02.10.06 · analysis / harmonic

The Bochner-Minlos theorem and characteristic functionals on nuclear spaces

shipped3 tiersLean: none

Anchor (Master): Gel'fand-Vilenkin *Generalized Functions* Vol. 4 §IV.3-4 (Minlos' theorem, cylinder measures, nuclearity); Hida *Brownian Motion* (Springer, 1980) Ch. 3 (white noise as a measure on $\mathcal{S}'$); Glimm-Jaffe *Quantum Physics: A Functional Integral Point of View* (Springer, 2nd ed. 1987) §6.1, §6.6 (Minlos and the free field); Simon *The $P(\phi)_2$ Euclidean (Quantum) Field Theory* (Princeton, 1974) §I

Intuition Beginner

A probability law on the line can be summarised by a single function. Hand the law a frequency, ask for the average of a rotating phase at that frequency, and write down the answer. As the frequency sweeps across all values you trace out one function, the characteristic function of the law. The remarkable fact is that this one function remembers everything: two laws with the same characteristic function are the same law. The characteristic function is the Fourier transform of the probability, and Fourier transforms can be undone.

Which functions arise this way? Bochner answered exactly. A function is the characteristic function of some probability law on the line if and only if it is continuous, equals one at frequency zero, and is positive-definite — a mild self-consistency condition that any genuine average of phases must satisfy. So the recipe runs both ways: every law gives a nice function, and every nice function comes from a law.

Now make the random object bigger. Instead of a single random number, imagine a random field: a random value attached to every point of space at once. You can still ask for the average of a rotating phase, but now the frequency is itself a whole test-function. The question Minlos answered is when one of these field-level functions still comes from an honest probability law.

Visual Beginner

Picture a dial. At the top sits a probability law, a cloud of likelihood. From the dial hangs one curve, the characteristic function, swept out as you turn a frequency knob. Bochner's picture says the dial and the curve are two faces of one object: turn the curve into the law by an inverse Fourier transform, turn the law into the curve by a forward one.

In the second panel the single knob is replaced by a small dashboard of knobs, one for each test-function. The field-level curve becomes a functional: feed it a smooth test-function, read off a number. The catch the picture hides is the home of the law. For a single number the law lives on the line. For a field the law refuses to live on any ordinary function space; it spreads out onto a larger space of generalised functions. The smaller and gentler the test-dashboard, the larger and roomier the home the law needs — and only when the dashboard is gentle enough in a precise sense does a home exist at all.

Worked example Beginner

Let us read off the characteristic function of the most familiar law, the standard bell curve on the line. The bell curve assigns likelihood that falls off like a Gaussian. Its characteristic function is computed by averaging a rotating phase against the bell.

Step 1. The average of the phase at frequency $t$ against a standard bell curve is itself a Gaussian in $t$ : the answer is $e^{- t^{2} /2}$ .

Step 2. Check the three Bochner conditions. At frequency zero the value is $e^{0} = 1$ , as required. The function is continuous, since the exponential of a continuous expression is continuous. And it is positive-definite, the self-consistency property every characteristic function shares.

Step 3. Read the recipe backwards. Bochner guarantees that any continuous, positive-definite function equal to one at zero is the characteristic function of a law. So $e^{- t^{2} /2}$ had to come from some law — and it does, the bell curve we started from.

Step 4. Now picture the field analogue. Replace the single frequency $t$ by a test-function $f$ , and replace $t^{2}$ by the squared size of $f$ . The functional $e^{- ∥ f ∥^{2} /2}$ is the field-level analogue of the bell curve. Minlos' achievement is the statement that this functional, too, is the characteristic functional of a genuine probability law — the law of white noise, the random field that powers Brownian motion and the free quantum field.

What this shows: the Gaussian recipe survives the jump from one number to a whole field, and the resulting law is the cornerstone of probability in infinite dimensions.

Check your understanding Beginner

Exercise (easy, multiple choice).

What does the characteristic function of a probability law record?

