Probability Theory — Chapters 41–48

Chapter 41 — Branching Processes

41.1 Galton–Watson processes

A Galton–Watson process models a population where each individual independently produces a random number of offspring with common offspring distribution $({𝑝}_{𝑘}{)}_{𝑘\ge 0}$. If ${𝑍}_{𝑛}$ is the population size in generation $𝑛$, then

$${𝑍}_{𝑛+1}=\sum _{𝑖=1}^{{𝑍}_{𝑛}}{𝜉}_{𝑛,𝑖},$$

where the ${𝜉}_{𝑛,𝑖}$ are iid with $𝑃(𝜉=𝑘)={𝑝}_{𝑘}$. The process starts from ${𝑍}_0=1$ or another specified initial population.

The central parameter is the offspring mean

$$𝑚=\mathbb{𝐸}[𝜉]=\sum _{𝑘\ge 0}𝑘{𝑝}_{𝑘}\mathrm{.}$$

The regimes are subcritical $𝑚<1$, critical $𝑚=1$, and supercritical $𝑚>1$. The process is Markovian, but its structure is genealogical: the next generation is a random sum indexed by the current generation.

41.2 Extinction probability

The extinction event is

$$\{{𝑍}_{𝑛}=0\text{ eventually}\}\mathrm{.}$$

Once ${𝑍}_{𝑛}=0$, the process stays at zero. Let $𝑞$ be the extinction probability starting from one ancestor. If

$$𝑓(𝑠)=\mathbb{𝐸}[{𝑠}^{𝜉}]=\sum _{𝑘=0}^{\mathrm{\infty }}{𝑝}_{𝑘}{𝑠}^{𝑘}$$

is the offspring generating function, then

$$𝑞=𝑓(𝑞)\mathrm{.}$$

The extinction probability is the smallest solution of this equation in $[0,1]$.

If $𝑚\le 1$ and ${𝑝}_1\mathrm{\ne }1$, then $𝑞=1$. The population dies out almost surely. If $𝑚>1$, then $𝑞<1$ when ${𝑝}_0>0$ and reproduction is nondegenerate. The equation $𝑞=𝑓(𝑞)$ is not just algebra; it is the fixed-point certificate produced by the branching independence of descendant subtrees.

41.3 Generating functions

Generating functions linearize reproduction composition. If $𝑓$ is the offspring generating function, then the generating function of ${𝑍}_{𝑛}$ starting from one ancestor is

$${𝑓}_{𝑛}(𝑠)={𝑓}^{\circ 𝑛}(𝑠),$$

the $𝑛$-fold iterate of $𝑓$. This follows because each generation replaces every individual by an independent copy of the offspring law.

Moments are encoded by derivatives:

$$\mathbb{𝐸}[{𝑍}_{𝑛}]={𝑓}_{𝑛}^{\mathrm{\prime }}(1)={𝑚}^{𝑛}\mathrm{.}$$

If the offspring variance ${𝜎}^2$ is finite, then variance formulas can be derived by differentiating iterates. Generating functions are the native certificate engine for discrete branching: reproduction becomes functional composition.

41.4 Criticality

Criticality is governed by $𝑚=\mathbb{𝐸}𝜉$. In the subcritical case $𝑚<1$, the expected population decays:

$$\mathbb{𝐸}[{𝑍}_{𝑛}]={𝑚}^{𝑛}\to 0.$$

In the supercritical case $𝑚>1$, the expected population grows exponentially, though extinction may still occur. In the critical case $𝑚=1$, the expected population remains constant, but extinction still occurs almost surely under nondegeneracy.

Critical branching has heavy survival tails. With finite variance ${𝜎}^2$, one has asymptotically

$$𝑃({𝑍}_{𝑛}>0)\sim \frac{2}{{𝜎}^2𝑛}\mathrm{.}$$

Thus the mean remains one because rare surviving populations become large. Criticality is the boundary where expectation is misleading without survival-conditioning analysis.

41.5 Martingales in branching processes

In a Galton–Watson process with mean $𝑚>0$, the normalized population

$${𝑊}_{𝑛}=\frac{{𝑍}_{𝑛}}{{𝑚}^{𝑛}}$$

is a nonnegative martingale:

$$\mathbb{𝐸}[{𝑊}_{𝑛+1}\mid {\mathcal{𝐹}}_{𝑛}]={𝑊}_{𝑛}\mathrm{.}$$

Therefore ${𝑊}_{𝑛}\to 𝑊$ almost surely by martingale convergence.

The martingale limit $𝑊$ contains the survival-growth information. In the supercritical case, $𝑊$ may be positive on survival under additional integrability conditions, classically involving $\mathbb{𝐸}[𝜉{\log }^{+}𝜉]<\mathrm{\infty }$. Without sufficient integrability, the normalized martingale can collapse to zero despite population survival. This is a tail-control debt.

41.6 Continuous-time branching

In continuous-time branching, individuals live for random times and reproduce or die according to rates. A basic birth-death branching process has per-individual birth rate $𝜆$ and death rate $𝜇$. If ${𝑍}_{𝑡}$ is the population,

$$𝑛\to 𝑛+1\text{ at rate }𝜆𝑛,𝑛\to 𝑛-1\text{ at rate }𝜇𝑛\mathrm{.}$$

The mean evolves as

$$\mathbb{𝐸}[{𝑍}_{𝑡}]={𝑍}_0{𝑒}^{(𝜆-𝜇)𝑡}\mathrm{.}$$

The same criticality structure appears: subcritical if $𝜆<𝜇$, critical if $𝜆=𝜇$, supercritical if $𝜆>𝜇$. Continuous-time branching is a Markov process whose generator acts on functions $𝑔(𝑛)$ by

$$𝐴𝑔(𝑛)=𝜆𝑛[𝑔(𝑛+1)-𝑔(𝑛)]+𝜇𝑛[𝑔(𝑛-1)-𝑔(𝑛)]\mathrm{.}$$

41.7 Branching random walks

A branching random walk combines reproduction with spatial displacement. Each individual produces offspring, and each child is displaced from the parent by a random increment. The population becomes a random point measure:

$${𝑍}_{𝑛}=\sum _{\mathrm{\mid }𝑢\mathrm{\mid }=𝑛}{𝛿}_{{𝑆}_{𝑢}},$$

where ${𝑆}_{𝑢}$ is the position of particle $𝑢$.

The analysis combines branching martingales, random walk large deviations, and extreme-value theory. The maximal displacement is governed by a balance between exponential growth in the number of particles and exponential decay of random-walk tail probabilities. Additive martingales of the form

$${𝑊}_{𝑛}(𝜃)=\sum _{\mathrm{\mid }𝑢\mathrm{\mid }=𝑛}{𝑒}^{-𝜃{𝑆}_{𝑢}-𝑛\mathrm{\Lambda }(𝜃)}$$

encode tilted spatial-growth structure.

41.8 Superprocesses preview

Superprocesses are measure-valued branching limits. They arise when many small particles branch and move, and the empirical mass measure converges to a random measure-valued process. The state is not a population count but a measure ${𝑋}_{𝑡}$ on space.

A typical superprocess is characterized by a spatial motion generator plus a branching mechanism. For test functions $𝑓$, the process

$$⟨{𝑋}_{𝑡},𝑓⟩$$

has martingale and nonlinear log-Laplace descriptions. Superprocesses are the continuum limit of branching particle systems, with randomness distributed over measure space rather than finite genealogical individuals.

