Phase Transitions Criticality

The unifying principle

Take many units (spins, neurons, traders, cells) that copy their neighbors with strength $K$ (coupling / gain / connection probability). Each unit balances a local aligning force against noise. Define an order parameter $m$ (net alignment, fraction connected, fraction active). A crude mean-field self-consistency gives

$m = \tanh(K m + h)$

• For $K < K_c$: only solution is $m \approx 0$ (disordered, local noise wins).
• For $K > K_c$: nonzero $m$ appears (ordered, global structure).
• At $K = K_c$: the two regimes meet — a continuous transition.

Expanding for small $m$ (Landau form):

$F(m) \sim a(K_c-K)\,m^2 + b\,m^4$

the quadratic coefficient flips sign at $K_c$, so the minimum slides continuously off zero. Key universal signatures near $K_c$:

• Diverging correlation length: $\xi \sim |K-K_c|^{-\nu}$ — everything talks to everything.
• Power-law observables (no scale): cluster sizes $P(s)\sim s^{-\tau}$.
• Diverging susceptibility / critical slowing down: response to a tiny perturbation blows up, recovery time $\to \infty$.

The magic: exponents $\nu,\tau$ are shared across wildly different systems (universality) because they depend only on symmetry and dimension, not microscopic detail.

How it shows up in each field

Physics — the ferromagnet (Ising model)

The literal origin. Spins $s_i=\pm1$ with energy $H = -J\sum_{\langle ij\rangle} s_i s_j - h\sum_i s_i$. Tune temperature $T$; the knob is $K = J/k_B T$.

• Above the Curie temperature $T_c$: spins random, magnetization $m=0$.
• Below $T_c$: spontaneous alignment, $m \neq 0$, with $m \sim (T_c-T)^{\beta}$, $\beta\approx 0.33$ in 3D.
• At $T_c$: domains of all sizes, susceptibility $\chi \sim |T-T_c|^{-\gamma}$ diverges.

AI-ML — percolation & the "edge of chaos"

Connectivity/gain plays the role of $K$.

• Random networks: as edge probability $p$ crosses $p_c \approx 1/N$, a giant connected component snaps into existence (percolation) — the graph goes from fragments to globally connected.
• Deep nets: signal propagation depends on weight variance $\sigma_w^2$. There's a critical $\sigma_w^2$ separating vanishing gradients (ordered) from exploding gradients (chaotic); trainable deep nets live near this edge of chaos, where the Jacobian's mean squared singular value $\approx 1$.
• Emergence in LLMs: capabilities appear suddenly as scale (params/data/compute) crosses thresholds — a transition-like jump in a capability order parameter.

Stock-Market — crashes & self-organized criticality

Coupling = herding / leverage / correlated positioning.

• Traders imitate neighbors; when effective coupling $K$ rises (leverage, common risk models), the market approaches instability. Below threshold, shocks stay local; above it, a single sell triggers a global cascade (crash).
• Empirically markets sit near criticality: return distributions have power-law tails $P(|r|>x)\sim x^{-\alpha}$ ($\alpha\approx 3$), and volatility clusters — hallmarks of scale-free avalanches (cf. sandpile self-organized criticality).
• Some crashes show log-periodic precursors, the signature of a system tuning toward a critical point (Sornette).

Biology — neural avalanches & criticality of the brain

Coupling = synaptic gain; order parameter = fraction of firing neurons.

• Cortical activity propagates as neuronal avalanches whose sizes follow $P(s)\sim s^{-3/2}$ — the exact exponent of a critical branching process (branching ratio $\sigma=1$: each spike triggers on average one more).
• $\sigma<1$: activity dies (subcritical); $\sigma>1$: seizure-like runaway (supercritical); $\sigma\approx1$: maximal dynamic range and information transmission.
• Also literal thresholds: gene-regulatory switches, epidemic spread ($R_0=1$), flocking onset.

Why this bridge matters

• Universality lets intuition transfer. Because exponents depend only on symmetry/dimension, an intuition from magnets ($\xi$ diverges → global sensitivity) directly predicts brain dynamic range, network trainability, and market fragility.
• The branching ratio $\sigma\approx1$ is the same statement as $K\approx K_c$ and $R_0\approx1$: systems that compute or transmit best sit right at the edge — powerful design principle for both brains and neural nets.
• Early-warning signals are shared. Critical slowing down — rising autocorrelation and variance as recovery time diverges — is used to forecast ecosystem collapse, epileptic seizures, and possibly market crashes. One diagnostic, four fields.
• Danger is shared too. Living near criticality buys sensitivity but courts catastrophe: the same knob that makes a network expressive makes a market crash-prone and a brain seizure-prone.

Connections

• 01-Ising-Model-and-Curie-Temperature
• 02-Landau-Theory-and-Order-Parameters
• 03-Percolation-and-Giant-Components
• 04-Edge-of-Chaos-in-Deep-Networks
• 05-Emergent-Abilities-and-Scaling-Laws
• 06-Self-Organized-Criticality-Sandpiles
• 07-Log-Periodic-Crash-Precursors
• 08-Neuronal-Avalanches-and-Branching-Processes
• 09-Critical-Slowing-Down-Early-Warnings
• 10-Universality-and-Critical-Exponents

#bridge