Emergence in Complex Systems

Or, Rare Stability in the Chaos

Aykut C. Satici

Rare Stability in the Chaos

Only five positions in the Sun–Earth system allow a stable equilibrium relative to both bodies: the Lagrange points.

This is a rare stable solution to the restricted three-body problem.

Is the generalized three-body problem solvable analytically? No, and that gap is the subject of this lecture.

The five Lagrange points of a two-body system (e.g., Sun–Earth).

• Open with the surprising fact: continuum of positions, only five (L1–L5) give a stable/quasi-stable equilibrium for a negligible-mass third body
• “Restricted” three-body problem (one mass negligible) → rare closed-form family of solutions
• The hook: general three-body problem (comparable masses, mutual interaction) has no closed-form solution at all — Poincaré, 1890s, one of the first rigorous demonstrations that deterministic systems can be unpredictable in closed form despite well-understood underlying laws
• → Thread for the whole lecture: analytic predictability is the exception, not the rule; stability, when it occurs, is rare and precarious

Newton and Complex Systems

Sir Isaac Newton (1642–1727).

Newton’s laws describe planetary motion and fluid dynamics, and, through sensitivity to initial conditions, the seeds of chaos theory.

The three-body problem is a direct consequence of Newtonian mechanics applied to more than two interacting bodies.

Newton’s laws are exact. Their consequences, for three or more interacting bodies, are not analytically solvable.

• Newton’s laws = founding language of complex systems, even though Newton wasn’t thinking in those terms
• Planetary motion: multiple gravitating bodies = already a complex system; 3-body problem is the direct consequence of F=ma + gravity applied to 3+ bodies
• Same equations underlie fluid dynamics (weather, ocean currents) and chaos theory (sensitivity to initial conditions is a consequence of nonlinear Newtonian dynamics, not a departure from it)
• Key point to land: laws are exact per pairwise interaction → aggregate behavior of 3+ bodies is not analytically solvable. Exact micro-laws ≠ tractable macro-behavior
• → First hint of a theme the whole course returns to

Laplace’s Demon

“We may regard the present state of the universe as the effect of its past and the cause of its future. An intellect which… knew all forces… and all positions of all items… would embrace in a single formula the movements of the greatest bodies of the universe and those of the tiniest atom… for such an intellect nothing would be uncertain and the future, just like the past, would be present before its eyes.”

— Pierre-Simon Laplace, A Philosophical Essay on Probabilities

Pierre-Simon Laplace (1749–1827).

• Classical, pre-complexity worldview at its most extreme — read the quote aloud, slowly
• “Laplace’s Demon” = pure determinism: exact position/momentum of every particle + all forces → entire future (and past) perfectly computable, zero uncertainty
• Laplace earned this claim: solar-system stability (not just 3 bodies), pioneered probability theory / Bayesian updating, analytical mechanics, tides from Earth-Moon-Sun interaction, perturbation theory for planetary deviations — genuinely sophisticated many-body complex-systems work
• → His demon is the target the rest of the lecture dismantles — not because determinism is wrong, but because even perfectly deterministic systems can be unpredictable in closed form, in practice and in principle

Analytic vs. Algorithmic

Definition

Physics is generally analytic; complex systems are generally algorithmic. Both can produce quantitative predictions that are experimentally testable.

• Complex systems exhibit a rich phase structure and huge variety of macrostates that often cannot be inferred from the elements’ properties alone: this is emergence
• The theory of complex systems is the quantitative, predictive, and experimentally testable science of generalized matter interacting through generalized interactions

Reference

Thurner, Hanel, & Klimek (2018), Introduction to the Theory of Complex Systems.