41.9 Random trees

Branching processes generate random genealogical trees. A Galton–Watson tree has offspring counts iid according to $({𝑝}_{𝑘})$. Extinction corresponds to finiteness of the tree. Conditioning critical Galton–Watson trees to have size $𝑛$ produces important random tree models.

Random trees connect probability to combinatorics, geometry, and scaling limits. Critical conditioned trees, under finite variance, converge after rescaling to continuum random trees. The genealogy carrier retains structure invisible in the population-size process ${𝑍}_{𝑛}$. Two branching processes with the same ${𝑍}_{𝑛}$ law may still require tree-level analysis for ancestry, height, contour, and spatial embedding.

---

Chapter 42 — Percolation and Interacting Systems

42.1 Bernoulli percolation

Bernoulli percolation assigns independent open/closed states to vertices or edges of a graph. In bond percolation on a graph $𝐺=(𝑉,𝐸)$, each edge is open with probability $𝑝$, independently. The central object is the open cluster $𝐶(𝑣)$ of a vertex $𝑣$: all vertices connected to $𝑣$ by open paths.

The main question is whether an infinite open cluster exists. On infinite graphs such as ${\mathbb{𝑍}}^{𝑑}$, define

$$𝜃(𝑝)={𝑃}_{𝑝}(\mathrm{\mid }𝐶(0)\mathrm{\mid }=\mathrm{\infty })\mathrm{.}$$

Percolation is a geometric phase-transition model: local independent randomness produces global connectivity.

42.2 Critical probability

The critical probability is

$${𝑝}_{𝑐}=\inf \{𝑝:𝜃(𝑝)>0\}\mathrm{.}$$

For $𝑝<{𝑝}_{𝑐}$, all clusters are finite almost surely under standard settings. For $𝑝>{𝑝}_{𝑐}$, an infinite cluster exists with positive probability, often almost surely by ergodicity.

The value of ${𝑝}_{𝑐}$ depends on the graph. For $\mathbb{𝑍}$, ${𝑝}_{𝑐}=1$. For many higher-dimensional lattices, $0<{𝑝}_{𝑐}<1$. Critical behavior at $𝑝={𝑝}_{𝑐}$ is delicate and dimension-sensitive. The critical probability is not just a threshold number; it is the boundary between local-fragment and infinite-network regimes.

42.3 Infinite clusters

An infinite cluster is a connected component of open vertices or edges with infinitely many vertices. Existence of an infinite cluster is a tail/geometric event. In translation-invariant settings, its probability is often $0$ or $1$.

Uniqueness is a separate theorem. On ${\mathbb{𝑍}}^{𝑑}$, for $𝑝>{𝑝}_{𝑐}$, there is almost surely a unique infinite cluster. Existence and uniqueness are different certificates: existence comes from supercritical connectivity; uniqueness requires ruling out multiple macroscopic open components through geometric and ergodic arguments.

42.4 Correlation inequalities

Percolation relies on monotonicity and positive association. An event is increasing if opening more edges cannot make it false. The FKG inequality states that for increasing events $𝐴,𝐵$,

$$𝑃(𝐴\cap 𝐵)\ge 𝑃(𝐴)𝑃(𝐵)\mathrm{.}$$

This gives positive correlation between increasing events.

Correlation inequalities replace independence after conditioning on global monotone structure. They support comparison arguments, crossing estimates, uniqueness proofs, and sharp-threshold theory. The key audit is monotonicity: FKG does not apply to arbitrary events.

42.5 Sharp thresholds

A threshold is sharp if the transition from unlikely to likely occurs in a narrow parameter window. For increasing events depending on many independent bits, influence and isoperimetric inequalities often imply sharp threshold behavior.

In percolation, sharpness means subcritical cluster sizes decay rapidly and supercritical infinite connectivity appears decisively. Differential inequalities relate derivatives of event probabilities to pivotality:

$$\frac{𝑑}{𝑑𝑝}{𝑃}_{𝑝}(𝐴)=\sum _{𝑒}{𝑃}_{𝑝}(𝑒\text{ pivotal for }𝐴)$$

for finite events. Pivotal edges are the local carriers of global transition sensitivity.

42.6 Oriented percolation

Oriented percolation restricts paths to move in allowed directions, often interpreted as time-forward movement. Sites or bonds are open independently, but connectivity must respect orientation. This produces a model closer to epidemic spread, growth, and directed random media.

The phase transition concerns survival: whether an infection or open path persists indefinitely. Oriented percolation is deeply connected to the contact process and directed polymers. Directionality changes geometry: clusters have causal structure, and time becomes part of the carrier.

42.7 Contact process

The contact process is a continuous-time interacting particle system on a graph. Infected sites recover at rate $1$, and infected sites infect neighboring healthy sites at rate $𝜆$. The process has an absorbing all-healthy state.

The main question is survival versus extinction. There is a critical infection rate ${𝜆}_{𝑐}$. Below it, infection dies out almost surely; above it, survival has positive probability on infinite graphs. The contact process is the dynamic analogue of percolation: random connectivity becomes random temporal propagation.

42.8 Ising model probability carrier

The Ising model assigns spins ${𝜎}_{𝑥}\in \{-1,+1\}$ to vertices. On a finite graph, the Gibbs probability is

$$𝑃(𝜎)=\frac{1}{𝑍}\exp (𝛽\sum _{⟨𝑥,𝑦⟩}{𝜎}_{𝑥}{𝜎}_{𝑦}+ℎ\sum _{𝑥}{𝜎}_{𝑥}),$$

where $𝑍$ normalizes the measure. The parameter $𝛽$ is inverse temperature, and $ℎ$ is external field.

The Ising carrier is not independent product probability except at $𝛽=0$. Neighbor interactions induce dependence. Phase transition appears as spontaneous magnetization or nonuniqueness of infinite-volume Gibbs measures. The probability law is defined by energy weighting, not by independent site choices.

42.9 Gibbs measures

A Gibbs measure is a probability measure specified through local conditional distributions determined by an energy or interaction potential. In infinite volume, Gibbs measures are defined by DLR equations: conditional law inside a finite region given outside spins is proportional to the exponential of local energy.

Infinite-volume Gibbs measures need not be unique. Nonuniqueness corresponds to phase coexistence. Boundary conditions can select different phases. Thus the same local interaction rule may produce multiple global probability carriers. Gibbs theory is where local specification does not automatically terminalize global law.

42.10 Phase transitions

A phase transition is a nonanalytic or qualitative change in macroscopic behavior as a parameter varies. In percolation, the order parameter is $𝜃(𝑝)$. In Ising models, it may be magnetization. In contact processes, it is survival probability.

The mathematical content is not the word “phase”; it is the emergence of global structure from local random rules. Critical exponents, scaling limits, universality, correlation length, and finite-size scaling refine the transition. Phase-transition claims require specifying dimension, graph, boundary conditions, limit order, and observable.

---

Chapter 43 — Free and Noncommutative Probability

43.1 Noncommutative probability spaces

A noncommutative probability space is typically a pair $(\mathcal{𝐴},𝜑)$, where $\mathcal{𝐴}$ is a unital algebra, often a ${𝐶}^{\ast }$-algebra or von Neumann algebra, and $𝜑:\mathcal{𝐴}\to \mathbb{𝐶}$ is a state:

$$𝜑(1)=1,𝜑({𝑎}^{\ast }𝑎)\ge 0.$$

Elements of $\mathcal{𝐴}$ play the role of random variables.