A. The most likely outcome of the law B. The average of a rotating phase at each frequency — the Fourier transform of the law C. The variance of the law only D. A list of the law's values at each point

Hint

The unit built it as a Fourier transform that can be undone, recovering the whole law.

Answer

B. The average of a rotating phase at each frequency — the Fourier transform of the law.

Feedback-correct: yes. Because it is a Fourier transform, it can be inverted, so it determines the law completely. Feedback-wrong: option A gives only the mode, option C only one moment, and option D describes a function rather than a law; none of these pins the law down the way the characteristic function does.

Formal definition Intermediate+

Let $Φ$ be a nuclear space 02.14.04 — for concreteness the Schwartz space $S (R^{n})$ — with topological dual $Φ^{'}$ , the space of tempered distributions $S^{'} (R^{n})$ . The pairing of $ξ \in Φ^{'}$ with $f \in Φ$ is written $⟨ ξ, f ⟩$ .

Definition (characteristic functional). Let $μ$ be a probability measure on $Φ^{'}$ (equipped with the $σ$ -algebra generated by the cylinder sets, defined below). Its characteristic functional is $$ L_\mu(f) = \int_{\Phi'} e^{i\langle \xi, f\rangle}, d\mu(\xi), \qquad f \in \Phi. $$ It is the infinite-dimensional analogue of the characteristic function: for fixed $f$ the integrand is the rotating phase $e^{i ⟨ ξ, f ⟩}$ , and $L_{μ} (f)$ is its $μ$ -average.

Definition (positive-definite functional). A functional $L : Φ \to C$ is positive-definite if for every finite collection $f_{1}, \dots, f_{m} \in Φ$ and complex scalars $c_{1}, \dots, c_{m}$ , $$ \sum_{j,k=1}^m c_j \bar c_k, L(f_j - f_k) \ge 0. $$ Equivalently, every matrix $[L (f_{j} - f_{k})]_{j, k}$ is positive semidefinite.

Definition (cylinder set and cylinder measure). A cylinder set in $Φ^{'}$ is a set of the form $$ C = {\xi \in \Phi' : (\langle\xi, f_1\rangle,\dots,\langle\xi, f_n\rangle) \in B}, $$ for some $f_{1}, \dots, f_{n} \in Φ$ and Borel $B \subseteq R^{n}$ . A cylinder measure assigns consistent finite-dimensional probability laws to the projections $ξ \mapsto (⟨ ξ, f_{1} ⟩, \dots, ⟨ ξ, f_{n} ⟩)$ , but is not assumed countably additive on the full cylinder $σ$ -algebra a priori. The central question is when a cylinder measure is in fact a genuine (countably additive) probability measure on $Φ^{'}$ .

Notation and conventions

$Φ \subset H \subset Φ^{'}$ : a Gel'fand triple 02.14.05; $H$ a Hilbert space, the inclusions continuous with dense range.
$L (0) = 1$ is the normalisation forcing $μ$ to be a probability (total mass one) rather than a general finite measure.
"Continuous" for $L$ means continuous for the nuclear topology on $Φ$ ; this is the hypothesis that does the work.

The finite-dimensional case: Bochner's theorem

For $Φ = R^{n}$ the construction reduces to the classical statement.

Theorem (Bochner; Bochner 1932 ^{[Bochner 1932]}). A function $L : R^{n} \to C$ is the characteristic function $L (t) = \int e^{i t \cdot x} d μ (x)$ of a finite positive measure $μ$ on $R^{n}$ if and only if $L$ is continuous and positive-definite. The measure is a probability measure precisely when $L (0) = 1$ .

This is covered as the Originator material of 02.10.04. The Bochner-Minlos theorem is the extension of this equivalence from $R^{n}$ to an infinite-dimensional nuclear $Φ$ , with the measure relocated to the dual $Φ^{'}$ .