• Methodological pivot of the whole lecture — be precise about vocabulary
• Analytic: write equations, solve in closed form, exact prediction. Newton-through-Laplace paradigm; works for two-body gravity, simple circuits
• Algorithmic: no closed-form solution, but specify local rules precisely, simulate forward, reliably/reproducibly/testably expect macro patterns to emerge
• Both are genuinely scientific — simulated CA/agent-based/genetic-algorithm predictions are just as quantitative/falsifiable as solving a differential equation, just derived differently
• → Formal definition (Thurner, Hanel, Klimek) worth students copying verbatim: quantitative, predictive, testable science of generalized matter interacting through generalized interactions — abstract enough to cover particles, neurons, bees, or traders

Algorithmic Thinking

• Cellular automata: Conway’s Game of Life. Simple grid rules produce unpredictable, complex patterns
• Agent-based modeling: simple agent rules lead to emergent traffic, market, or ecosystem dynamics
• Genetic algorithms: evolution-inspired search over vast, hard-to-navigate solution spaces

• Network algorithms: e.g. PageRank, navigating and ranking complex networks
• Self-organizing systems: local rules produce global order with no external guidance (animal coat patterns, galaxy structure)
• Chaos & fractals: small changes in initial conditions produce vastly different, self-similar outcomes

• Walk through as a toolkit, not a checklist — “algorithmic” is a family united by one philosophy: simple local rules, run forward, observe what emerges
• Cellular automata (Conway’s Game of Life): gliders, oscillators, self-replicating structures emerge from simple neighbor-counting rules
• Agent-based modeling: heterogeneous agents, individual rules → traffic, markets, ecosystems (economics/sociology/ecology)
• Genetic algorithms: evolution’s variation-and-selection loop as general-purpose optimization for vast search spaces
• Network algorithms (PageRank): ranking/navigating a network is itself algorithmic, not analytic
• Self-organizing systems: animal coat patterns, galactic structure, social organization — global order, no central planner
• Chaos & fractals: ties back to Lagrange points from the opening — sensitivity to initial conditions + self-similar structure = same algorithmic, non-analytic regime

Machine Learning and Emergence

Predictive algorithms: deep learning models complex systems by learning patterns directly from data (climate, markets, language).

Emergence in simulation: each bird follows three local, distance-based rules. No bird perceives “the flock.”

Separation: avoid crowding nearby neighbors.

Alignment: match nearby neighbors’ heading.

Cohesion: steer toward neighbors’ center.

Run all three on every bird at once, and flock-level movement emerges, unwritten in any single bird’s rules.

• Two closing examples for the algorithmic-thinking arc
• Machine learning: “algorithmic, not analytic” taken to its modern extreme — deep learning learns statistical patterns directly from data (markets, climate), no interpretable closed-form law produced
• Flocking birds (more conceptually important) — cleanest demonstration of true emergence. Craig Reynolds’s 1986 “boids” rules, still the standard teaching example
• Each bird reacts only within its local perception radius (gray disc) — a distance-based rule, neighbors outside are invisible to it
  • Separation: steer away from neighbors too close (avoid collisions)
  • Alignment: match average heading of nearby neighbors
  • Cohesion: steer toward average position of nearby neighbors
• No rule mentions “the flock” — no bird perceives the flock, only a handful of neighbors
• → Run all three at once on every bird: coordinated flock movement appears, unwritten in any individual bird’s behavior. Sharpest way to state emergence: the property exists at a level of description the components don’t have access to. Good answer when a student asks if emergence is just a vague word for “complicated”

The Real Thing

A murmuration of starlings. No bird is in charge. Credit: Walter Baxter, CC BY-SA 2.0.

• Pause on this photo before moving on
• Every coherent shape here = the three boids rules running independently in thousands of birds at once, no leader, no bird sees the shape it’s part of
• Each bird only reacts to a handful of neighbors within its own perception radius
• Wave-like ripples, sudden turns, the cloud “breathing” — none of it planned or centrally coordinated
• → Same phenomenon as the braid-pattern animation (lecture 1) and the sand-pile avalanches coming up: global structure, no blueprint, purely local distance-based interaction

Biology: Never at Equilibrium

Living matter constantly uses energy and performs work, driven by energy gradients, always out-of-equilibrium.

Evolution increases and destroys diversity: both a creative and a destructive process.

The Darwinian scenario alone fails to explain boom-and-bust phases, where diversity changes radically over short periods.

The fossil record: a story of both radiation and mass extinction.