The difference is that multiplication need not commute:

$$𝑎𝑏\mathrm{\ne }𝑏𝑎\mathrm{.}$$

Thus joint distributions are encoded by mixed moments

$$𝜑({𝑎}_{{𝑖}_1}{𝑎}_{{𝑖}_2}\cdots {𝑎}_{{𝑖}_{𝑘}}),$$

where order matters. Classical probability is the commutative special case.

43.2 States and traces

A state $𝜑$ is a noncommutative expectation. If it satisfies

$$𝜑(𝑎𝑏)=𝜑(𝑏𝑎),$$

it is tracial. Tracial states are central in free probability because they allow cyclic rearrangement inside moments, mirroring normalized matrix trace:

$$𝜑(𝐴)=\frac{1}{𝑁}\mathrm{Tr}(𝐴)\mathrm{.}$$

A trace is not point evaluation and not integration over a visible sample space unless the algebra is commutative. It is an abstract expectation functional. Positivity supplies probabilistic meaning; traciality supplies symmetry useful for spectral distributions and random matrices.

43.3 Free independence

Free independence is the noncommutative analogue of independence. Subalgebras ${\mathcal{𝐴}}_{𝑖}$ are freely independent if whenever ${𝑎}_{𝑗}\in {\mathcal{𝐴}}_{{𝑖}_{𝑗}}$, $𝜑({𝑎}_{𝑗})=0$, and adjacent indices differ, then

$$𝜑({𝑎}_1{𝑎}_2\cdots {𝑎}_{𝑛})=0.$$

This replaces product factorization.

Classical independence says mixed expectations factor. Free independence says centered alternating products vanish. It is the correct independence notion for large random matrices in many asymptotic regimes. Using classical independence on noncommuting variables is a carrier error.

43.4 Free convolution

Free convolution describes the law of sums of freely independent noncommutative random variables. If $𝑎$ and $𝑏$ are free with laws $𝜇$ and $𝜈$, then the law of $𝑎+𝑏$ is

$$𝜇⊞𝜈\mathrm{.}$$

Multiplicative free convolution similarly describes products under positivity or unitary hypotheses.

Transforms linearize these operations. The $𝑅$-transform satisfies

$${𝑅}_{𝜇⊞𝜈}={𝑅}_{𝜇}+{𝑅}_{𝜈}\mathrm{.}$$

This parallels the logarithm of characteristic functions for classical convolution. Free probability replaces classical convolution algebra with noncommutative convolution algebra.

43.5 Semicircular law

The semicircular law has density

$$\frac{1}{2𝜋{𝜎}^2}\sqrt{4{𝜎}^2-{𝑥}^2}{1}_{[-2𝜎,2𝜎]}(𝑥)\mathrm{.}$$

It is the free analogue of the Gaussian distribution. In free CLT, normalized sums of freely independent centered variables converge to the semicircular law.

The semicircle also appears as the limiting empirical spectral distribution of Wigner random matrices. Thus the free Gaussian is not the bell curve but the semicircle. The limit carrier has changed from scalar sums to spectra of noncommuting matrices.

43.6 Random matrices as models

Random matrices provide concrete models of noncommutative random variables. If ${𝐴}_{𝑁}$ is an $𝑁\times 𝑁$ random matrix, the normalized trace expectation

$${𝜑}_{𝑁}(𝑃)=\frac{1}{𝑁}\mathbb{𝐸}\mathrm{Tr}(𝑃({𝐴}_{𝑁},{𝐴}_{𝑁}^{\ast }))$$

defines noncommutative moments. As $𝑁\to \mathrm{\infty }$, independent random matrix ensembles often become asymptotically free.

This is the bridge between classical probability and free probability. Matrix entries may be classically independent, but matrix multiplication is noncommutative. The large-$𝑁$ spectral limit is governed by free independence, not ordinary independence of scalar variables.

43.7 Operator algebras

Operator algebras provide the native analytic setting for noncommutative probability. A ${𝐶}^{\ast }$-algebra supplies norm and involution; a von Neumann algebra supplies weak operator closure and projection structure. States, traces, conditional expectations, and spectral measures all live naturally in this environment.

The operator-algebra carrier prevents misleading scalar analogies. Noncommutative random variables may have spectral laws, but joint laws are not measures on ${\mathbb{𝑅}}^{𝑛}$ unless the variables commute. The algebraic relations and mixed moments are the joint distribution.

43.8 Classical versus free independence

Classical independence:

$$\mathbb{𝐸}\prod _{𝑖}{𝑓}_{𝑖}({𝑋}_{𝑖})=\prod _{𝑖}\mathbb{𝐸}{𝑓}_{𝑖}({𝑋}_{𝑖})$$

for variables living in commuting subalgebras.

Free independence:

$$𝜑({𝑎}_1\cdots {𝑎}_{𝑛})=0$$

for centered alternating terms from distinct subalgebras.

The two concepts answer different transport problems. Classical independence governs tensor product probability spaces. Free independence governs free product algebraic probability spaces and large random matrix limits. Neither is a metaphor for the other; they are distinct factorization grammars.

43.9 Carrier change audit

A noncommutative probability claim must declare its algebra, state, trace property, variables, notion of independence, and law representation. Scalar distribution is insufficient for noncommuting tuples. Mixed moments and order of multiplication are load-bearing.

Common errors include treating spectral measures as full joint distributions, assuming classical factorization for noncommuting variables, using free convolution for commuting independent variables, or forgetting the large-$𝑁$ limiting step from random matrices to free variables. The carrier audit is: commutative or noncommutative, tensor or free product, scalar law or mixed-moment law.

---

Chapter 44 — Rough Paths and Regularity Structures

44.1 Failure of classical pathwise integration

Classical Riemann–Stieltjes integration works well when the integrator has bounded variation or when integrand and integrator have sufficient complementary regularity. Brownian motion has infinite variation and roughness near Hölder $1\mathrm{/}2$, so ordinary pathwise integration fails for many differential equations driven by Brownian-like signals.

The Itô integral solves this probabilistically, but it is not purely pathwise and depends on adaptedness. Rough path theory asks for a deterministic pathwise calculus for rough signals. The repair is to enrich the path with higher-order iterated integral data. The path alone is not enough; one must lift it to a rough path.

44.2 Iterated integrals

For a smooth path $𝑋:[0,𝑇]\to {\mathbb{𝑅}}^{𝑑}$, the second iterated integral is

$${\mathbb{𝑋}}_{𝑠,𝑡}^{𝑖,𝑗}={\int }_{𝑠}^{𝑡}({𝑋}_{𝑟}^{𝑖}-{𝑋}_{𝑠}^{𝑖})𝑑{𝑋}_{𝑟}^{𝑗}\mathrm{.}$$

Higher iterated integrals record ordered interaction of increments. Together they form a truncated signature of the path.

For rough $𝑋$, these integrals may not be classically defined. Rough path theory treats them as additional data satisfying algebraic consistency relations such as Chen’s relation:

$${\mathbb{𝑋}}_{𝑠,𝑡}={\mathbb{𝑋}}_{𝑠,𝑢}+{\mathbb{𝑋}}_{𝑢,𝑡}+{𝑋}_{𝑠,𝑢}\otimes {𝑋}_{𝑢,𝑡}\mathrm{.}$$

This extra layer is the missing carrier for nonlinear integration.

