Skip to content
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
81 changes: 81 additions & 0 deletions notes/self-similarity.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,81 @@
# Self-similarity across scales: the mathematics of repetition

Exploration, not peer reviewed, no new result claimed.

## 1. Notions of dimension

Let X be a metric space, N_epsilon(X) the minimum number of balls of radius epsilon covering X.

Box-counting (Minkowski) dimension:
D_B = lim_{epsilon -> 0} log N_epsilon(X) / log(1/epsilon), if it exists.

Hausdorff dimension: for s >= 0, H^s(X) = lim_{delta -> 0} inf{ sum_i (diam U_i)^s : X subset union U_i, diam U_i <= delta }.
dim_H X = inf{ s : H^s(X) = 0 } = sup{ s : H^s(X) = infinity }.

For self-similar sets satisfying the open set condition, dim_H = dim_B = the similarity dimension: if X = union_{i=1}^m f_i(X) with
contractions of ratios r_i, then sum_i r_i^s = 1 solves for s. Example: middle-thirds Cantor set, r = 1/3 twice, 2*(1/3)^s = 1, s = ln 2 / ln 3 = 0.6309.

Correlation dimension (Grassberger-Procaccia): for points x_i, C(r) = (2/(N(N-1))) sum_{i<j} Theta(r - |x_i - x_j|). Then
D_2 = lim_{r -> 0} log C(r) / log r.

Information dimension D_1 = lim_{epsilon->0} (sum p_i log p_i)/log epsilon. In general D_0 (box) >= D_1 >= D_2 for multifractals, with equality for monofractals.

## 2. Multifractals and the spectrum

For a measure mu on X, the local dimension alpha(x) = lim_{r->0} log mu(B_r(x))/log r. The Hausdorff dimension of the set where alpha = alpha' is f(alpha). The partition function Z(q, epsilon) = sum_i mu(B_i)^q ~ epsilon^{tau(q)}, and the Legendre transform gives f(alpha) = q alpha - tau(q), alpha = d tau / d q. Monofractal: f(alpha) = D delta(alpha - D).

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking. f is a dimension, so it cannot be a delta function. For a monofractal, τ(q) = (q − 1)D, and the spectrum collapses to the single point α = f(α) = D.


Generated by Claude Code


## 3. Cosmic structure: from fractal to homogeneous

N(<r) = 1 + xibar(r), with xi(r) = integral k^2 P(k) j0(kr) dk / (2 pi^2), xibar(r) = integral k^2 P(k) [3 j1(kr)/(kr)] dk / (2 pi^2).

Correlation dimension of the point process:
D_2(r) = 3 + d ln (1 + xibar)/d ln r = 3 + 3 (xi - xibar)/(1 + xibar), since d xibar/d ln r = 3 (xi - xibar).

P(k) = A k^{n_s} T(k)^2 D(z)^2, T(k) Eisenstein-Hu (1998) transfer function, normalised to sigma_8.

Galaxy bias in redshift space: xi_g = b^2 (1 + 2 beta/3 + beta^2/5) D(z)^2 xi_0, beta = Omega_m^{0.55}/b (Kaiser 1987). Homogeneity scale R_H where D_2 = 2.97 (within 1% of 3) or N within 1% of 1.

WiggleZ measured R_H = 71 +/- 8, 70 +/- 5, 81 +/- 5, 75 +/- 4 h^-1 Mpc at z = 0.2, 0.4, 0.6, 0.8. LCDM predicts 76-85. The fractal window closes at ~70-80 h^-1 Mpc.
Comment on lines +36 to +38

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking. β = f(z)/b with f ≈ Ω_m(z)^0.55 is Linder's (2005) growth index; Kaiser 1987 used Ω^0.6, and Ω_m needs its redshift. WiggleZ used fitted β values. "LCDM predicts 76-85" mixes two definitions. By D₂, with the fitted bias, the values are 76, 78, 81 and 78; 80–85 is the N(<r) definition. "The fractal window closes at ~70-80" should say "at a 1% threshold".


Generated by Claude Code


## 4. Critical phenomena and scaling

Ising model: H = -J sum_{<ij>} s_i s_j, s_i = +/-1. At T_c, correlation length xi diverges: xi ~ |t|^{-nu}, t = (T - T_c)/T_c, nu = 1 (2D Ising, exact). Correlation function G(r) ~ r^{-(d-2+eta)} at T_c, eta = 1/4 (2D). Hyperscaling: d nu = 2 - alpha; scaling relations: gamma = nu(2 - eta), beta = (d - 2 + eta) nu / 2, delta = (d + 2 - eta)/(d - 2 + eta).

Onsager exact: T_c = 2 J / (k_B ln(1+sqrt2)); at T_c, free energy has a logarithmic singularity (alpha = 0 with log). Energy per spin u = -coth(2K)[1 + (2/pi)(2 tanh^2 2K - 1) K_1(k)], k = 2 sinh 2K / cosh^2 2K.