• Second of the three classical disciplines, sharper framing than last lecture
• Living matter permanently out of thermodynamic equilibrium — survives by continuously importing energy and doing work; equilibrium = death
• Out-of-equilibrium requirement + self-replication + localized “Carnot-cycle”-like energy processing = deep structural feature, not incidental
• Evolution: more nuanced than “survival of the fittest” — both creative (novel diversity via variation) and destructive (extinguishing lineages)
• Simple Darwinian gradualism doesn’t explain the fossil record: boom phases (rapid diversification) + crash phases (mass extinction), visible in the timeline
• If a student pushes back (“isn’t that just selection responding to an external shock, e.g. continents drifting?”): that’s the classical-physics move — treat environment as a fixed backdrop, occasionally perturbed. Two reasons that’s not sufficient (Thurner et al. §1.3.3, p.13):
  • Selection alone is gradualist by construction (variation + differential reproduction → smooth incremental change); it has no built-in mechanism for why diversification/extinction events follow a heavy-tailed, avalanche-like statistic (many small, occasional huge) rather than a steady rate
  • No external trigger is even required in principle: species reshape each other’s fitness landscapes as they evolve (new predator/symbiosis/niche), so boom-bust can emerge endogenously from co-evolution — sets up next slide’s point that the environment itself isn’t fixed
• → “Boom and bust” irregularity is itself a complex-systems signature — looks like punctuated, threshold-driven dynamics (self-organized criticality, coming shortly), not smooth optimization

Evolutionary Dynamics vs. Physics

Evolutionary dynamics differs from physics in two fundamental ways:

1. Boundary conditions can’t be fixed. A wolf pack’s fate depends on climate, disease, and human activity, none of which are stable “boundary conditions” you can hold constant.
2. The phase space itself isn’t fixed. New elements (traits, species, technologies) emerge and change the space of future possibilities for everyone else in the system.

• Sharp contrast: classical-physics worldview (first half of lecture) vs. biological/evolutionary worldview
• Physics (even the three-body problem): fixed phase space + fixed boundary conditions, ask how system evolves within it. Evolutionary systems break both
• 1. Boundary conditions can’t be fixed — wolf pack example: even with complete genetics/environment/resource knowledge, can’t predict future trajectory because climate, disease, human encroachment are themselves unpredictable, external, not fixed parameters
• 2. The phase space itself isn’t fixed — bird trait space (beak size, wing shape, coloration): food source shifts → fitness landscape reshapes; new traits/species open entirely new dimensions
• Callback to previous slide’s boom/bust question: this is why Darwinian selection alone doesn’t explain radiation/extinction statistics — species co-evolve each other’s fitness landscapes, so the “environment” a Darwinian model treats as fixed is itself part of the dynamics
• → Fundamentally different from a fixed configuration space with dynamics on top — the “board” itself is redrawn as the game is played

The Adjacent Possible

The set of all states the world could reach in the next step, given its present state.

Introduces path-dependence into the stochastic dynamics of phase space: what’s reachable next depends entirely on where you are now.

A game of Go mid-play: the legal next moves, the “adjacent possible,” depend entirely on the current board.

Example: Building a House

A single-level base’s adjacent possible: add a second level, add windows, add a roof. Once you have two levels and windows, the adjacent possible shifts to include a chimney, a balcony, a garage.

• Stuart Kauffman’s “adjacent possible” — makes the previous slide’s “changing phase space” idea precise and useful
• Adjacent possible = the (typically much smaller) set of states reachable in the very next step from wherever the system is now — not all possible states ever, just next-door neighbors
• Builds path-dependence in: identical eventual reachable spaces, but different near-term trajectories, because current adjacent possibles differ
• Go board: legal next moves = small, well-defined set from the current position, not the space of all Go positions
• Construction example: single-level base → adjacent possible = second level, windows, roof. Once you have those → shifts to chimney, balcony, garage — options not reachable (or meaningful) before
• → Becomes important later: how architectural decisions constrain what future decisions are even available

Robust and Adaptive: Edge of Chaos

Definitions

Robustness = trajectories in a λ < 0 regime, where disturbances decay and function is preserved: a genuine stability margin.
Adaptability = a large-enough disturbance triggers a bounded, transient excursion toward λ ≥ 0, sampling nearby configurations before the system re-settles into a stable regime.