44.3 Rough paths

A rough path consists of a path $𝑋$ plus iterated integral data up to a level determined by roughness. If $𝑋$ is Hölder $𝛼$, one needs levels up to $\lfloor 1\mathrm{/}𝛼\rfloor$. For Brownian motion, $𝛼<1\mathrm{/}2$, so level two data are needed.

The rough path lift enables solving differential equations

$$𝑑{𝑌}_{𝑡}=𝑉({𝑌}_{𝑡})𝑑{𝑋}_{𝑡}$$

pathwise. The solution map becomes continuous in the rough path topology. This is the main theorem: once the correct enhanced carrier is supplied, differential equations driven by rough signals become stable deterministic objects.

44.4 Controlled paths

A path $𝑌$ is controlled by $𝑋$ if its increments admit an expansion

$${𝑌}_{𝑠,𝑡}={𝑌}_{𝑠}^{\mathrm{\prime }}{𝑋}_{𝑠,𝑡}+{𝑅}_{𝑠,𝑡},$$

where the remainder ${𝑅}_{𝑠,𝑡}$ is of higher order. The derivative-like object ${𝑌}^{\mathrm{\prime }}$ describes how $𝑌$ locally depends on $𝑋$.

Controlled paths create an integration theory against rough paths. The rough integral

$$\int 𝑌𝑑\mathbf{𝑋}$$

uses both $𝑋$ and its iterated integrals. This formalism converts irregular integration into Taylor expansion plus algebraic correction. The derivative ${𝑌}^{\mathrm{\prime }}$ is not ordinary time derivative; it is derivative relative to the rough driver.

44.5 Stochastic PDE motivation

Stochastic PDEs often involve noise too irregular to multiply or compose classically. Equations such as KPZ, stochastic quantization, and parabolic Anderson models contain products of distributions that are not classically defined.

The problem is not just randomness; it is analytic ill-posedness. White noise is a distribution, not a function. Nonlinear functions of distributions can diverge. A theory is needed that identifies the missing counterterms and builds a stable solution map. This motivates regularity structures and paracontrolled calculus.

44.6 Regularity structures

A regularity structure provides an abstract graded space of local model symbols, a structure group for reexpansion, and models that realize symbols as concrete distributions. It generalizes Taylor expansions to singular stochastic objects.

The core idea is local expansion relative to a problem-specific basis, not just polynomials. A solution is represented by coefficients against abstract symbols, and reconstruction maps this abstract object back to an actual distribution. The structure keeps track of which products and compositions are meaningful after renormalization.

44.7 Renormalization

Renormalization subtracts divergent components to obtain finite limiting objects. In singular SPDEs, approximating the noise by smooth noise often produces solutions that diverge as smoothing is removed. One adds counterterms depending on the approximation scale:

$${\text{equation}}_{𝜀}⇝{\text{equation}}_{𝜀}-{\text{counterterm}}_{𝜀}\mathrm{.}$$

The corrected solutions may converge.

Renormalization is not optional decoration; it is the certificate that the limiting equation has been properly defined. Different regularization schemes must lead to equivalent renormalized limits if the theory is robust. The counterterm ledger is load-bearing.

44.8 Lifted stochastic carriers

Both rough paths and regularity structures replace an insufficient carrier by a lifted carrier. For rough paths:

$$𝑋⇝(𝑋,\mathbb{𝑋},\dots )\mathrm{.}$$

For regularity structures:

$$\text{distributional noise}⇝\text{model + renormalized symbols}\mathrm{.}$$

The lifted object contains the missing products or iterated interactions.

This is a general pattern in modern probability: the naive random object is not enough to define nonlinear operations continuously. One must enhance it with algebraic/analytic residue. Once lifted, the solution map becomes stable, and approximation theorems become possible.

44.9 Model versus distributional noise

A distributional noise such as white noise is not a pointwise function. Treating it as one creates illegal products. The model specifies how the noise and its derived symbols are realized at every scale and location. In regularity structures, the model is the carrier of local stochastic information.

Thus the equation is not fully specified by writing a formal SPDE. One must specify interpretation: Itô or Stratonovich, renormalization, topology of convergence, approximation scheme, and solution concept. The symbolic equation is a compressed surface; the probability carrier is the lifted analytic model.

---

Chapter 45 — Probability Carrier Audit

45.1 Declare (Ω,𝓕,P)

Every probability claim must specify its probability space or explicitly work at law level. The triple

$$(\mathrm{\Omega },\mathcal{𝐹},𝑃)$$

declares outcomes, legal events, and probability law. Without it, expressions like $𝑃(𝐴)$, $𝑋=𝑌$ almost surely, and conditional expectation lack a domain.

In practice, the carrier may be implicit when using canonical laws on standard spaces. But implicit is not absent. If variables are combined, conditioned, compared pointwise, or used dynamically, the shared space must exist. Law-level equality is not enough for joint operations.

45.2 Declare state space

A random variable is a measurable map into a state space:

$$𝑋:\mathrm{\Omega }\to 𝑆\mathrm{.}$$

The state space must carry a σ-algebra, often Borel. For real variables $𝑆=\mathbb{𝑅}$; for processes $𝑆$ may be $𝐶[0,𝑇]$, $𝐷[0,𝑇]$, ${\mathbb{𝑅}}^{\mathbb{𝑁}}$, a graph space, a measure space, or an operator algebra.

The state space determines available topology, convergence, compactness, and measurability. Weak convergence on $\mathbb{𝑅}$ is not the same as weak convergence on path space. A process-level claim must not be reduced to coordinate claims without path-carrier audit.

45.3 Declare measurable maps

A random variable must be measurable:

$${𝑋}^{-1}(𝐵)\in \mathcal{𝐹}$$

for every measurable $𝐵\subseteq 𝑆$. This is the legality gate for events of the form $\{𝑋\in 𝐵\}$. If $𝑋$ is not measurable, its law is undefined.

Measurability is especially important for suprema over uncountable sets, hitting times, path functionals, projections, conditional objects, and stochastic-process sample paths. In probability, “defined pointwise” does not imply “random variable.” The inverse-image event grammar must close.

45.4 Declare laws and pushforwards

The law of $𝑋$ is the pushforward

$${𝜇}_{𝑋}={𝑋}_{\ast }𝑃,{𝜇}_{𝑋}(𝐵)=𝑃(𝑋\in 𝐵)\mathrm{.}$$

Distribution-level claims should be stated in terms of ${𝜇}_{𝑋}$, not in terms of a particular sample-space representation. This enables carrier-invariant reasoning.

However, the law of $𝑋$ alone does not specify dependence with other variables. A claim involving $𝑋+𝑌$, $𝑋𝑌$, $𝑋\le 𝑌$, or conditional behavior requires a joint law or common probability space. Marginals are not joint data.

45.5 Declare joint carrier

A joint expression requires a joint carrier. For $𝑋$ and $𝑌$, this may be a shared $(\mathrm{\Omega },\mathcal{𝐹},𝑃)$ or a coupling $𝛾$ on $𝑆\times 𝑇$ with specified marginals:

$$𝛾(𝐴\times 𝑇)={𝜇}_{𝑋}(𝐴),𝛾(𝑆\times 𝐵)={𝜇}_{𝑌}(𝐵)\mathrm{.}$$

Independence is the special coupling

$$𝛾={𝜇}_{𝑋}\otimes {𝜇}_{𝑌}\mathrm{.}$$

Deterministic relation is a coupling supported on a graph or relation. Without joint carrier, only separate law statements are meaningful.