Counterexamples to common slips

Positive-definiteness alone is not enough at the field level: a continuous positive-definite functional on an infinite-dimensional Hilbert space need not be a characteristic functional. The Gaussian $e^{- ∥ f ∥^{2} /2}$ on a separable Hilbert space is positive-definite and continuous, yet defines only a cylinder measure that is not countably additive on the Hilbert space itself. Countable additivity is recovered only after enlarging the home to a space where the embedding is Hilbert-Schmidt — exactly what nuclearity supplies.
The measure does not live on $Φ$ or on $H$ ; it lives on the strictly larger $Φ^{'}$ . White noise paths are distributions, not functions.
"Characteristic functional" is the Fourier transform of a measure on $Φ^{'}$ , not on $Φ$ ; the roles of space and dual are swapped relative to the naive finite-dimensional reading.

Key theorem with proof Intermediate+

Theorem (Bochner-Minlos; Minlos 1959; Gel'fand-Vilenkin Vol. 4 §IV.3 ^{[Minlos 1959]}). Let $Φ$ be a nuclear space and $L : Φ \to C$ a functional. Then $L$ is the characteristic functional of a unique probability measure $μ$ on the dual $Φ^{'}$ if and only if

$L$ is continuous on $Φ$ ,
$L$ is positive-definite, and
$L (0) = 1$ .

Proof. (Necessity.) If $μ$ is a probability measure on $Φ^{'}$ and $L = L_{μ}$ , then $L (0) = μ (Φ^{'}) = 1$ ; positive-definiteness is the identity $$ \sum_{j,k} c_j\bar c_k L(f_j - f_k) = \int_{\Phi'} \Big|\sum_j c_j e^{i\langle\xi, f_j\rangle}\Big|^2 d\mu(\xi) \ge 0, $$ and continuity follows because $f \mapsto e^{i ⟨ ξ, f ⟩}$ is continuous and bounded, so dominated convergence transfers continuity through the integral.

(Sufficiency — the content.) Fix any finite tuple $f_{1}, \dots, f_{n} \in Φ$ and consider the map $T : R^{n} \to C$ , $T (t) = L (\sum_{j} t_{j} f_{j})$ . Linearity of $f \mapsto \sum t_{j} f_{j}$ shows $T$ is continuous and positive-definite on $R^{n}$ with $T (0) = 1$ , so classical Bochner gives a probability measure $μ_{f_{1}, \dots, f_{n}}$ on $R^{n}$ whose characteristic function is $T$ . These finite-dimensional laws form a projective (consistent) family indexed by finite subsets of $Φ$ : marginalising one coordinate corresponds to setting one $t_{j} = 0$ , and the matching of characteristic functions makes the consistency conditions automatic. By Kolmogorov's extension argument this family defines a cylinder measure on $Φ^{'}$ with the prescribed projections, and $L$ is its characteristic functional on each cylinder.

The one missing step is countable additivity of this cylinder measure on the full cylinder $σ$ -algebra of $Φ^{'}$ . This is where nuclearity is indispensable. Continuity of $L$ at $0$ on the nuclear $Φ$ gives, for each $ε$ , a continuous Hilbertian seminorm $p$ with $∣1 - L (f) ∣ < ε$ whenever $p (f) \leq 1$ . Nuclearity means the inclusion of the completion $Φ_{p}$ into a coarser Hilbertian completion $Φ_{q}$ is Hilbert-Schmidt. The Hilbert-Schmidt property forces the finite-dimensional laws to be tight: their mass does not escape to infinity as the dimension grows, because the Hilbert-Schmidt norm controls $\sum_{k} E [⟨ ξ, e_{k} ⟩^{2}]$ along an orthonormal system ${e_{k}}$ and keeps it summable. Tightness of the projective family is exactly Prokhorov's criterion for the cylinder measure to extend to a genuine countably additive probability measure on $Φ^{'}$ . Uniqueness is inherited from the finite-dimensional case: two measures with the same characteristic functional have the same projections, hence agree on every cylinder, hence on the generated $σ$ -algebra. $□$

