Source
Xiangyi Meng, Benjamin Piazza, Csaba Both, Baruch Barzel and Albert-László Barabási (2026), “Surface optimization governs the local design of physical networks”, Nature 649, 315–322, LIT-069, read in full with the Supplementary Information in NOTE-075.
What was actually shown
The departure from Steiner (Fig. 1). Length-minimal (Steiner) trees have only degree-3 nodes, planar bifurcations and 120° angles. In six classes of 3D networks, k = 4 nodes are 17.0% (human neurons), 6.5% (fly), 8.3% (one arterial tree), 17.7% (tropical trees), 8.7% (corals) and 4.6% (Arabidopsis); P(θ) has no peak at 2π/3; bifurcations are close to, but not exactly, planar.
The model (Eqs. 1–2). Each link is a smooth sleeve; total surface area is minimised subject to every non-contractible loop around a link having length at least w. Length minimisation has no scale; this has two, χ = w/r and ρ = w′/w.
Trifurcations (Fig. 3g; SI Fig. 13). For a regular tetrahedron of terminals the intermediate link length λ = l/w drops sharply to 0 at χ ≈ 0.83: two close junctions cost more area than one merged junction. Empirical P(λ → 0) is positive in all six datasets.
Sprouts (Fig. 4g; SI Fig. 19). With two equal branches and a thinner third, the thin branch leaves perpendicular to a straight thick path (Ω₁→₂ = 0) for ρ below about 0.6, and the angle rises linearly above. A plain-cylinder model (Zamir) instead approaches 90° only as ρ → 0. Breakpoints fitted per dataset lie at 0.52–0.92, and sprouts are 11–53% of equal-branch bifurcations. In human neurons 3,911 of 4,003 sprouts (97.7%) end on a synapse (the caption says 92%).
What could have come out otherwise. The data could have looked like Steiner trees: degree 3, 120°, planar. They do not, in all six classes. The sharp thresholds could have been absent in the simulations; for the two symmetric motifs they are present.
What this does not say
- That real networks are surface-optimal. Agreement is qualitative. Real double-bifurcation motifs are 1.20–1.35 times the Steiner length, which the model, predicting straight cylinders away from junctions, does not explain. “A stereotypical principle that is not overwritten by functional or global need” (p. 320) is an assertion.
- That the theory was solved exactly. Strebel’s exact minimal-area solution ignores immersibility in three dimensions and has conical singularities. The numbers come from a locally minimising, penalised quad-mesh optimisation, and SI §5 concedes “a globally distinct surface configuration may still achieve a lower total area”.
- That trifurcations are measured independently of the analysis. Two bifurcations with intermediate link at most w/2π are merged into one trifurcation (SI Eq. 43), and the one skeletonisation check changes P(λ → 0) from 0.276 to 1.200. The “roughly 15%” k = 4 figure is pooled and 96% human-neuron nodes.
- That the random-terminal curve came from the minimiser. Running it on 10⁶ quadruplets was “un-scalable”; λ was mapped through a sigmoid calibrated on the tetrahedron, and the empirical Steiner reference is the distribution of external legs, whose relation to internal ones has Spearman ρ of only 0.11–0.42.
- That the sprouting transition is sharp in the data. Breakpoints are chosen to maximise a slope jump, with no null; the tree dataset’s jump is 0.002.
- That surface is the right cost. For neurons, corals and trees the authors say volume must also count; the volume check is one motif.
- Anything about looped networks. Loops are excluded by construction.
Connections
- Graph curvature (THEORY-456, from LIT-004; THEORY-502, from LIT-020). Those curvatures live on an abstract graph’s edges and are blind to the thickness ratios χ and ρ that drive this account. NOTE-075 draws the contrast.
- What it leaves open. NOTE-075 suggests the natural question this seeds: at what scale material-cost optimisation stops predicting morphology, local junction geometry against the 25% excess in global layout.