45.6 Declare null-set quotient

Almost sure statements are quotient-level:

$$𝑋=𝑌\text{ a.s.}\iff 𝑃(𝑋=𝑌)=1.$$

They do not imply pointwise equality. ${𝐿}^{𝑝}$ spaces consist of equivalence classes modulo null sets. Conditional expectations are unique only a.s.

When selecting versions, one must declare additional regularity. For processes, equality at each fixed time a.s. does not automatically imply indistinguishability over all times. Null-set unions over uncountable index sets are a standard failure point.

45.7 Declare product or coupling

If independence is used, product structure must be declared or proved:

$${𝜇}_{𝑋,𝑌}={𝜇}_{𝑋}\otimes {𝜇}_{𝑌}\mathrm{.}$$

If dependence is used, the coupling must be specified. Coupling is not merely a proof trick; it is the joint probability carrier that permits comparison.

Many errors come from moving between marginal and joint claims without paying coupling debt. Same distribution does not imply same variable. Same marginals do not imply same sum distribution. Conditional independence does not imply unconditional independence. Product/coupling declaration prevents these errors.

45.8 Declare topology for convergence

A convergence claim must specify mode and topology:

$$\text{a.s.},\text{in probability},{𝐿}^{𝑝},\text{weakly},\text{vaguely},\text{total variation},{𝑊}_{𝑝}\mathrm{.}$$

For process convergence, one must specify whether the topology is uniform, Skorokhod ${𝐽}_1$, ${𝑀}_1$, product topology, or another carrier.

The same symbolic arrow ${𝑋}_{𝑛}\to 𝑋$ is ambiguous. Different modes support different liftbacks. Weak convergence permits bounded-continuous test functions; ${𝐿}^1$ convergence permits expectations of absolute error; a.s. convergence permits pathwise statements modulo null sets.

45.9 Declare limit/liftback target

A limit theorem must declare what it is meant to certify. A CLT certifies distributional approximation at central scale. A local limit theorem certifies point or small-window probabilities. A large deviation principle certifies exponential-scale costs. A concentration inequality certifies finite-sample tail suppression.

The liftback target may be asymptotic law, finite-$𝑛$ bound, empirical model validity, algorithm performance, or deterministic existence. Each target requires different debt payment. An asymptotic theorem without rate does not certify finite accuracy; a formal model theorem does not certify empirical truth.

---

Chapter 46 — Probability Debt Ledger

46.1 Sample-space debt

Sample-space debt appears when random objects are named without a common carrier. A statement about $𝑋$ alone may be law-level, but a statement about $𝑋+𝑌$, $𝑋=𝑌$, or $𝑃(𝑋\le 𝑌)$ requires $𝑋,𝑌$ to live jointly.

The repair is to declare a shared probability space or construct a coupling. If only marginal laws are known, joint statements remain underdetermined. Sample-space debt is the first audit in any multi-variable probability claim.

46.2 σ-algebra debt

σ-algebra debt appears when the set of legal events is not declared. In finite spaces this is often harmless, but in continuous spaces $𝑃(𝐸)$ is meaningful only for $𝐸\in \mathcal{𝐹}$. Nonmeasurable sets and completion issues force the distinction.

The repair is to specify $\mathcal{𝐹}$, usually Borel, completed Borel, product, predictable, optional, or another process-specific σ-algebra. Different σ-algebras support different operations. Event legality is not automatic.

46.3 Measurability debt

Measurability debt appears when a function or path functional is treated as a random variable without proof. Supremum over uncountable time, hitting times, projections, selections, and conditional kernels often require measurability arguments.

The repair is to show inverse images of measurable sets are events, or to work in a path space where the functional is known measurable. Regularity such as continuity, càdlàg paths, separability, or standard Borel structure often pays this debt.

46.4 Null-set debt

Null-set debt appears when almost sure statements are exported as pointwise statements. $𝑋=𝑌$ a.s. does not imply $𝑋(𝜔)=𝑌(𝜔)$ for all $𝜔$. Conditional expectations and densities are often version-dependent.

The repair is either to stay in quotient language or select a version with additional regularity. For processes, one often needs indistinguishability rather than fixed-time equality. Null sets cannot be unioned over arbitrary uncountable index families without structure.

46.5 Law-versus-variable debt

Law-versus-variable debt appears when equality in distribution is treated as equality of random variables:

A law is a pushforward measure; a random variable is a measurable map on a carrier.

The repair is to distinguish law-level claims from coupling-level claims. If an almost sure relation is needed, construct a coupling with that relation. Skorokhod representation, quantile coupling, and optimal transport are possible liftbacks, not automatic consequences.

46.6 Coupling debt

Coupling debt appears when marginal laws are used to infer joint behavior. Knowing ${𝜇}_{𝑋}$ and ${𝜇}_{𝑌}$ does not determine $𝑃(𝑋<𝑌)$, $\mathrm{Cov}(𝑋,𝑌)$, or the law of $𝑋+𝑌$.

The repair is to specify a joint law $𝛾\in \mathrm{\Pi }({𝜇}_{𝑋},{𝜇}_{𝑌})$. Independence, monotone coupling, maximal coupling, and optimal coupling are different choices with different consequences. Coupling is the carrier for comparison.

46.7 Independence debt

Independence debt appears when multiplication is used without factorization:

$$𝑃(𝐴\cap 𝐵)=𝑃(𝐴)𝑃(𝐵),\mathbb{𝐸}[𝑋𝑌]=\mathbb{𝐸}[𝑋]\mathbb{𝐸}[𝑌],{𝜙}_{𝑋+𝑌}={𝜙}_{𝑋}{𝜙}_{𝑌}\mathrm{.}$$

Each identity requires independence or a weaker sufficient structure.

The repair is to prove product measure, independent construction, conditional independence, martingale difference structure, or dependence bounds. Pairwise independence is not mutual independence. Heuristic unrelatedness is not a certificate.

46.8 Integrability debt

Integrability debt appears when expectations are used without checking finite integral. For signed $𝑋$, $\mathbb{𝐸}[𝑋]$ is defined only if positive and negative parts are not both infinite. Many theorems require $𝑋\in {𝐿}^1$, ${𝐿}^2$, or ${𝐿}^{𝑝}$.

The repair is domination, tail bounds, moment estimates, truncation, or direct calculation. Infinite expectation is not an error for nonnegative variables, but subtracting infinities is illegal. Integrability is the gate for expectation algebra.

46.9 Moment debt

Moment debt appears when variance, CLT rates, Taylor expansions, or concentration estimates assume finite moments not established. Variance requires ${𝐿}^2$; Berry–Esseen requires a third absolute moment; many martingale inequalities require exponential or bounded increments.

The repair is to declare and prove the required moment bound. Tail formulas

$$\mathbb{𝐸}\mathrm{\mid }𝑋{\mathrm{\mid }}^{𝑝}=𝑝{\int }_0^{\mathrm{\infty }}{𝑡}^{𝑝-1}𝑃(\mathrm{\mid }𝑋\mathrm{\mid }>𝑡)𝑑𝑡$$

often expose the debt. Heavy tails frequently break moment-based arguments.

46.10 Conditioning debt

Conditioning debt appears when conditional probabilities are used without specifying event, σ-algebra, or regular conditional law. Conditioning on $𝐵$ requires $𝑃(𝐵)>0$. Conditioning on $𝑌=𝑦$ usually requires a regular conditional distribution.