Bridge. The Bochner-Minlos theorem builds toward the entire measure-theoretic foundation of constructive quantum field theory and appears again in 08.14.01, where Wiener measure is the characteristic-functional construction applied to Brownian paths. This is exactly the infinite-dimensional shadow of classical Bochner: the foundational reason a random field exists at all is that its characteristic functional is continuous on a nuclear test-space, and nuclearity is the central insight that converts a merely finitely-consistent family of laws into a single countably additive measure on $Φ^{'}$ . The theorem generalises the $R^{n}$ equivalence "positive-definite continuous function $\leftrightarrow$ finite measure" to the nuclear pair $Φ \subset Φ^{'}$ , and the Hilbert-Schmidt tightness mechanism is dual to the nuclearity that powers the Schwartz kernel theorem of 02.14.04. Putting these together, the bridge is the recognition that the same Grothendieck nuclearity which collapses separately continuous bilinear forms to kernels also confines an infinite-dimensional Gaussian's mass onto $S^{'}$ — one structural hypothesis, two cornerstone theorems.

Exercises Intermediate+

Exercise (medium, symbolic).

Let $C$ be a continuous positive operator on the test-space and set $L (f) = e^{- \frac{1}{2} ⟨ C f, f ⟩}$ . Show that $L$ is positive-definite.

Hint

A Gaussian functional is positive-definite because it is itself a limit/average of the rank-one positive-definite functions $e^{i ⟨ ξ, \cdot ⟩}$ ; alternatively, $e^{- \frac{1}{2} ⟨ C (f - g), (f - g)⟩}$ is the characteristic function of a finite-dimensional Gaussian once you restrict to the span of the $f_{j}$ .

Answer

Restrict to the finite-dimensional span $V = span (f_{1}, \dots, f_{m})$ . On $V$ the quadratic form $⟨ C f, f ⟩$ is a positive semidefinite quadratic form, so $f \mapsto e^{- \frac{1}{2} ⟨ C f, f ⟩}$ is the characteristic function of a (possibly degenerate) centred Gaussian measure on $V^{*}$ by classical Bochner. Every characteristic function is positive-definite by the necessity computation $\sum_{j, k} c_{j} \overset{c}{ˉ}_{k} L (f_{j} - f_{k}) = \int ∣ \sum_{j} c_{j} e^{i ⟨ ξ, f_{j} ⟩} ∣^{2} d μ \geq 0$ . Since the $f_{j}$ were arbitrary, $L$ is positive-definite on all of $Φ$ .

Exercise (medium, short answer).

Explain why the Gaussian functional $e^{- ∥ f ∥^{2} /2}$ on a separable infinite-dimensional Hilbert space $H$ fails to define a countably additive measure on $H$ itself, while it succeeds on the dual of a nuclear space.

Hint

Compute the would-be expected squared norm $\sum_{k} E [⟨ ξ, e_{k} ⟩^{2}]$ along an orthonormal basis.

Answer

For the standard Gaussian cylinder measure each coordinate $⟨ ξ, e_{k} ⟩$ is a standard normal, so $E [⟨ ξ, e_{k} ⟩^{2}] = 1$ for every $k$ and $\sum_{k} E [⟨ ξ, e_{k} ⟩^{2}] = \infty$ . A countably additive measure on $H$ would force $∥ ξ ∥^{2} = \sum_{k} ⟨ ξ, e_{k} ⟩^{2} < \infty$ almost surely, contradicting the divergent expectation; the mass "escapes to infinity in norm." On the dual of a nuclear $Φ$ the embedding $Φ_{p} ↪ Φ_{q}$ is Hilbert-Schmidt, so the relevant sum is taken with summable Hilbert-Schmidt weights $\sum_{k} λ_{k}^{2} < \infty$ , and the field sits in the larger $Φ^{'}$ where the weighted squared norm is finite. Nuclearity is exactly the mechanism that restores tightness.

Exercise (hard, short answer).

Identify the characteristic functional of the free Euclidean field of mass $m$ on $R^{d}$ and state which operator plays the role of the covariance.

Hint

The free-field covariance is the Green's function of $- Δ + m^{2}$ ; the functional is Gaussian.

