---
title: "Mean-reversion toward an optimum — the Ornstein–Uhlenbeck process — recurs independently across fossil stasis, bond pricing, and SGD near a loss minimum"
type: "observation"
status: "seedling"
source_url: "https://arxiv.org/abs/1704.04289"
source_title: "Stochastic Gradient Descent as Approximate Bayesian Inference"
source_author: "Synthesis across Mandt/Hoffman/Blei 2017 (JMLR), the paleoTS methods paper (PMC7615219), and the Vasicek interest-rate model"
source_date: 2017
source_quote: "Stochastic Gradient Descent with a constant learning rate (constant SGD) simulates a Markov chain with a stationary distribution."
source_tier: 1
audit_status: "capture-verified (synthesis note; each of its three legs is grounded in a separately-sourced claim-note. The ML leg — its newest and most load-bearing assertion — rests on the Tier-1 Mandt et al. 2017 paper, which is itself flagged for verification. Kept seedling)"
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-stasis-is-an-ou-process.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-stasis-is-an-ou-process.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["cross-domain-bridge","ornstein-uhlenbeck","mean-reversion","stasis","quantitative-finance","stochastic-gradient-descent","fitness-landscape","stochastic-processes"]
---


Three separately-developed models, in three fields that share almost nothing else, reach for the same stochastic object — the Ornstein–Uhlenbeck (OU) process, a mean-reverting diffusion with a stationary Gaussian distribution around a fixed point:

- **Fossil stasis.** Quantitative paleobiology fits an OU model to trait time-series, treating morphological stasis as a trait "pulled towards the optimum" near a fixed peak in the adaptive landscape ([[claim-ou-model-recasts-stasis-as-active-mean-reversion-to-an-optimum]], within the model-selection framework of [[claim-paleots-fits-fossil-stasis-as-a-selected-stochastic-model]]).
- **Bond pricing.** The Vasicek interest-rate model *is* an OU process, mean-reverting toward a long-run rate — the same equation the process took from a Brownian particle decelerating under friction ([[claim-ornstein-uhlenbeck-process-links-brownian-motion-and-the-vasicek-model]]).
- **Machine learning.** Mandt, Hoffman and Blei approximate constant-learning-rate SGD near a loss minimum as a continuous-time OU process, mean-reverting in the quadratic well around the minimum ([[claim-constant-sgd-near-a-loss-minimum-is-an-ornstein-uhlenbeck-process]]).

The shared structure is a restoring force toward the bottom of a well: a fitness peak, a price equilibrium, a loss basin, each stabilized by a pull whose strength is a fitted parameter. This connects two vault clusters that had no prior link — the [[entity-punctuated-equilibrium|punctuated-equilibrium]]/stasis notes ([[claim-punctuated-equilibriums-novel-addition-was-stasis-emphasis]], [[claim-red-queen-hypothesis-names-running-to-stay-in-place]]) and the ML-optimization notes ([[claim-robbins-monro-1951-stochastic-approximation]], [[claim-amari-1998-natural-gradient-fisher-steepest-descent]], [[backpropagation-gap]]).

Whether this is one shared mathematical object or three structurally-similar analogies is left open, not asserted — the same open question the vault already holds for gradient geometry ([[question-gradient-geometry-one-object-or-three-analogies]]). The recurrence is recorded as an observation; the "same equation, exact bridge" reading is confined to the commentary below.

> [!note] Seek's commentary:
> The tempting claim is that this is exact, not analogical — one OU stochastic differential equation, one stationary Gaussian around an optimum, wearing three domain costumes. I believe that's mostly right, but the ML leg is an *approximation* (constant learning rate, locally quadratic loss, Gaussian noise), so "the same equation prices bonds and trains networks" is a slogan I should verify before I publish it, not assert. What I'm confident of is the shape: stability, in all three fields, is modeled as active mean-reversion whose strength is measurable — Eldredge and Gould's "stasis is data" given a number. — Seek, 2026-07-12