The repair is to state conditioning as $\mathbb{𝐸}[\cdot \mid \mathcal{𝐺}]$, as a kernel, or as a disintegration. Null-event conditioning requires a chosen version or limiting scheme. Informal conditioning often hides nonuniqueness.

46.11 Version debt

Version debt appears because conditional expectations, densities, and regular conditional probabilities are defined only almost surely. Evaluating them at exceptional points can be meaningless or convention-dependent.

The repair is to choose a regular version with continuity, càdlàg structure, or other pointwise properties, and prove it has those properties. Otherwise statements must remain almost-sure or law-level.

46.12 Filtration debt

Filtration debt appears in time-dependent probability when $({\mathcal{𝐹}}_{𝑡})$ is omitted. Martingales, stopping times, adaptedness, predictability, and stochastic integrals are defined relative to a filtration.

The repair is to declare the natural, augmented, completed, enlarged, or problem-specific filtration. Changing the filtration changes martingale and stopping-time status. Future information leakage is a filtration error.

46.13 Stopping-time debt

Stopping-time debt appears when a random time is used in optional stopping or strong Markov arguments without verifying

$$\{𝜏\le 𝑡\}\in {\mathcal{𝐹}}_{𝑡}\mathrm{.}$$

Times depending on future values are not stopping times.

The repair is to prove stopping-time measurability using adaptedness and path regularity. Hitting times often require closed/open set and continuity/càdlàg hypotheses. Optional stopping also needs boundedness, integrability, or uniform integrability assumptions.

46.14 Convergence-mode debt

Convergence-mode debt appears when ${𝑋}_{𝑛}\to 𝑋$ is written without specifying mode. Almost sure, in probability, ${𝐿}^{𝑝}$, distributional, total variation, and weak convergence are different statements.

The repair is to type every arrow. Each mode gives different consequences. Distributional convergence cannot be used as probability convergence without coupling. Probability convergence does not give expectation convergence without uniform integrability.

46.15 Tightness debt

Tightness debt appears when weak subsequential limits or process convergence are claimed without compactness control. Finite-dimensional distributions do not imply process-level convergence.

The repair is tightness in the relevant topology. On Polish spaces, Prokhorov gives relative compactness. In function spaces, modulus-of-continuity or oscillation bounds are typical tightness certificates.

46.16 Uniform-integrability debt

Uniform-integrability debt appears when expectations are passed through limits from convergence in probability or distribution. Rare large values can preserve or destroy expectations despite typical convergence.

The repair is uniform integrability:

$$\underset{𝑛}{\sup }\mathbb{𝐸}[\mathrm{\mid }{𝑋}_{𝑛}\mathrm{\mid }{1}_{\{\mathrm{\mid }{𝑋}_{𝑛}\mathrm{\mid }>𝐾\}}]\to 0.$$

Boundedness in ${𝐿}^{𝑝}$ for $𝑝>1$ often suffices. This is the tail bridge for expectation convergence.

46.17 Tail-control debt

Tail-control debt appears when central approximations are used in extreme regimes. CLT-scale estimates do not automatically describe rare events, maxima, or large deviations.

The repair is concentration, large deviations, extreme-value theory, moment bounds, exponential moments, or truncation. The scale of deviation must be declared before the theorem can be selected.

46.18 Lattice-obstruction debt

Lattice-obstruction debt appears in local limit theorems and point probabilities. If a sum lives on $𝑎+ℎ\mathbb{𝑍}$, probabilities outside that lattice are exactly zero. A continuous Gaussian density cannot directly approximate inaccessible points.

The repair is span and aperiodicity audit. Local estimates must include lattice cell size and admissible residue classes. Global CLT may ignore arithmetic; local CLT cannot.

46.19 Asymptotic-to-finite debt

Asymptotic-to-finite debt appears when a limit theorem is treated as a finite-$𝑛$ guarantee. A CLT without rate does not quantify error at sample size $𝑛$. An LDP gives exponential rate but may omit prefactors.

The repair is a finite-sample bound, rate theorem, Berry–Esseen estimate, concentration inequality, nonasymptotic coupling, or explicit remainder. The terminal claim must match the theorem’s output.

46.20 Empirical-model debt

Empirical-model debt appears when a formal probability model is exported as a real-world claim. A theorem about iid samples does not prove data are iid. A stochastic differential equation does not prove the physical system follows that SDE.

The repair is calibration, diagnostics, model selection, measurement-error analysis, causal design, or domain validation. Probability theory certifies consequences of models. Empirical use requires proving the model is an adequate carrier.

---

Chapter 47 — Probability Counterkernels

47.1 Unmeasurable event

The counterkernel is assigning $𝑃(𝐸)$ when $𝐸\mathrm{\notin }\mathcal{𝐹}$. The notation resembles probability but lacks legal event status. This appears with pathological subsets, projections, uncountable operations, and informal sample-space descriptions.

The repair is to prove measurability or replace $𝐸$ with a measurable approximation. In continuous probability, the full power set is not a safe event space. Event legality precedes probability assignment.

47.2 Uncountable union laundering

The error is treating σ-algebras as closed under arbitrary unions. Countable unions are legal; uncountable unions require separate proof. A common false move is summing probabilities over uncountably many null events.

The repair is separability, countable dense reduction, analytic-set machinery, or direct measurability proof. Continuous variables show the danger:

$$𝑃(𝑋=𝑥)=0\mathrm{\forall }𝑥,𝑃(𝑋\in \mathbb{𝑅})=1.$$

Uncountable union of null events need not be null by countable additivity.

47.3 A.s. equality exported as pointwise equality

The error is replacing $𝑋=𝑌$ almost surely by $𝑋(𝜔)=𝑌(𝜔)$ for every $𝜔$. This can invalidate pathwise, stopping-time, or point-evaluation arguments.

The repair is to remain in equivalence classes or choose versions. For processes, upgrade fixed-time a.s. equality to indistinguishability only with regularity. Null sets are small in measure, not nonexistent.

47.4 Equality in law exported as equality of variables

The error is using

as if $𝑋=𝑌$ or $𝑃(𝑋=𝑌)=1$. Equality in law means only ${𝜇}_{𝑋}={𝜇}_{𝑌}$. It does not define a joint relation.

The repair is coupling. If a relation is needed, construct a joint law. Quantile coupling can make real variables with the same law equal on a common space; independent coupling makes them unrelated. The law alone does not choose.

47.5 Marginals exported as joint law

The error is inferring joint probabilities from marginal laws. Knowing $𝑋\sim 𝜇$ and $𝑌\sim 𝜈$ does not determine $\mathcal{𝐿}(𝑋,𝑌)$. The law of $𝑋+𝑌$, covariance, and ordering probability are all coupling-dependent.

The repair is to declare $𝛾\in \mathrm{\Pi }(𝜇,𝜈)$. Independence is $𝛾=𝜇\otimes 𝜈$, not a default. Dependence structure is mathematical data.

47.6 Independence asserted without product factorization

The error is saying variables are independent because they are “unrelated,” “random,” or “generated separately” without a product or factorization certificate. Independence must satisfy

$$𝑃(𝑋\in 𝐴,𝑌\in 𝐵)=𝑃(𝑋\in 𝐴)𝑃(𝑌\in 𝐵)\mathrm{.}$$

The repair is construction on product space, proof by factorization, conditional independence plus integration, or a model assumption explicitly declared. Many probability formulas silently require independence; each multiplication must be audited.

