Feng Xiaoping
中 EN
Book a conversation
01Home 02About 03Research 04Papers 05Writing 06Contact
Book a conversation

Write: geo@qiyanword.cn
阅读中文

Theory ~8 min

The Edge: A Gap Left by Partial Observability

Turning “edge and generation” into a problem that already has a name: the state is hidden, so you maintain a belief; when the belief moves, the policy is reset.

The previous notes all argued one thing: prediction is failing. This one turns the question around — not why it fails, but what to do after it does, written in a language another person can argue with.

I did not invent that language. It already has a name: partially observable stochastic optimal control (POMDP). My contribution is smaller: pointing "edge" and "generation" at its parts, one at a time.

1. The gap: you never see the state, only a shadow

Classical decision models assume you know what state you are in. In the world this almost never holds. What you observe — prices, orders, filings, headlines — is not the state. It is a projection of the state through noise, delay and selective disclosure.

Once you accept that, the problem changes shape. Not "what will happen next", but "what should I believe, given that I only see a shadow". The first asks for a forecast; the second asks for a belief.

The edge sits where the shadow is thinnest and the consequences are heaviest: low signal-to-noise, structure not yet settled, a region where mainstream models give confident but unreliable answers.

2. The move: from forecasting to believing

Folding observations into a distribution π = p(state | history) has a standard name: Bayesian filtering. The Kalman filter is the linear-Gaussian case; particle filters approximate the high-dimensional nonlinear case.

The point is not the algorithm but the act: every observation rewrites what you believe. Not "collect more information", but "admit that information changes the structure of judgment".

That is also where generation gets its place. Generation means the policy is reset after the belief updates — not forecasting the future, but reorganizing action now.

Edge–Generation Theory: Technical Architecture From philosophical premises to partially observable stochastic optimal control — every concept is anchored to a computable object v1 · 2026-11 Related fields Stochastic control Information theory Complex systems Strategic management Philosophical basis Strategic philosophy Value created under uncertainty Treat volatility as raw material, not error Peripheral innovation Value sits where attention is thin Weakest consensus, largest option The edge is not the laggard OODA loop Observe–orient–decide–act Faster cycles win Write feedback into process Mathematical core System model State process dx = F(x)·dt + Σ(x)·dW Persistent disturbance ① Observation process dz = H(x)·dt + dV Only a noisy shadow is seen ② Peripheral region E 低 μ(x)、高 Σ(x) Where structure is undetermined Estimation theory Belief state πₜ = p(xₜ | z₁:ₜ) Observations folded into a distribution ③ Bayesian filtering πₜ ∝ p(zₜ|xₜ)·∫p(xₜ|xₜ₋₁)πₜ₋₁ Every observation rewrites belief Approximate solvers Kalman / particle filter Fallback when dimensions grow Optimization theory Value function V(π) = max E[∫ r·e^-ρt dt] Value defined on belief space ④ HJB equation ρV = max_u { r + ℒV } Hamilton–Jacobi–Bellman on beliefs Optimal policy u* = argmax_u { r + ℒV } Policy reset as belief updates Partially observable stochastic optimal control (POMDP) State hidden → maintain belief → optimize on belief belief update Key concepts ① Non-equilibrium forcing dW keeps driving the system off equilibrium anchored → state process ② Weak signals Low SNR: small H, large dV anchored → observation ③ Attention density Concentration of belief: entropy, Σ anchored → belief state ④ Generative value The ΔV that belief updates create anchored → value function Edge–Generation Theory Edge–Generation Theory Applications Strategic management Technological innovation Organizational design Investment decisions Edge = the epistemic gap created by partial observability: we see the shadow, not the state. Generation = the reset of policy after belief updates: not forecasting the future, but reorganizing action now. constitutes / depends on feedback loop ①–④ mark the anchors: each concept maps to the numbered object on the left
Figure 1　Technical architecture of edge–generation theory The labels are dense; open the vector original for full size.

3. Four words, each anchored to an object

The figure is easy to misread as a collage of concepts, so each word points to a numbered object on it:

  • ① Non-equilibrium forcing — the disturbance term dW in the state equation that never decays. The system is always being pushed away from equilibrium; there is no day on which it settles.
  • ② Weak signals — low signal-to-noise in the observation equation: small H, large dV. A weak signal is not little information; it is information diluted by noise.
  • ③ Attention density — how concentrated the belief distribution is, measured by entropy or covariance. Attention is not interest; it is belief sharp enough to bet on.
  • ④ Generative value — the increment in the value function created by a belief update. Value is not what you found; it is what you can now do because your judgment changed.

If the four words can be pointed at like this, they stop being rhetoric and become discussable. If they cannot, they are just pleasant.

4. What the figure can and cannot do

It can supply a common coordinate system. Strategy, investment and organizational design stop being four parallel fields and become instances of one observe–believe–act loop at different scales. It also makes vague judgments accountable: when someone says their intuition is good, is the prior sharper, or the observation weighted higher?

It cannot produce a single estimable parameter. The dimension of the belief space, the form of H, the solution of the value function — the figure has symbols, not numbers. Which means it is currently a language of analysis, not a computable model.

Put plainly: the figure cannot tell you what to buy. What it can tell you is what you mean when you say you see something others do not — whether your observation is better, or your prior is.

5. So what is it, for now

A working hypothesis. The four concepts have definitions but no measurement; whether the language pays off depends on whether someone writes it into something that runs — myself included.

I publish it because a framework that can be refuted is more useful than a posture that happens to be right. If one of the anchors is wrong, write and tell me the number.

← All notes