The flowers' box-counting exponent log (u + v) / log u holds for minimum box covers of the finite graphs themselves, uniformly across every resolving scale: from single vertices, where it is the mass-scaling law log |V_g| / log diam G_g → d (HasBoxDimension.tendsto_log_card_div_log_diam), up to any scale that grows more slowly than the diameter.
A complete Lean 4 and Mathlib (v4.34.1) proof that the (u,v)-flowers of Rozenfeld, Havlin and ben-Avraham have fractal dimension log (u + v) / log u for 1 < u ≤ v: as the limit of their counting recurrences, as the log-ratio dimension of the explicit graphs, and as their box-counting dimension. There is no sorry and no custom axiom. Documentation: docs.projectnavi.ai/fd-formalization.
flowerDimension(FdFormal/FlowerDimension.lean): for the vertex countN_gand hub distanceL_g, both defined by recurrences,log N_g / log L_gtends tolog (u + v) / log u.flowerGraph_dist_hub0_hub1(FdFormal/FlowerConstruction.lean): in the constructed graphflowerGraph u v g hu huvonFin N_g, the distance betweenhub0andhub1(indices 0 and 1) isu ^ g.flowerGraph_hasLogRatioDimension(FdFormal/FlowerGraphDimension.lean): so these graphs have log-ratio dimensionlog (u + v) / log u(HasLogRatioDimension).flowerGraph_hasBoxDimension(FdFormal/FlowerBoxDimension.lean): these graphs also have box-counting dimensionlog (u + v) / log u(HasBoxDimension). Boxes are vertex sets of diameter< ℓ(Song, Havlin and Makse), the diameters diverge, and the limit oflog N_B(G_g, ℓ_g) / log (diam G_g / ℓ_g)is taken along every scale sequenceℓ_gwithdiam G_g / ℓ_g → ∞, so it does not depend on a choice of scales and is unique. This is the dimension the paper identifies with the log-ratio limit.
Not covered: u = 1, the transfractal case. There the hub distance is 1 ^ g = 1 and the diameter grows only linearly in g, so there is no length scale factor u > 1 whose logarithm could serve as the denominator.
To our knowledge, this is the first machine-checked computation, in any proof assistant, of the intrinsic box-covering dimension (network box dimension) of a recursively growing family of finite combinatorial graphs. We prove in Lean 4 that, for 1 < u ≤ v, the Rozenfeld–Havlin–ben-Avraham (u,v)-flowers have box-counting dimension log (u + v) / log u, using minimum covers by vertex sets of intrinsic diameter less than ℓ in the Song–Havlin–Makse sense. The limit holds along every positive integer scale sequence ℓ_g with diam G_g / ℓ_g → ∞, so it does not depend on a preferred sequence of recursive scales. This is a large-generation dimension for a sequence of finite graphs, taken as the generation tends to infinity; it is not the classical box dimension of a single bounded space as the scale tends to zero. Rozenfeld, Havlin and ben-Avraham obtained the same exponent from the flowers' self-similar scaling in 2007; it is also consistent with the general dimension theorem for deterministic iterated graph systems of Neroli (2024), which is stated for a Gromov–Hausdorff scaling limit rather than for minimum covers of the finite graphs. The search behind this claim, and what it found, is recorded in Prior art; corrections are welcome.
lake exe cache get
lake build --wfail
lake env lean -DwarningAsError=true FdFormal/Verify.leanFdFormal/Verify.lean prints the axioms of 78 declarations, including all four results; CI requires each to use only propext, Classical.choice and Quot.sound.
Formalization by Nelson Spence. Aristotle (Harmonic) proved leaf lemmas and simplified proofs, and proved the box-counting modules against hand-written definitions and statements; Claude assisted with Lean. For the mathematics, cite H. D. Rozenfeld, S. Havlin and D. ben-Avraham, "Fractal and transfractal recursive scale-free nets", New J. Phys. 9, 175 (2007), for box covering C. Song, S. Havlin and H. A. Makse, "Self-similarity of complex networks", Nature 433, 392 (2005), and for the dimension theory of iterated graph systems Z. Neroli, "Fractal dimensions for iterated graph systems", Proc. R. Soc. A 480, 20240406 (2024), doi:10.1098/rspa.2024.0406. Apache 2.0; see LICENSE.