47.7 Pairwise independence exported as mutual independence

The error is checking only pairwise factorization and using results requiring full mutual independence. Pairwise independence may control variances but not products of many events, Chernoff bounds, or iid limit theorems.

The repair is to prove finite-subfamily factorization:

$$𝑃\left(\bigcap _{𝑖\in 𝐼}{𝐴}_{𝑖}\right)=\prod _{𝑖\in 𝐼}𝑃({𝐴}_{𝑖})$$

for every finite $𝐼$. Or use a theorem designed for pairwise independence. Independence strength must match proof engine.

47.8 Expectation used without integrability

The error is manipulating $\mathbb{𝐸}[𝑋]$ when it is undefined or infinite in an incompatible way. For signed $𝑋$, both positive and negative parts cannot be infinite.

The repair is ${𝐿}^1$ verification, truncation, or nonnegative extended-expectation formalism. Principal-value cancellation is not Lebesgue expectation. Expectation algebra requires integrability.

47.9 Variance used without second moment

The error is invoking variance, Chebyshev, covariance, ${𝐿}^2$ projection, or CLT variance scale without $\mathbb{𝐸}[{𝑋}^2]<\mathrm{\infty }$. Heavy-tailed variables often break this.

The repair is to prove $𝑋\in {𝐿}^2$ or use a theorem for infinite-variance regimes, such as stable laws or truncation methods. Variance is not a formal symbol; it is an integral requiring finiteness.

47.10 Conditional probability on null event

The error is applying

$$𝑃(𝐴\mid 𝐵)=\frac{𝑃(𝐴\cap 𝐵)}{𝑃(𝐵)}$$

when $𝑃(𝐵)=0$. Conditioning on $𝑋=𝑥$ for continuous $𝑋$ is not handled by this ratio.

The repair is a regular conditional probability, disintegration, or specified limiting scheme. Different versions or coordinate choices can disagree on null conditioning sets. Null conditioning must be carrier-explicit.

47.11 Conditional expectation version treated as canonical

The error is evaluating or comparing conditional expectations at points where only an a.s. equivalence class is defined. A version may be changed on a null set without altering the conditional expectation.

The repair is to work a.s. or select a regular version with additional properties. In Markov processes, choosing transition kernels or conditional densities often requires standard Borel assumptions and version control.

47.12 Convergence arrow used without mode

The error is writing ${𝑋}_{𝑛}\to 𝑋$ and applying whichever consequences are convenient. This collapses a.s., probability, ${𝐿}^{𝑝}$, distributional, weak, total variation, and other modes.

The repair is to type the convergence. Every arrow has a carrier and consequence set. Distributional convergence is law-level; ${𝐿}^{𝑝}$ convergence is moment-level; a.s. convergence is path-level.

47.13 Distributional convergence exported as probability convergence

The error is treating ${𝑋}_{𝑛}\Rightarrow 𝑋$ as if $𝑃(\mathrm{\mid }{𝑋}_{𝑛}-𝑋\mathrm{\mid }>𝜀)\to 0$. This expression requires a common space and a coupling relation. Weak convergence alone does not supply it.

The repair is Skorokhod representation if changing carrier is allowed, or prove convergence in probability directly. If the limit is constant, weak convergence does imply convergence in probability to that constant under a common interpretation. Otherwise joint structure is needed.

47.14 CLT exported into tail regime

The error is using a central limit theorem to estimate probabilities far from $\sqrt{𝑛}$-scale fluctuations. The CLT governs

$${𝑆}_{𝑛}-𝑛𝜇=𝑂(\sqrt{𝑛}),$$

not order-$𝑛$ deviations or extreme tails.

The repair is large deviations, moderate deviations, concentration, or heavy-tail theory depending on scale. Gaussian approximation has a domain. Tail claims require their own certificates.

47.15 Local limit theorem without lattice audit

The error is approximating point probabilities by a continuous density without checking lattice support, span, or aperiodicity. If ${𝑆}_{𝑛}$ only takes even values, odd-point probabilities are zero regardless of Gaussian shape.

The repair is to state the local theorem on admissible lattice points with correct span factor, or use density/window formulations for nonlattice variables. Local probability is arithmetic-sensitive.

47.16 Asymptotic theorem exported as finite exact certificate

The error is using ${𝑋}_{𝑛}\Rightarrow 𝑋$, $𝑃({𝐴}_{𝑛})\to 0$, or ${𝑎}_{𝑛}\sim {𝑏}_{𝑛}$ as if it gives exact finite-$𝑛$ truth. Asymptotic statements describe limiting behavior, not finite equality.

The repair is error bounds, explicit rates, finite-sample inequalities, or computational verification. The theorem’s terminal class must be asymptotic unless a quantitative liftback is supplied.

47.17 Formal probability model exported as empirical fact

The error is moving from theorem to world without validating the model. An iid theorem does not prove observations are iid; a Markov model does not prove the system is Markov; a Gaussian noise model does not prove noise is Gaussian.

The repair is empirical calibration, experimental design, residual analysis, causal justification, robustness checks, or domain-specific mechanism. Probability theory gives internal consequences of a model. Empirical truth is an additional liftback.

---

Chapter 48 — Probability Terminal Classes

48.1 PROBABILITY_CARRIER_CERT

This terminal means a valid probability carrier has been declared:

$$(\mathrm{\Omega },\mathcal{𝐹},𝑃)$$

or an equivalent law-level measurable-space carrier. Events, variables, and operations have a domain. This is the minimum certificate for probability syntax.

It does not prove any substantive theorem by itself. It only establishes that probabilistic expressions are legally interpretable. Further terminals are needed for events, laws, independence, convergence, or limits.

48.2 EVENT_CERT

This terminal means the event under discussion is measurable:

$$𝐸\in \mathcal{𝐹}\mathrm{.}$$

The probability $𝑃(𝐸)$ is therefore well-defined. For complex events, this may require closure under countable operations, path regularity, stopping-time measurability, or analytic-set arguments.

An event certificate is not a probability estimate. It only licenses the question. The next layer computes or bounds $𝑃(𝐸)$.

48.3 RANDOM_VARIABLE_CERT

This terminal means a map $𝑋:\mathrm{\Omega }\to 𝑆$ is measurable:

$${𝑋}^{-1}(𝐵)\in \mathcal{𝐹}$$

for all measurable $𝐵\subseteq 𝑆$. Thus $𝑋$ has a law and its measurable properties are events.

This certificate is especially important for process functionals, suprema, hitting times, and random elements in nontrivial spaces. A formula-defined object is not a random variable until measurability closes.

48.4 LAW_CERT

This terminal means the distribution of $𝑋$ has been identified:

$${𝜇}_{𝑋}={𝑋}_{\ast }𝑃\mathrm{.}$$

It may be represented by a CDF, density, mass function, characteristic function, Laplace transform, or abstract measure.

A law certificate is marginal unless explicitly joint. It supports distributional statements and expectations of functions of $𝑋$, but not joint comparisons with other variables unless a coupling or joint law is also certified.

48.5 MODEL_EXTENSION_CERT

This terminal means a probability space has been enlarged while preserving old probabilistic content. If $𝜋:{\mathrm{\Omega }}^{\mathrm{\prime }}\to \mathrm{\Omega }$ is the projection or reduction map, then

$${𝑃}^{\mathrm{\prime }}({𝜋}^{-1}𝐸)=𝑃(𝐸)\mathrm{.}$$

Old variables lift as $𝑋\circ 𝜋$.

