The Fragmentation Hysteresis Theorem: Restoration Debt, a Latent Heat of Habitat Loss, and the Information Price of Putting a Landscape Back

Paper supported by machine-verified Lean 4. Conservation policy worldwide rests on an untested premise: that habitat destroyed can be compensated by habitat restored, at some multiplier set by administrative judgment. We show that the degrade–restore cycle is thermodynamically irreversible, and we derive the multiplier. Destruction requires no knowledge of where a landscape's connective backbone lies; restoration does, because habitat placed off the backbone is area without connectivity. Encoding this asymmetry as a siting efficiency qin(0,1] and taking the percolation giant-component fraction S(a) as the connectivity order parameter, we prove five results. (i) The restoration branch lies strictly below the degradation branch for every q<1, so the cycle encloses a hysteresis loop whose area is a dissipated connectivity work, non-negative always and zero only in the perfectly targeted (quasi-static) limit—a Second Law in landscape form. (ii) The required restoration area exceeds the lost area by exactly 1/q, so the correct biodiversity offset multiplier is m=1/q: an information premium, not a risk premium. Observed offset practice averaging 4.9:1 implies a real-world siting efficiency qapprox0.20. (iii) Because the connective backbone is observable only while a giant component survives, siting quality collapses from q_1 to q_0 at the percolation threshold, producing a latent heat of fragmentation L=(1/q_0-1/q_1)Delta A—a first-order irreversibility riding on a continuous transition. We name the threshold the fragmentation horizon. (iv) The latent heat is repayable in bits: each bit of siting information halves the restoration overshoot, up to a hard saturation depth I^ast=log_2(q_1/q_0), yielding a closed-form optimal survey depth. (v) Bits cost time, and the acquisition rate is capped by the Intelligence Bound dI/dtle PD/(k_BTln 2), giving a survey race condition and a minimum sensing power that diverges as a landscape nears its horizon. We name this operator the Mender. All results are numerically verified (71/71) against Newman–Ziff percolation with finite-size scaling. Three falsifiable predictions follow, one of which is testable today on existing offset registries.