Thurner et al.’s resolution: a system that operates near this boundary gets both, from two different phases of its own trajectory.

The Mechanism

Every dynamical system has a Lyapunov exponent λ, measuring how fast nearby trajectories diverge. Time spent with λ < 0 gives robustness: perturbations shrink back out. A disturbance strong enough to push the system toward λ ≥ 0 gives adaptability: a bounded excursion through nearby configurations, followed by re-settling — cheap to trigger, but recoverable, much like a trust-region exploration step in reinforcement learning.

• The “apparent paradox” structuring the rest of the lecture: robustness (resist change) vs. adaptability (embrace change) sound like opposites
• Resolution (Thurner, Hanel & Klimek §1.3.4, framed here in control/RL terms): robustness and adaptability come from two different phases of the same system’s trajectory, not from a single fixed operating point
• Every dynamical system has a max Lyapunov exponent λ: |δX(t)| ∼ e^(λt)|δX(0)|
  • λ < 0 → trajectories converge to an attractor — this is where robustness lives, a real stability margin in the control-theory sense
  • λ > 0 → trajectories diverge, chaotic — a system passing through this regime is exploring: sampling alternative nearby configurations
• Adaptability = a disturbance large enough to carry the system through a λ ≥ 0 excursion, then back to a λ < 0 regime (possibly a different attractor than before) — bounded, transient exploration, analogous to a trust-region or bounded-perturbation exploration step in RL: enough movement to discover something new, not so much that the system fails to re-converge
• → Operating near the λ = 0 boundary, rather than deep in one regime, is what keeps this kind of bounded excursion cheap to trigger
• Self-organized criticality (a few slides ahead) is one mechanism that positions systems close to this boundary with no external tuning — most triggered excursions stay small, only rarely large, which is what keeps the exploration bounded in practice

Case Study: Vehicle CO₂ Emissions

A single anthropogenic input, vehicle CO₂, propagates through a coupled human–natural system at three levels:

1. Component (plant physiology): altered photosynthesis, stomatal regulation, short-term adaptive plasticity
2. Ecosystem (forest): diversity provides redundancy; species respond unevenly; succession shifts over time
3. Earth system (climate–carbon coupling): feedback loops (temperature ↔︎ CO₂ ↔︎ vegetation ↔︎ albedo), nonlinear interactions, possible tipping points

Local adaptation may coexist with global instability. Robustness at one scale does not guarantee stability at another.

• Separate, complementary point from edge-of-chaos — a systems-engineering observation, not from Thurner’s book: robustness at one level tells you nothing about stability at another. Three nested levels — take it slowly
• Component (plants): elevated CO₂ changes photosynthesis/stomatal regulation — genuine adaptive plasticity, purely local
• Ecosystem (forest): species diversity gives redundancy, but species respond unevenly — composition/succession shift over time (structural adaptation + functional robustness mixed, not “no change”)
• Earth system: temperature–CO₂–vegetation–albedo feedback loops dominate, strongly nonlinear, vulnerable to tipping points
• → Not “plants adapt, problem solved” — local adaptation can coexist with global instability. Robustness at one scale tells you nothing about stability at another. Previews why systems thinking insists on checking multiple levels, not declaring victory after one

Self-Organized Criticality

Dynamical, out-of-equilibrium systems that evolve toward a critical point as an attractor, with no external tuning required.

Characterized by (approximate) scale invariance: no characteristic scale, often showing up as power laws in the associated probability distributions.

Laboratory footage: grains poured onto a real sand-pile, forming and repeatedly shedding its critical slope. Credit: Sergei Klishin.

The system is simultaneously robust and adaptive: it returns to its critical angle of repose after every avalanche, by reorganizing.