Answer

The free Euclidean field of mass $m$ is the centred Gaussian measure $μ_{C}$ on $S^{'} (R^{d})$ with characteristic functional $$ L(f) = \exp!\Big(-\tfrac12 \langle f, C f\rangle\Big), \qquad C = (-\Delta + m^2)^{-1}. $$ The covariance operator $C$ is the resolvent of the Laplacian, whose kernel is the Euclidean (Yukawa) propagator. Continuity of $L$ on $S$ holds because $C$ maps $S$ continuously into itself for $m > 0$ , and positive-definiteness holds because $C$ is positive; Bochner-Minlos then delivers the measure on $S^{'}$ . This measure is the starting point of constructive (Euclidean) quantum field theory.

Exercise (hard, short answer).

Where in the proof of Bochner-Minlos is nuclearity used, and what would go wrong without it?

Hint

Trace the logic: Bochner gives finite-dimensional laws, Kolmogorov glues them into a cylinder measure, and one final property is needed.

Answer

Nuclearity is used only at the last step: promoting the consistent cylinder measure to a countably additive measure on $Φ^{'}$ . Continuity of $L$ supplies a Hilbertian seminorm controlling $L$ near $0$ , and nuclearity makes the inclusion of the corresponding Hilbert completions Hilbert-Schmidt. The Hilbert-Schmidt norm bounds the total expected squared size of the field along an orthonormal system, which is precisely the tightness (Prokhorov) condition for the projective family. Without nuclearity the family can fail to be tight — as on an infinite-dimensional Hilbert space — and then the cylinder measure is only finitely additive, with no genuine probability measure on the space behind it.

Advanced results Master

The Bochner-Minlos theorem is the entry point to a circle of constructions. Three deserve statement.

White noise and Hida calculus. Taking $Φ = S (R)$ and $L (f) = exp (- \frac{1}{2} ∥ f ∥_{L^{2}}^{2})$ produces the standard Gaussian measure on $S^{'} (R)$ , white noise. Hida's white-noise analysis ^{[Hida 1980]} builds a full differential calculus — Hida derivatives, the Wick product, the $S$ -transform — on the $L^{2}$ space of this measure, realising it as a Gel'fand triple $(S) \subset L^{2} (S^{'}, μ) \subset (S)^{'}$ of test and generalised Brownian functionals. The Brownian path of 08.14.01 is the indefinite integral $B_{t} = ⟨ ξ, 1_{[0, t]} ⟩$ of white noise, so Wiener measure is recovered as a pushforward of the Bochner-Minlos measure.

The free Euclidean field. With $C = (- Δ + m^{2})^{- 1}$ the functional $exp (- \frac{1}{2} ⟨ f, C f ⟩)$ gives the Gaussian free field on $S^{'} (R^{d})$ . Its moments are the Euclidean Wightman/Schwinger functions of a free scalar field, satisfying the Osterwalder-Schrader axioms; reflection positivity of $C$ is what allows analytic continuation back to a relativistic quantum field. This is the measure-theoretic backbone on which the $P (ϕ)_{2}$ and $ϕ_{3}^{4}$ interacting theories are built as absolutely continuous (or singular limit) perturbations ^{[Glimm-Jaffe]}.

Lévy and non-Gaussian fields. Replacing the Gaussian exponent by a Lévy-Khinchine exponent $ψ$ — itself the logarithm of a positive-definite functional — yields generalised random fields with independent increments, the field analogue of Lévy processes. The Bochner-Minlos theorem certifies their existence on $S^{'}$ from the single datum of a continuous positive-definite functional, which is why the characteristic-functional method is the universal existence tool for random distributions.

