摘要

中文摘要

自然选择与认知系统通常被视为两个不同的领域：前者是盲目的后果驱动过程，后者是有目标的真值寻求过程。本文提出一个统一的形式框架——T-Π-O——以揭示两者共享的底层结构，并解释"盲目的进化如何产生有目标的认知"。

T-Π-O 框架由三个基本概念构成：T（真实约束/环境）、Π（投影机制，即系统将 T 转化为可观测结果的结构）、O（后果，即 Π 在 T 中运行后产生的实际结果）。三者构成一个基本循环：T → Π → O → Π'。本文给出该框架的严格公理化定义，并证明三条核心公理：（1）投影盲性——Π 的构造不依赖 T 的显式信息；（2）后果唯一判据——系统的更新仅由 O 驱动；（3）T 的隐式存在——T 只在 O 的生成中进入系统。

在此框架下，本文重新表述了进化论的两个核心数学定理。Price 方程被重写为 Π-O 空间中的 L0 层会计恒等式，并证明其右边不显含 T——它只分解"总变化 = 选择项 + 传递项"，不解释进化是否在逼近真实世界。Fisher 基本定理被重定位为 Price 方程在无变异、无环境变化条件下的特殊退化形式——它是"O 曲面上的局部爬坡"，而非"向 T 的全局收敛"。两者的共同盲区在于：它们都不涉及 Π 与 T 的真值关系。

本文进一步形式化了进化受制的三条硬边界：R（分辨率界）——O 无法区分过小的 Π 差异；D（深度界）——复制机制无法保持过深的信息；A（可达性界）——Π 空间被死亡谷分割为不连通的区域。三条边界统一为一个信息论不等式：\( \frac{dI_{\Pi}(T)}{dt} \leq \min\{\mathcal{R}, \mathcal{D}, \mathcal{A}\} \)，即 Π 向 T 累积信息的速度受最紧的那条边界限制。

在此基础上，本文给出认知涌现的精确数学判据：当且仅当 \( \left|\frac{\partial O}{\partial \hat T} \cdot \frac{\partial \hat T}{\partial \Pi}\right| > \left|\frac{\partial O}{\partial \Pi}\right| \) 时，系统从"纯自然选择"分岔出"认知系统"——其中 \(\hat T\) 是 Π 内部对 T 的估计。这意味着：认知不是对自然选择的超越，而是自然选择在 R、D、A 边界约束下发现的一种更高效的 Π 更新策略。 认知系统本质上是在 Π 内部重新运行 T-Π-O 循环，将原本只发生在种群层面的"投影→后果→筛选"过程内化，从而实现比自然选择快数百万倍的适应速度。

本文通过三个案例验证框架的解释力：视觉系统的进化（展示 R、D、A 如何塑造感官结构）、大脑与认知的进化（展示认知涌现的五个条件如何在灵长类谱系中被满足）、以及强化学习中无模型方法与基于模型方法的选择（展示 CLR > 1 判据如何统一生物进化与人工智能设计）。T-Π-O 为理解"盲目的进化如何产生能够认识世界的结构"提供了一个精确、统一、可计算的形式语言。

关键词：T-Π-O 框架；自然选择；认知涌现；Price 方程；Fisher 定理；分辨率界；深度界；可达性界；信息论；进化认识论

English Abstract

Natural selection and cognitive systems are typically treated as two distinct domains: the former a blind consequence-driven process, the latter a goal-directed truth-seeking process. This paper proposes a unified formal framework——T-Π-O——to reveal the underlying structure shared by both, and to explain how blind evolution can give rise to goal-directed cognition.

The T-Π-O framework consists of three primitive concepts: T (reality/environmental constraints), Π (projection mechanisms, i.e., the structures by which a system maps T into observable outcomes), and O (outcomes, the actual results produced when Π operates within T). Together they form a fundamental cycle: T → Π → O → Π'. We provide a rigorous axiomatization of this framework and prove three core axioms: (1) Projection Blindness——the construction of Π does not depend on explicit information from T; (2) Outcome as Sole Criterion——system updates are driven solely by O; (3) Implicit T——T enters the system only through the generation of O.

Within this framework, we reformulate two central mathematical theorems of evolutionary theory. The Price equation is rewritten as a Level-0 accounting identity in Π-O space, and we prove that its right-hand side contains no explicit T——it decomposes "total change = selection term + transmission term" but does not explain whether evolution is approaching reality. Fisher's fundamental theorem is re-positioned as a special degenerate case of the Price equation under conditions of no mutation and no environmental change——it describes "local hill-climbing on the O-surface," not "global convergence toward T." The shared blind spot of both theorems is that neither addresses the truth-relation between Π and T.

We further formalize three hard boundaries that constrain all evolutionary systems: R (Resolution)——O cannot distinguish sufficiently small differences in Π; D (Depth)——replication mechanisms cannot maintain excessively deep information; A (Accessibility)——Π-space is partitioned by "valleys of death" into disconnected regions. These three boundaries are unified into a single information-theoretic inequality: \( \frac{dI_{\Pi}(T)}{dt} \leq \min\{\mathcal{R}, \mathcal{D}, \mathcal{A}\} \), meaning that the rate at which Π accumulates information about T is limited by the tightest of the three boundaries.

On this foundation, we derive an exact mathematical criterion for the emergence of cognition: a system bifurcates from "pure natural selection" to a "cognitive system" if and only if \( \left|\frac{\partial O}{\partial \hat T} \cdot \frac{\partial \hat T}{\partial \Pi}\right| > \left|\frac{\partial O}{\partial \Pi}\right| \), where \( \hat T \) is the internal model of T within Π. This implies that cognition is not a transcendence of natural selection, but rather a more efficient Π-updating strategy discovered by natural selection under the constraints of the R, D, and A boundaries. A cognitive system is essentially a T-Π-O cycle re-run internally within Π——internalizing the "projection→consequence→selection" process that originally occurred only at the population level, thereby achieving adaptation speeds millions of times faster than natural selection.

We validate the framework through three case studies: the evolution of visual systems (demonstrating how R, D, and A shape sensory structures), the evolution of brains and cognition (showing how the five conditions for cognitive emergence were satisfied in the primate lineage), and the choice between model-free and model-based methods in reinforcement learning (demonstrating how the CLR > 1 criterion unifies biological evolution and artificial intelligence design). T-Π-O provides a precise, unified, and computable formal language for understanding how blind evolution can produce structures that can, in turn, "know" the world.

Keywords: T-Π-O Framework; Natural Selection; Cognitive Emergence; Price Equation; Fisher's Fundamental Theorem; Resolution Boundary; Depth Boundary; Accessibility Boundary; Information Theory; Evolutionary Epistemology