• Self-organized criticality (SOC) — the richest idea in the lecture, spend real time here
• Defining feature: system drives itself to a critical state, purely internal dynamics, no external tuning
• At the critical state: scale invariance — no characteristic event size, small/medium/enormous avalanches all occur following a power law, not clustering around a bell-curve “average”
• Sand-pile (canonical demo, pictured): pour grains slowly — early accumulation, then slope hits the critical angle of repose, avalanches begin (some tiny, some pile-reshaping)
• No one tunes the slope — continuous slow driving + local instability rules is enough to find and maintain criticality
• → “Robust and adaptive” from two slides ago, made concrete: robust (returns to the same angle of repose every time) and adaptive (reorganizes its exact shape every time) — robustness in the aggregate statistic, adaptability in the microscopic detail, same system at once

Slow Driving, Sudden Release

Sand-pile: grains added slowly and continuously; each grain adds stress to the system.

Amazon delivery network: customer orders arrive continuously (the slow driving signal); each order adds workload or stress to the network.

Both systems are continuously driven by external input they do not themselves control, and both can respond with avalanches of any size.

• Deliberate analogy-building: map the sand-pile mechanism onto a system students interact with, to show SOC isn’t just a physics-lab curiosity
• Sand-pile: grains arrive slowly/continuously, add local stress, threshold crossed somewhere → avalanche redistributes stress, unpredictable size
• Amazon delivery network: orders play the same role as grains — continuous slow driving input, each adds workload somewhere in fulfillment/logistics, network absorbs gradually
• Occasionally a capacity threshold crosses (hub, shipping lane, warehouse) → disruption cascades, unpredictable “avalanche” size
• → Not just a loose analogy — same underlying math (SOC) on a socio-technical system instead of a physical granular one. Why real supply chains suffer disruptions wildly disproportionate to their trigger

Three Properties, One System

• Emergence of criticality: a characteristic slope forms naturally as grains accumulate. The system progresses toward avalanches without being told to
• Structural robustness: the pile maintains integrity despite local disturbances, reorganizing to preserve overall slope and shape
• Adaptive reshaping: the pile continuously reshapes as new grains arrive. Stable global behavior emerges from constant local change

• Synthesis checkpoint before social science — ask students to state all three properties back before continuing
• Criticality = the process (pile organizes toward critical angle, no external tuning)
• Robustness = the aggregate outcome (statistical behavior/characteristic slope preserved, whatever local avalanche damage occurs)
• Adaptive reshaping = the mechanism underneath both (grain-by-grain shape constantly changing — that local reshaping is what produces stable global behavior)
• → General lesson, well beyond sand: don’t confuse “stable” with “static.” Global order and local churn aren’t in tension — local churn is often how global order is maintained

Social Science: Qualitative

The science of social interactions and their implications for society. Usually neither quantitative nor predictive in the way physics is; largely qualitative and descriptive.

Social processes are hard to model mathematically: they’re evolutionary, path-dependent, out-of-equilibrium, context-dependent, high-dimensional, and multi-level.

Early social science built its methods around exactly these difficulties.

• More honest accounting than lecture 1 of social science’s methodological challenges
• Social science doesn’t produce sharp, falsifiable Newton/Laplace-style predictions — not a failure, a structural consequence of what it studies
• Social processes are: evolutionary (path-dependent, no clean reset), out-of-equilibrium (always in flux), context-dependent (same intervention, different outcomes), high-dimensional, multi-level (individual/group/institutional/societal)
• → These features make Laplace’s-Demon-style closed-form prediction essentially impossible — why social science leans qualitative/descriptive
• Sets up next slide’s payoff: complexity science’s algorithmic/network toolkit gives a new way to make quantitative progress anyway

Social Systems as Networks

Social systems can be modeled as time-varying, multilayer (multiplex) networks:

• Nodes: individuals or institutions
• Links: interactions of different types
• Interactions change over time

A multilayer network: the same nodes, connected differently across layers (relation types) and time.