Synthesis. The Bochner-Minlos theorem is the foundational reason infinite-dimensional probability has a home: it is the unique bridge that turns a continuous positive-definite functional on a nuclear test-space into a countably additive law on the dual, and this is exactly the local-to-global pattern that recurs across the corpus — finite-dimensional Bochner laws are glued by consistency, then nuclearity confines their joint mass to $Φ^{'}$ . The central insight is that nuclearity is dual to the Hilbert-Schmidt tightness which prevents an infinite-dimensional Gaussian's mass from escaping to infinity, and this is exactly the same Grothendieck nuclearity that powers the Schwartz kernel theorem of 02.14.04; putting these together, white noise, Wiener measure, and the free Euclidean field are three faces of one construction. The theorem generalises classical Bochner from $R^{n}$ to $Φ \subset Φ^{'}$ , and the bridge to physics is that the free-field measure built here is the measure constructive QFT perturbs — so the structure assembled in this unit appears again in 08.14.01 as Wiener measure and is the silent foundation under every Euclidean functional integral.

Full proof set Master

Proposition (characteristic functionals separate measures). Two probability measures $μ, ν$ on $Φ^{'}$ with $L_{μ} = L_{ν}$ are equal.

Proof. For any finite tuple $f_{1}, \dots, f_{n} \in Φ$ the pushforward of $μ$ under the projection $π (ξ) = (⟨ ξ, f_{1} ⟩, \dots, ⟨ ξ, f_{n} ⟩)$ has characteristic function $t \mapsto L_{μ} (\sum_{j} t_{j} f_{j})$ , and likewise for $ν$ . Equality of $L_{μ}, L_{ν}$ forces equal finite-dimensional characteristic functions, so by classical Bochner uniqueness on $R^{n}$ the pushforwards agree. Thus $μ, ν$ agree on every cylinder set. The cylinder sets form a $π$ -system generating the cylinder $σ$ -algebra, so by the $π$ - $λ$ theorem $μ = ν$ . $□$

Proposition (necessity of positive-definiteness). If $L = L_{μ}$ for a probability measure $μ$ on $Φ^{'}$ , then $L$ is positive-definite.

Proof. For $f_{1}, \dots, f_{m} \in Φ$ and $c_{1}, \dots, c_{m} \in C$ , $$ \sum_{j,k=1}^m c_j\bar c_k, L(f_j - f_k) = \int_{\Phi'} \sum_{j,k} c_j\bar c_k, e^{i\langle\xi, f_j - f_k\rangle},d\mu(\xi) = \int_{\Phi'}\Big|\sum_{j=1}^m c_j e^{i\langle\xi,f_j\rangle}\Big|^2 d\mu(\xi), $$ using $e^{i ⟨ ξ, f_{j} - f_{k} ⟩} = e^{i ⟨ ξ, f_{j} ⟩} \overline{e^{i ⟨ ξ, f_{k} ⟩}}$ and the product measure structure of 02.07.07 to justify interchanging the finite sum and the integral. The integrand is a squared modulus, hence non-negative, so the integral is non-negative. $□$

Proposition (Gaussian functional yields a Gaussian measure). For a continuous positive symmetric bilinear form $B$ on $Φ$ , the functional $L (f) = e^{- \frac{1}{2} B (f, f)}$ is the characteristic functional of a unique centred Gaussian measure on $Φ^{'}$ , provided $Φ$ is nuclear and $B$ is continuous.

Proof. Continuity of $B$ gives continuity of $L$ ; restricting $B$ to any finite-dimensional subspace shows each $T (t) = L (\sum t_{j} f_{j}) = e^{- \frac{1}{2} B (\sum t_{j} f_{j}, \sum t_{k} f_{k})}$ is a Gaussian characteristic function on $R^{n}$ , hence continuous and positive-definite with $T (0) = 1$ , so $L$ is positive-definite with $L (0) = 1$ . Bochner-Minlos then supplies a unique probability measure $μ$ on $Φ^{'}$ with $L_{μ} = L$ . Its finite-dimensional projections are centred Gaussians, so $μ$ is by definition a centred Gaussian measure; the covariance is the operator induced by $B$ . $□$

Proposition (tightness from Hilbert-Schmidt embedding). If continuity of $L$ at $0$ is witnessed by a Hilbertian seminorm $p$ whose canonical map $Φ_{p} \to Φ_{q}$ into a coarser completion is Hilbert-Schmidt, then the cylinder measure determined by $L$ is tight, hence countably additive.

