grammar.
Context-sensitive L-systems: A symbol’s production depends on its neighbors in the string, modeling cell signaling.
Parametric L-systems: Symbols carry numerical parameters (segment length, taper ratio, branch angle), enabling continuous variation.
Open L-systems: The rewriting process is coupled to environmental signals (light direction, gravity, obstacles), allowing simulated tropisms.

Emergent Behavior

Plant morphology (observed). Simple L-system rules produce specific, recognizable plant architectures --- not generic plant-like shapes. Prusinkiewicz and Lindenmayer’s 1990 book The Algorithmic Beauty of Plants demonstrates L-system derivations for dozens of named species: Stachys sylvatica, Mycelis muralis, Cornus florida, Hibiscus rosa-sinensis. A rule with three productions and one angle parameter captures branching architecture that botanical illustrators recognize.

Fractal geometry (proven). Several canonical fractals are L-systems. The Koch snowflake: F -> F+F—F+F with delta = 60 degrees. The Sierpinski triangle: A -> B+A+B, B -> A-B-A with delta = 60 degrees. The Hausdorff dimension of Koch’s curve is exactly log(4)/log(3) approximately 1.2619, provable from the self-similar structure encoded in the production rule (Mandelbrot, 1982). The fractal property follows directly from the recursive rewriting structure.

Approximate fractal dimension of real plants (observed). The box-counting dimension of tree branch structures falls reliably in the range 1.5 to 2.0, depending on species (Mandelbrot, 1982; West, Brown, and Enquist, 1997). This self-similarity is consistent with the hypothesis that the same developmental program executes at every scale of growth, which is exactly what an L-system produces.

The Mechanism

The mechanism is recursive self-similar expansion through parallel application of identical rules at every scale.

Each production rule specifies how a single element (stem segment, cell) transforms into a more complex sub-structure at the next generation. Because the same rule applies to every instance of that element simultaneously, the sub-structure at generation n contains copies of the sub-structure at generation n-1, which contains copies of generation n-2, and so on. The result is a structure that is self-similar by construction: every branch looks like a miniature version of the whole, because it was generated by the same rule.

The branching geometry of real plants has approximately the same property, because real developmental programs execute the same gene-regulatory logic at every growth point. L-systems do not approximate plant geometry from the outside --- they replicate the developmental program that produces it.

Transferable Principle

When a system grows by applying the same local transformation rule simultaneously at every active site, the resulting structure is self-similar across scales, regardless of whether the elements are cells, line segments, or architectural modules.

Formal Properties

Proven:

The class of deterministic, context-free L-systems (D0L systems) generates a proper superset of the context-free languages when interpreted as string generators (Rozenberg and Salomaa, 1980).
The word problem for D0L systems is decidable: given a D0L system and a string, it is decidable whether the string appears in the derivation sequence (Rozenberg and Salomaa, 1980).
The growth function of a D0L system --- the length of the string at generation n --- is always eventually either polynomial or exponential. The class is determined by the eigenvalues of the production matrix (Rozenberg and Salomaa, 1980).
The Hausdorff dimension of fractal L-systems is computable from the production rules and can be expressed as a ratio of logarithms of the scaling factor (Mandelbrot, 1982).

Observed / conjectured:

Parametric L-systems fitted to measured plant data (internode lengths, branching angles, phyllotactic patterns) reproduce the statistical distributions of real plant architecture to visual and quantitative accuracy (Prusinkiewicz et al., 2001).
Fibonacci phyllotaxis (the arrangement of leaves and florets in Fibonacci spirals) emerges from L-system models incorporating diffusion-based inhibitory signals at the apical meristem, consistent with the Douady-Couder physical experiments (1992). No rigorous proof connects the L-system formalism to the underlying molecular mechanism.

Cross-Domain Analogues

Procedural vegetation in film and games. Scene elements are symbols; production rules are branching specifications; the rendered tree is the geometric interpretation. Transfer is formal: the same L-system grammar that models a biological plant generates the digital asset. The analogy is exact by construction --- SpeedTree and similar tools implement L-system grammars directly. It “breaks” only in the sense that artistic tools add manual overrides that violate the grammar.

