A process is a family of random variables
Objective: Explain A process is a family of random variables to a learner who has not seen the notation. Decode X_t means the random variable selected at time t; X_t(omega) is its value in outcome omega; t is not random here. derive Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists. reproduce The path values are (3,6,-3), while the marginal variances are (1,4,1). and identify the first failed line under Choosing a fresh unrelated probability space for every t does not yet provide a common sample path.
Unit 1, chapter 1 asks a single concrete question: What changes when one Gaussian random variable is indexed by time? The answer is not allowed to begin with unexplained notation. In ordinary language, A stochastic process is a rule that gives one random output for each allowed time; a path is what one realization draws across all those times. This sentence identifies the observable idea before any probability space, covariance operator, kernel, or integral is introduced.
Read the symbols aloud before manipulating them: X_t means the random variable selected at time t; X_t(omega) is its value in outcome omega; t is not random here. A superscript H is a parameter label, not an exponent applied to a random variable. E means average over repeated realizations under the declared law. Equality may mean equality of numbers, random variables, finite-dimensional laws, processes in law, or sample paths; the chapter states which one is intended.
The formal definition used here is: A real stochastic process on T is a jointly indexed family X=(X_t)_{t in T} of measurable real random variables on one probability space. A definition names an object and the data needed to identify it. It does not assert continuity, differentiability, independence, a stochastic integral, a causal mechanism, or suitability for real data unless those properties are proved separately.
The bounded theorem target is: Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists. Every hypothesis belongs to the conclusion. The parameter range, time domain, centering convention, covariance normalization, version of the process, integrability order, topology, and endpoint exclusions remain visible when the result is reused.
The proof route has four ledgers. The algebra ledger checks identities and powers. The covariance ledger checks positive semidefiniteness and every pairwise expectation. The path ledger states whether a claim concerns one realization, almost every realization, or only finite-dimensional distributions. The limit ledger records the approximation and mode of convergence. No conclusion crosses from one ledger to another without a stated theorem.
The finite model is deliberately small: Let (X_0,X_1,X_2)=(Z,2Z,-Z) for one standard Gaussian Z; compute one outcome when Z=3 and list the three marginal variances. Keep exact fractions, powers, units, covariance entries, and the order of increments. This finite computation does not prove an infinite-dimensional theorem, but it exposes a wrong sign, missing factor, confused normalization, or unsupported independence claim before those errors are hidden by notation.
Two independent routes must reach the checkpoint The path values are (3,6,-3), while the marginal variances are (1,4,1). One route expands the declared formula directly. The other uses a covariance matrix, an increment identity, a scaling table, or a conditional projection. Agreement is evidence that the finite ledger is internally consistent; disagreement means the chapter stops and repairs the first failed line.
The required failure mutation is: Choosing a fresh unrelated probability space for every t does not yet provide a common sample path. A mutation is not an afterthought. It identifies the exact hypothesis that makes the theorem work and prevents a familiar Brownian-motion rule from being carried into fractional Gaussian noise merely because the symbols look similar.
Keep construction, calculation, and interpretation separate. A covariance formula can construct a Gaussian law after positive semidefiniteness is established. A sample-path theorem can then describe regularity for a chosen version. Neither step shows that measured observations were generated by this law, that H is identifiable from a short record, or that persistent correlation is a causal mechanism.
Close by reconstruction rather than recognition. Without looking back, restate the question in plain language, decode every symbol, rebuild the finite ledger twice, state the theorem with all hypotheses, activate the mutation, and report the strongest surviving conclusion. Recognizing the words “Hurst,” “self-similar,” or “long memory” is not yet understanding.
Maintain a local symbol dictionary beside the page. For each symbol in “X_t means the random variable selected at time t; X_t(omega) is its value in outcome omega; t is not random here.,” record its type, units, permitted range, dependence on H or time, and whether it is observed, deterministic, or random. Rebuild the displayed formula once with words replacing symbols. This prevents an exponent, covariance, random variable, and probability law from being treated as interchangeable objects.
Use an equality ladder whenever two expressions are compared: exact algebraic equality; equality almost surely for fixed times; indistinguishability of paths; equality of finite-dimensional laws; equality of path-space laws; convergence in L2; convergence in probability; or convergence in distribution. The step justified in this chapter is tied to Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists.; no stronger rung is silently substituted.
Bounded theorem: Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists. No endpoint, estimation, simulation, application, or causal claim is included unless explicitly stated.
Type every object in A real stochastic process on T is a jointly indexed family X=(X_t)_{t in T} of measurable real random variables on one probability space. Mark deterministic times, random variables, vectors, kernels, Hilbert-space elements, and probability laws separately. State whether the process starts at zero and whether time is one-sided or two-sided.
Decode X_t means the random variable selected at time t; X_t(omega) is its value in outcome omega; t is not random here. Write one concrete numerical instance before using the general formula. Check that both sides have the same physical units and that a variance is never mistaken for a standard deviation.
Construct the finite covariance or increment table required by Let (X_0,X_1,X_2)=(Z,2Z,-Z) for one standard Gaussian Z; compute one outcome when Z=3 and list the three marginal variances. Verify symmetry, nonnegative diagonal entries, and each off-diagonal entry. If a Gaussian vector is claimed to exist, verify nonnegative quadratic forms rather than inspecting entries one by one.
Derive the decisive identity behind Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists. Expand expectations line by line, cancel only matching terms, preserve every factor one-half, and state where stationarity, self-similarity, Gaussianity, or continuity is actually used.
Connect the finite identity to the theorem through the appropriate result: Gaussian laws determined by mean and covariance, Kolmogorov continuity, a variation criterion, a Hilbert-space isometry, Young integration, rough-path continuity, or Malliavin duality. Naming the result without checking its hypotheses is not a proof.
Recompute the checkpoint independently and obtain The path values are (3,6,-3), while the marginal variances are (1,4,1). Record exact arithmetic and any rounding interval. If a simulation is later added, it must be labeled as a random illustration and cannot replace this exact calculation.
Activate Choosing a fresh unrelated probability space for every t does not yet provide a common sample path. Locate the first invalid equality or theorem hypothesis. The argument proves only Every finite time list produces a random vector, and statements about the process law must agree on overlapping finite lists. It does not prove every endpoint, arbitrary Gaussian processes, non-Gaussian long memory, statistical identifiability, causal interpretation, or domain validity.
Reconstruct this finite ledger exactly: Let (X_0,X_1,X_2)=(Z,2Z,-Z) for one standard Gaussian Z; compute one outcome when Z=3 and list the three marginal variances. Give two independent calculations and retain the interpretation of every number.
- Copy the exact finite input from Let (X_0,X_1,X_2)=(Z,2Z,-Z) for one standard Gaussian Z; compute one outcome when Z=3 and list the three marginal variances. Label times, increments, covariance entries, exponents, and units before calculating.
- Apply the definition once, without shortcuts. Expand every variance or covariance as an expectation of a product and keep the order of the two increments visible.
- Simplify powers and fractions exactly. For a covariance matrix, check a quadratic form or eigenvalue certificate; for scaling, compare ratios at two declared scales.
- Recompute through the independent route specified in the chapter—direct expansion versus matrix multiplication, or increment variance versus covariance identity.
- Both routes must give The path values are (3,6,-3), while the marginal variances are (1,4,1). Then apply the failure mutation, name the first unsupported step, and weaken the conclusion instead of silently importing a Brownian rule.
Result: Audited checkpoint: The path values are (3,6,-3), while the marginal variances are (1,4,1). This certifies the stated finite calculation only under the declared conventions.