ETF Construction via Banach Fixed-Point Theory 

An exchange-traded fund assigns portfolio weights w ∈ Δ_m to m assets, where Δ_m is the probability simplex {w ≥ 0, 1ᵀw = 1}. Classical construction (Markowitz 1952, Black–Litterman 1991) optimises a quadratic objective over an estimated covariance matrix — an approach whose solutions are notoriously unstable under small perturbations.

We encode pairwise asset relationships as a weighted graph G = (V, E, w) with graph Laplacian L = D − A, where D is the diagonal degree matrix and A is the weighted adjacency matrix. The portfolio update rule is:

where μ is the expected return vector, γ > 0 is a step size, and Π₋ is the Euclidean projection onto Δ_m. T is a contraction mapping on (Δ_m, ‖·‖₂) for any γ ∈ (0, 1/λ_max(L)), with contraction factor κ = 1 − γλ₂(L).

The central quantity is the Fiedler value λ₂(L) — the second-smallest eigenvalue of L, also called the algebraic connectivity of G. It simultaneously governs how fast the iteration converges, how tight the risk bound is, and how well-connected the ETF basket is.

By the Banach fixed-point theorem, T has a unique fixed point w* ∈ Δ_m, and w⁽ⁿ⁺¹⁾ = T(w⁽ⁿ⁾) converges from any starting point with exponential rate κ.

The fixed point has a closed-form characterisation via the Moore–Penrose pseudoinverse L† of the Laplacian:

The Kirchhoff law has a direct financial reading: at the fixed point, the net correlation-weighted flow at each asset node exactly equals its negative return excess. Every asset is in equilibrium with its graph neighbours.

Using the normalised Laplacian Σ = L/λ_max as a covariance proxy (positive semidefinite, bounded spectrum), portfolio risk σ(w*) = √(w*ᵀΣw*) satisfies:

This formalises a key intuition: diversification is not merely “holding many assets.” It is increasing algebraic connectivity. A sparse asset graph (low λ₂) can carry higher risk than a small dense one, even with more assets. Validated across 180 random ETF instances with zero bound violations.

Assets whose return series share harmonic frequency relationships form subgraphs with significantly higher Fiedler values than the full asset universe. These harmonic clusters are natural ETF baskets identified by spectral alignment rather than sector labels or arbitrary index membership.

Adding a cross-cluster edge between two dense subgraphs always lowers the full-graph λ₂ (Cauchy interlacing), loosening the risk bound. This gives a precise criterion for which cross-asset linkages are beneficial: those that raise λ₂ of the resulting combined graph.

In a d-dimensional bounded phase space (d = 3: price, volume, momentum), the number of distinguishable market-state trajectories of depth n is:

For d = 3: 𝒯(n, 3) = 3 · 4ⁿ⁻¹. Each additional market cycle multiplies the state count by exactly 4. At depth n₀ = 8, there are 49,152 states — fitting in 98 MB for m = 500 assets (L3-cache-friendly). At n₀ = 10, the table covers 786,432 states in 1.57 GB.

This transforms ETF rebalancing from an iterative optimisation (re-solved at each market event) into a hash-table lookup, enabling sub-microsecond execution at scale. The geometric ratio property 𝒯(n+1, d)/𝒯(n, d) = d+1 means the state table can be extended depth-by-depth without recomputation of earlier levels.

Select the number of assets m (4–12) and graph density ρ, then press Run. The tool builds a random weighted correlation graph, computes the fixed-point portfolio w* by 3,000 Banach iterations, then animates fresh convergence from a random starting point w⁽⁰⁾ ∈ Δ_m. Node size in the graph encodes the current weight w_i⁽ⁿ⁾; dashed bars mark the analytical fixed point w*. The contraction rate κ and Fiedler estimate λ₂ are measured empirically from consecutive error ratios as the iteration unfolds.