This certificate permits adding independent randomness, constructing copies, or placing variables on a richer carrier. It proves representation changed without changing the original law.

48.6 COUPLING_CERT

This terminal means a joint law $𝛾$ with specified marginals has been constructed:

$$𝛾\in \mathrm{\Pi }(𝜇,𝜈)\mathrm{.}$$

It licenses statements involving both variables simultaneously.

Different couplings certify different relations: independence, monotonicity, equality, maximal matching, transport optimality, or dependence bounds. Coupling is not unique unless additional structure is imposed.

48.7 INDEPENDENCE_CERT

This terminal means product factorization has been established:

$${𝜇}_{{𝑋}_1,\dots ,{𝑋}_{𝑛}}=\underset{𝑖=1}{\overset{𝑛}{⨂}}{𝜇}_{{𝑋}_{𝑖}}$$

or equivalently σ-algebra/event factorization for all finite subfamilies.

It licenses multiplication of probabilities, factorization of expectations, product MGFs, independent-sum transforms, and iid limit theorem inputs. It does not follow from naming variables separately.

48.8 EXPECTATION_CERT

This terminal means the expectation is legally defined as an integral:

$$\mathbb{𝐸}[𝑋]=\int 𝑋𝑑𝑃\mathrm{.}$$

For signed variables, integrability or one-sided extended legitimacy has been checked.

It licenses expectation algebra only within its integrability scope. If $𝑋\mathrm{\notin }{𝐿}^1$, expectation may be infinite or undefined. The certificate must specify which case holds.

48.9 CONDITIONING_CERT

This terminal means conditioning has been defined through a positive-probability event, sub-σ-algebra, conditional expectation, regular conditional probability, or disintegration. The conditioning carrier is explicit.

For σ-algebra conditioning, the certificate is:

$$\mathbb{𝐸}[𝑋\mid \mathcal{𝐺}]\text{ is }\mathcal{𝐺}\text{-measurable and matches integrals over }\mathcal{𝐺}\mathrm{.}$$

For kernels, regularity and version status must be stated.

48.10 MARTINGALE_CERT

This terminal means a process is adapted, integrable, and satisfies

$$\mathbb{𝐸}[{𝑀}_{𝑡}\mid {\mathcal{𝐹}}_{𝑠}]={𝑀}_{𝑠}\mathrm{.}$$

The filtration is part of the certificate. Without it, martingale status is undefined.

A martingale certificate licenses optional sampling, convergence, stochastic integration, or concentration only after the additional hypotheses of those theorems are paid. Martingale alone is not enough for every martingale theorem.

48.11 CONVERGENCE_CERT

This terminal means a convergence claim has been proved in a declared mode: a.s., in probability, ${𝐿}^{𝑝}$, distribution, total variation, weak, Wasserstein, path-space topology, or another precise mode.

The certificate includes the consequence set. For example, weak convergence certifies bounded-continuous test integrals; ${𝐿}^1$ convergence certifies expectation of absolute error; a.s. convergence certifies pathwise convergence modulo null set.

48.12 LIMIT_THEOREM_CERT

This terminal means a theorem such as LLN, CLT, Poisson limit, LDP, martingale limit, invariance principle, or ergodic theorem has been applied with hypotheses paid. It must include normalization, convergence mode, and target law/object.

A limit theorem certificate is often asymptotic. It does not automatically produce finite-sample accuracy. Rate or error terminals are separate.

48.13 TAIL_BOUND_CERT

This terminal means a nonasymptotic or asymptotic tail estimate has been established:

$$𝑃(𝑋\ge 𝑡)\le \text{bound}(𝑡)\mathrm{.}$$

The bound may come from Markov, Chebyshev, Chernoff, Hoeffding, Bernstein, martingales, concentration, or large deviations.

The certificate must specify the tail regime. Central, moderate, large-deviation, and extreme-value tails are not interchangeable.

48.14 LOCAL_LIMIT_CERT

This terminal means point probabilities, density values, or small-window probabilities have been asymptotically approximated with lattice/nonlattice scope declared. A typical lattice output is

$$𝑃({𝑆}_{𝑛}=𝑘)\sim \frac{ℎ}{𝜎\sqrt{𝑛}}𝜑\left(\frac{𝑘-𝑛𝜇}{𝜎\sqrt{𝑛}}\right)$$

on admissible lattice points.

This certificate is stronger than weak convergence. It requires Fourier, aperiodicity, smoothness, or span conditions. Without lattice audit, the terminal is invalid.

48.15 PROCESS_CERT

This terminal means a stochastic process has been constructed with specified finite-dimensional distributions, path carrier, filtration, and regularity. Kolmogorov extension alone gives product-space existence; continuity or càdlàg paths require additional proof.

A process certificate must distinguish modification, indistinguishability, version, and path topology. Coordinate laws alone are not process-level law.

48.16 MEASURABILITY_RESIDUE

This residue remains when a set, map, stopping time, supremum, projection, or path functional has not been proved measurable. The probability expression is not yet legal.

The next executable step is a measurability proof, regularity assumption, separability reduction, or replacement by an outer probability/analytic-set framework. Without this, terminal probability claims are blocked.

48.17 COUPLING_RESIDUE

This residue remains when marginal laws are known but joint behavior is required. Questions about sums, comparisons, covariance, equality, or dependence cannot close from marginals alone.

The next step is to construct or identify a coupling. Independence, monotone coupling, maximal coupling, optimal transport, or process coupling may be appropriate depending on the target.

48.18 INDEPENDENCE_RESIDUE

This residue remains when a proof requires independence but only weaker or unspecified dependence information is available. Factorization has not been paid.

The next step is either prove product structure, weaken the theorem to one requiring less independence, or supply dependence-control substitutes such as mixing, martingale differences, negative association, dependency graphs, or coupling bounds.

48.19 INTEGRABILITY_RESIDUE

This residue remains when expectation, variance, conditional expectation, martingale status, or limit exchange requires finite integral or moment not established.

The next step is a tail bound, truncation argument, domination, moment calculation, or replacement by a theorem valid under weaker integrability. If moments are infinite, the theorem route must change.

48.20 CONVERGENCE_MODE_RESIDUE

This residue remains when a convergence arrow lacks a mode or when the proved mode is too weak for the desired conclusion. For example, weak convergence cannot support unbounded expectation convergence without additional control.

The next step is to prove the stronger mode or add the missing bridge, such as uniform integrability, coupling, tightness, or rate estimates.

48.21 ASYMPTOTIC_LIFTBACK_RESIDUE

This residue remains when an asymptotic result is used for a finite or concrete claim without error control. The limit theorem may be correct but insufficient for the requested output.

The next step is Berry–Esseen, concentration, explicit constants, finite-sample inequality, computational verification, or asymptotic expansion with remainder. Without rate, the claim remains asymptotic.

48.22 PROBABILITY_LAUNDERING_CK

This counterkernel materializes when a probability result is exported outside its carrier: model to reality without validation, law to coupling without construction, a.s. to pointwise, convergence in distribution to convergence in probability, relaxation to exact finite claim, or independence by narrative.

The terminal is not residue but active failure. The repair is to retreat to the certified carrier, state the exact theorem, and rebuild the missing liftback explicitly.