• The payoff promised last slide: complexity science’s network toolkit (from lecture 1) is a workable formalism for social systems too
• Model as a time-varying multilayer (“multiplex”) network: nodes = individuals/institutions; multiple simultaneous link layers (friendship, professional, family, financial), same nodes
• Nothing static — layers and time both vary, so who’s connected to whom, in what way, is itself dynamic (mirrors the evolving phase space from evolutionary biology earlier)
• → Genuinely useful research tool: e.g., does professional-layer influence predict friendship-layer influence six months later — quantitative, testable, made tractable by giving social science a representational vocabulary (networks) instead of demanding physics-style closed-form equations

Emergence, Revisited

Definition

Emergence: the whole exhibits behaviors and properties not predictable from the sum of its parts.

• Traditional analytical methods (physics, biology, social science alone) are insufficient to predict emergent behavior
• Emergent patterns can be expected through an algorithmic approach
• Complex systems science is a developing field, working to make emergence quantifiable, predictable, and experimentally testable

Examples across domains: Conway’s Game of Life, flocking birds, ant colony optimization, ecosystems, financial market bubbles and crashes, human cognition, social-network misinformation spread, urban development

• Close the loop on the lecture’s central concept — restate with everything students now have
• Emergence: system-level behavior that can’t be predicted from the parts, even with complete part-knowledge — exists only at the collective level (flocking-birds example made this vivid)
• No single discipline alone predicts it; algorithmic methods (simulate local rules, observe what emerges) can expect it without deriving it analytically
• “Spot the pattern” example list — read through quickly:
  • Game of Life, flocking birds — direct callbacks
  • Ant colony optimization — pheromone trails → near-optimal foraging, no ant sees the whole picture, inspired real routing/scheduling algorithms
  • Ecosystem biodiversity/resilience — from species interactions, not any one species
  • Financial bubbles/crashes — aggregate of simple trading rules
  • Human cognition/consciousness — billions of neurons, simple firing rules
  • Misinformation spread — network structure, not one person’s intent
  • Urban development — countless decisions, no central planner
• → Complex systems science is still actively working to make all of this quantifiable/testable — an open research frontier, not settled science. Tell students this explicitly

In-Class Activity: The ED as a Complex System

A hospital emergency department, working in groups, as a case study:

• Physics → nonlinear utilization, the capacity cliff, synchronization and blocking
• Biology → homeostasis, acute stress response, allostatic load
• Social science → local rationality vs. global externalities, correlated decisions, phase transitions
• This lecture → reinforcing/balancing feedback loops, self-organized criticality (boarding cascades), edge of chaos (robust and adaptive), networks (whole-hospital coupling)

Five discussion questions, one per lens plus resilience. Work the case, then we regroup and connect each finding back to a concept from today.

• Transition into the case-study handout (hospital_ed_complex_systems_case_study.pdf/tex, in case_studies/hospital_ed) — frame as the activity that operationalizes everything just covered, not a disconnected add-on
• Handout’s “Complex Systems Lens” maps onto this lecture’s concepts:
  • Utilization/capacity-cliff → edge of chaos, criticality, in a concrete operational setting
  • Homeostasis/allostasis → biology-never-at-equilibrium
  • Local rationality → global externalities → boids/flocking emergence (individually reasonable clinician decisions, no one sees “the system”)
  • Reinforcing feedback loops (boarding, workload-error-rework) → same slow-driving/sudden-release dynamic as the sand-pile and Amazon analogy
  • Final resilience-architecture discussion → robustness-and-adaptability, now as a design goal rather than observed phenomenon
• Run: pairs, 15-20 min on the five discussion questions, then regroup — each pair reports back mapping their answer to a specific lecture concept
• → Goal: students leave seeing complex-systems theory as a diagnostic lens on a real, unglamorous operational system, not a collection of physics/biology anecdotes

References

Reference

Thurner, S., Klimek, P., & Hanel, R. (2018). Introduction to the Theory of Complex Systems. Oxford University Press.

• Same reference that anchored lecture 1 — both lectures draw from one coherent framework, not a loose collection of anecdotes
• Everything covered (analytic/algorithmic, SOC, adjacent possible, evolving networks, emergence) traces back to this text’s unified treatment: the quantitative, predictive, testable science of generalized matter interacting through generalized interactions
• → From here: course moves into Part I, System Thinking — formal definition of architecture, the four-task discipline (form, function, relationships, emergent properties) that structures the rest of the semester