Fractal antenna design. Antenna segments are symbols; production rules specify the self-similar geometry; the resulting structure has broadband reception properties because its fractal geometry responds to multiple wavelength scales. Transfer is formal for the geometry but structural for the electromagnetic behavior --- the antenna’s frequency response is a consequence of the fractal shape, not a direct output of the L-system formalism. The analogy breaks for antenna geometries that are not self-similar.

Generative architecture. Building modules (rooms, facade elements, structural members) are symbols; production rules specify how modules compose and branch. Transfer is structural: the recursive, self-similar logic is shared, but buildings have structural, regulatory, and functional constraints that override the grammar. The analogy breaks when human design intent and building codes constrain the output to a narrow subset of what the grammar would generate.

Music composition. Notes or motifs are symbols; production rules generate melodic and rhythmic sequences with structure across multiple time scales. Transfer is structural: the recursive expansion produces long-range correlations that distinguish L-system music from random sequences. The analogy breaks because musical quality is aesthetic rather than structural --- self-similarity is necessary but not sufficient for good music.

Limits

Scope conditions. L-systems require a system that grows by repeated application of local rules to discrete elements. They are well-suited to organisms with modular growth (plants, corals, some colonial organisms) and poorly suited to systems with continuous deformation (fluid flow, elastic mechanics) or systems where the “rules” change over time (adaptive agents).

Known failures. L-systems without environmental coupling cannot model tropisms (growth toward light, gravity response) or competitive shading between branches. Pure L-systems do not model resource transport --- a branch grows according to its rule regardless of whether the root system can supply nutrients. Prusinkiewicz’s group addressed this by coupling L-systems to functional-structural plant models, but the resulting systems are no longer simple L-systems.

Common misapplications. Claiming that because an L-system can generate a structure that visually resembles a biological form, the L-system must be the mechanism by which the organism develops. Visual resemblance is not evidence of mechanistic identity. The L-system is a hypothesis about the developmental program; validation requires comparing predicted and observed growth trajectories, not just final shapes.

Connections

Methods: Procedural Generation --- L-systems are the foundational technique for rule-based procedural content generation, used wherever large volumes of structured content must be generated algorithmically.

Critiques: The Limits of Simple Models --- The strongest objection is that L-systems, as purely syntactic systems, have no semantics: they produce strings that can be interpreted as geometry, but the grammar itself contains no physics, no biology, and no mechanism. The response is that the grammar captures the logical organization of growth, which is the part that transfers across domains.

Related Models: Conway’s Game of Life --- Both are discrete systems where local rules produce complex spatial structure through iterated application; Life uses synchronous cellular automaton rules on a grid, L-systems use synchronous string rewriting. Ising Model --- Both exhibit scale-free behavior at critical parameter values; L-systems produce exact self-similarity by construction, while the Ising model produces statistical self-similarity at the critical temperature.

References

Lindenmayer, A., “Mathematical Models for Cellular Interactions in Development,” Journal of Theoretical Biology (1968). The foundational paper introducing parallel rewriting systems for biological development.
Prusinkiewicz, P. and Lindenmayer, A., The Algorithmic Beauty of Plants, Springer (1990). The definitive reference connecting L-system theory to plant morphology and computer graphics. Freely available at algorithmicbotany.org.
Rozenberg, G. and Salomaa, A., Mathematical Theory of L Systems, Academic Press (1980). The formal language theory foundations: decidability, growth functions, and the relationship to Chomsky grammars.
Mandelbrot, B. B., The Fractal Geometry of Nature, W. H. Freeman (1982). Establishes the fractal dimension framework that connects L-system output to natural geometry.
Douady, S. and Couder, Y., “Phyllotaxis as a Physical Self-Organized Growth Process,” Physical Review Letters (1992). Physical experiments demonstrating that Fibonacci phyllotaxis emerges from local inhibitory interactions at the meristem.