Binder cumulant U_L = 1 - <m^4>/(3 <m^2>^2) -> 1 - V_4/3 = 0.61069 at T_c (V_4 = 1.1679229). Finite-size scaling: U_L(T) = f((T - T_c) L^{1/nu}), so crossing of U_{L1}, U_{L2} locates T_c with corrections L^{-omega}.

Renormalization group: under rescaling x -> x/b, couplings flow. At a fixed point, observables scale as power laws; the fractal dimension of clusters is the anomalous dimension. The critical point is where the correlation length is infinite, i.e. the system is self-similar at all scales up to the system size.

## 5. Percolation

Site percolation on square lattice: p_c = 0.59274605079210(2) (Newman-Ziff 2000). At p_c, the incipient spanning cluster has fractal dimension D = 91/48 = 1.895833. The number of sites in the largest cluster scales as L^D. The percolation (hull) exponents: tau = 187/91, sigma = 36/91, beta = 5/36, gamma = 43/18, nu = 4/3, eta = 5/24, alpha = -2/3.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking. 0.59274605079210(2) is Jacobsen 2015's value, and its last digits are disputed by Yang and Zhou 2024. Newman and Ziff 2000 give 0.59274621(13), which critical/ uses. The listed τ = 187/91 = 1 + d/D and σ = 36/91 = 1/(νD) are bulk exponents. The hull has dimension 7/4 and the accessible perimeter 4/3. Suggest "2D percolation (bulk) exponents", and add the hull and perimeter values.


Generated by Claude Code


Open (free) boundary conditions shift finite-size crossings: R_L(p) = 1/2 + a_1 (p-p_c) L^{1/nu} + b_0/L + ... with b_0 = 0.322, a_1 = 0.765 (Ziff-Newman 2002). Crossing of R_{L1}, R_{L2} sits above p_c by (b_0/a_1)(1/L_1 - 1/L_2)/(L_2^{1/nu} - L_1^{1/nu}).

## 6. Quantum paths (Abbott-Wise)

A free particle's position measurement record has fractal dimension 2 between the Compton wavelength lambda_C = h/(mc) and the de Broglie wavelength lambda_dB = h/p. Below lambda_C, relativistic smoothness; above lambda_dB, classical. The witness number W = (m Delta x^2 / hbar) / tau_witness determines which regime: W < 1 wave, W > 1 classical. Decoherence rate for gas: Lambda = (8/3) n sqrt(2 pi m kT) a^2 (kT)^2 / hbar^2 (Epstein/Schlosshauer), vs Joos-Zeh table values ~10^36 m^-2 s^-1.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking. (1) This rate has units of kg s⁻³: √(2π m kT) is a momentum, and (kT)² over ħ² leaves an extra kT. Use Λ = 8 n √(2π m_g) a² (k_BT)^{3/2} / (3ħ²), with m_g the gas molecule's mass and a the object's radius, as in CORRECTIONS 5 and quantum-paths/paths.js. Credit Schlosshauer; Epstein is the drag law. (2) Joos–Zeh's table is in cm⁻² s⁻¹. Write "10³⁶ cm⁻² s⁻¹ (10⁴⁰ m⁻² s⁻¹) for a 10 µm grain in air". (3) Use ħ/(mc) and ħ/p, not h/(mc) and h/p, as CORRECTIONS 17 and the root README do. The Compton end is quantum-paths/'s model, not Abbott and Wise's, and W decides coherence across a gap, not the path's dimension.


Generated by Claude Code


## 7. Algorithm: box counting

function boxDimension(points, epsilons):
for eps in epsilons:
grid = set()
for p in points:
grid.add(floor(p.x/eps), floor(p.y/eps), floor(p.z/eps))

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Non-blocking. grid.add(a, b, c) raises a TypeError in Python, and in JavaScript Set.add keeps only the first argument, so the routine measures the x-projection. Use a tuple or string key built from the data's dimension. Fit only over a window where the local slopes are flat, and report that window.


Generated by Claude Code

N[eps] = len(grid)
# linear regression of log N vs log(1/eps)
return slope

For a time series, delay embedding (Takens): reconstruct attractor in R^m with x(t), x(t+tau), ..., x(t+(m-1)tau); then box-count. Correlation dimension via Grassberger-Procaccia is more efficient.

## 8. Where repetition breaks

Every fractal window is bounded by a mechanism:
- Quantum paths: c (light speed) below, witnesses above.
- Cosmic structure: gravitational collapse below, Hubble expansion above.
Comment on lines +76 to +77

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking. Neither edge is right. The quantum-path window's upper edge is the particle's own momentum (ħ/p); the surroundings only set whether the roughness is drawn. The cosmic window has no "Hubble expansion" edge: R_H is where clustering falls below a chosen threshold (77 h⁻¹ Mpc at 1%, 185 at 0.1% for matter today, in #26's model and in an independent recomputation), and it moves with time and with bias.


Generated by Claude Code

- Critical point: only at T_c is xi infinite; off T_c, xi ~ |t|^{-nu} closes the window.
- Coastlines: molecular size below, continental size above.

Self-similarity is a property of the window and the mechanism that bounds it, not of the object.
Loading