The Full Mathematical Flow of ULR

Open source document →

1. #1

   ULR mathematics does not begin by asserting that a common latent object exists. It begins with a limited alignment phenomenon and, before interpreting that phenomenon, specifies separate equivalence relations for coordinates, functions, relations, experiments, and time.

   $$\begin{aligned} \text{observed alignment} &\to \text{typed observable}\ (O,\mathcal G,N,\mathcal S)\\ &\to \text{architecture gauge and function fibre}\\ &\to \text{quotient, lift, gluing, assembly}\\ &\to \text{observer-relative response quotient}\\ &\to \text{strong typed baselines and carrier reduction}\\ &\to \text{object-relative formation test}\\ &\to \text{ontology verdict}. \end{aligned}$$
2. #2

   Write the item representation of model as a matrix . What ULR first observed was not agreement between the raw coordinates of , but a limited alignment of inter-item relation geometry. Define the centred Gram matrix by

   $$K_m=JH_mH_m^\top J, \qquad J=I-\frac{1}{n}\mathbf 1\mathbf 1^\top$$
3. #3

   Provided that and , linear CKA can be read as the cosine between two centred Gram matrices.

   $$\operatorname{CKA}(H_a,H_b) = \frac{\langle K_a,K_b\rangle_F} {\lVert K_a\rVert_F\,\lVert K_b\rVert_F}.$$
4. #4

   $$\text{specified pairs share a measured relation statistic above a specified null}.$$
5. #5

   Every representation observable declares at least the following triple.

   $$\boxed{(O,\mathcal G,N)}$$
6. #6

   Comparisons across models or times may require extending this contract to the following quadruple.

   $$(O,\mathcal G,N,\mathcal S).$$
7. #7

   Let denote parameters and the function implemented by the architecture. An admissible gauge may satisfy

   $$f_{g\theta}=f_\theta$$
8. #8

   An identity candidate must therefore be relative at least to the orbit space, not to the raw parameter space .

   $$[\theta]_{\mathcal G} = \{g\theta:g\in\mathcal G\}.$$
9. #9

   Even here, the two directions must be distinguished.

   $$g\theta\in[\theta]_{\mathcal G} \Rightarrow f_{g\theta}=f_\theta,$$
10. #10

    whereas

    $$f_{\theta'}=f_\theta \not\Rightarrow \theta'\in[\theta]_{\mathcal G}$$
11. #11

    A local change permitted by an equation at a node becomes a global gauge only if it lifts to an actual parameter reparameterisation, is compatible across branches and residual paths, and preserves the entire network function. This layer is organised by the filtration

    $$\mathcal G_{\mathrm{loc}} \supseteq \mathcal G_{\mathrm{lift}} \supseteq \mathcal G_{\mathrm{compat}} \supseteq \mathcal G_{\mathrm{global}}.$$
12. #12

    For a function , define its fibre by

    $$\mathcal F_f=\{\theta\in\Theta:f_\theta=f\}$$
13. #13

    The following inference is prohibited, however.

    $$\mathcal F_f/\mathcal G\neq\{*\} \quad\not\Rightarrow\quad \text{different meanings or circuits}.$$
14. #14

    Channel incidence asks whether off-diagonal pairs can be excluded in

    $$\mathcal I_{\mathrm{full}} \subseteq \Theta\times X_{\mathrm{ch}}\times X_{\mathrm{ch}}$$
15. #15

    The assembly side instead asks about the tangent and ramification of a quotient map on . Consequently,

    $$\text{channel injectivity}+\text{assembly unramifiedness}$$
16. #16

    Suppose the local certificate is sound, so that at each audited point. Even then, the existence of one good point in a fibre does not mean that the entire fibre is collision-free.

    $$\forall\theta\ \exists x\ \mathrm{Good}(\theta,x) \quad\not\Rightarrow\quad \forall\theta\ \forall x\ \mathrm{NoCollision}(\theta,x).$$
17. #17

    Because learning and inference can use the same local operator, the presence or absence of a weight update is not a sufficient definition. The current candidate typing is

    $$(\text{updated state},\ \text{information/objective source},\ \text{write scope}, \ \text{persistence},\ \text{reset/clone contract},\ \text{future causal role}).$$
18. #18

    For a linear system, let be the persistence law, the read port, and the preparation basis. The reset response is

    $$Y_R=HRB$$
19. #19

    If has full column rank and has full row rank—equivalently, has a left inverse and has a right inverse—then can be recovered from . At a partial port, however, even matrices of different ranks may produce the same .

    $$R\sim_{H,B}R' \iff HRB=HR'B.$$
20. #20

    Let be the role and the observer filtration. A deterministic exact classifier exists if and only if

    $$C\text{ is }\mathcal F^O\text{-measurable}$$
21. #21

    For probabilistic transcript laws , the equal-prior Bayes risk is

    $$R^*=\frac{1-\lVert P_L-P_I\rVert_{\mathrm{TV}}}{2}.$$
22. #22

    For a set of admissible policies, the optimal discrimination power of an adaptive observer is

    $$\sup_{\pi\in\Pi} \lVert P_L^\pi-P_I^\pi\rVert_{\mathrm{TV}}$$
23. #23

    If the internal transcript is a Markov garbling of the external transcript, then

    $$\lVert KP_L-KP_I\rVert_{\mathrm{TV}} \le \lVert P_L-P_I\rVert_{\mathrm{TV}}.$$
24. #24

    An incremental residual must at least satisfy

    $$\Delta(Q;B^*) = \operatorname{Risk}(B^*)- \operatorname{Risk}(B^*,Q) >0$$
25. #25

    Z0–Z7 form a registry of competing explanations. The present decision places Z0, which requires the least ontology, first; another hypothesis defeats Z0 only by making an independent prediction.

    $$\begin{array}{ll} Z0:&\text{typed pluralism/deflation},\\ Z1:&\text{physical response},\\ Z2:&\text{predictive state},\\ Z3:&\text{possibility field},\\ Z4:&\text{multiscale dynamics},\\ Z5:&\text{external relation},\\ Z6:&\text{causal circuit},\\ Z7:&\text{coherence formation}. \end{array}$$
26. #26

    The result is 0/7.

    $$\begin{aligned} \text{parameter quotient}&\to\text{symmetry/invariants},\\ \text{activation ensemble}&\to\text{empirical law/kernel/moment},\\ \text{relation transform}&\to\text{RSA/CKA/external relation},\\ \text{predictive state}&\to\text{PSR/minimal realisation},\\ \text{physical response}&\to\text{causal testing},\\ \text{dynamical carrier}&\to\text{state-space/RDS},\\ \text{external relation}&\notin\text{internal carrier}. \end{aligned}$$
27. #27

    If formation is defined only as the time at which a scalar becomes large, then choices of coordinates, metric, and threshold manufacture the event. Let be the declared time index. The conditional ULR definition first requires a time-indexed object package.

    $$\mathcal O_t= (X_t,\sim_t,\{\tau_{t\to s}\}_{s\in T},\operatorname{Type}_t), \qquad t\in T,$$
28. #28

    where is identity, is the declared family of cross-time transport maps, and is structural type. A formation event must be a change of structural stratum under transport, not merely a change in the quotient value.

    $$\operatorname{Type}_{t^-}(\mathcal O) \neq \operatorname{Type}_{t^+}(\mathcal O).$$
29. #29

    Therefore,

    $$\boxed{ \text{strong phenomena and boundary theorems} \quad\land\quad \text{no extra ontology required today} }$$