Proof. Let ${e_{k}}$ be an orthonormal basis of $Φ_{q}$ realised from the singular system of the Hilbert-Schmidt map, with singular values $λ_{k}$ satisfying $\sum_{k} λ_{k}^{2} < \infty$ . The second-moment estimate from continuity of $L$ gives $E [⟨ ξ, e_{k} ⟩^{2}] \leq c λ_{k}^{2}$ for a constant $c$ depending on the modulus of continuity of $L$ at $0$ . Hence $E [\sum_{k} ⟨ ξ, e_{k} ⟩^{2}] \leq c \sum_{k} λ_{k}^{2} < \infty$ , so the field lies almost surely in the Hilbert space $Φ_{q}^{'}$ of finite weighted norm. Bounded second moment on a Hilbert space gives tightness of the finite-dimensional marginals by Chebyshev plus the compactness of balls of the trace-class covariance, which is Prokhorov's criterion; the cylinder measure therefore extends to a countably additive measure on $Φ^{'}$ . $□$

Connections Master

Fourier transform and Bochner's theorem on $R^{n}$ 02.10.04. The classical Bochner theorem stated there is the finite-dimensional case: positive-definite continuous functions on $R^{n}$ are Fourier transforms of finite measures. This unit is the verbatim infinite-dimensional extension, with the measure relocated from $R^{n}$ to the dual $Φ^{'}$ of a nuclear test-space; the proof literally invokes the $R^{n}$ statement on every finite-dimensional projection.
Distributions, nuclear spaces, and the Schwartz kernel theorem 02.14.04. Nuclearity is the shared load-bearing hypothesis. There it collapses separately continuous bilinear forms to a single kernel on the product; here the same Hilbert-Schmidt embedding chain $Φ ↪ H ↪ Φ^{'}$ confines an infinite-dimensional Gaussian's mass onto $Φ^{'} = S^{'}$ . Without nuclearity both theorems fail, which is why Grothendieck's notion is the quiet engine of both.
The rigged Hilbert space (Gel'fand triple) and the nuclear spectral theorem 02.14.05. The triple $Φ \subset H \subset Φ^{'}$ that hosts the Bochner-Minlos measure is the same rigging that carries generalised eigenfunctions in the nuclear spectral theorem. White noise lives in $L^{2}$ of a measure on the rigged dual, and Hida's calculus erects a second Gel'fand triple of Brownian functionals on top of it, so the rigging appears twice: once for the field, once for functionals of the field.
Brownian motion, the Wiener measure, and the path integral 08.14.01. Wiener measure is the Bochner-Minlos construction applied with the white-noise functional $exp (- \frac{1}{2} ∥ f ∥_{2}^{2})$ : Brownian motion is the time-integral of white noise, and Wiener measure on path space is the pushforward of the Gaussian measure on $S^{'}$ . This unit supplies the existence theorem that the path-integral unit assumes when it places a measure on continuous paths.
Fubini-Tonelli and product measures 02.07.07. The interchange of the finite sum and the integral in the positive-definiteness computation, and the consistency of the projective family of finite-dimensional laws, both rest on the product-measure machinery; the projective family is the infinite-dimensional analogue of an infinite product of marginals stitched by Kolmogorov's theorem.

Historical & philosophical context Master

The characteristic-function method for probability laws on $R^{n}$ was perfected by Salomon Bochner in Vorlesungen über Fouriersche Integrale (Leipzig, 1932) ^{[Bochner 1932]}, where positive-definiteness was isolated as the exact condition for a continuous function to be the Fourier transform of a positive measure. The infinite-dimensional question — when does a positive-definite functional on a function space come from a measure? — was open until Robert Adol'fovich Minlos resolved it for nuclear spaces in his 1959 paper Generalized random processes and their extension to a measure (Trudy Moskov. Mat. Obšč. 8, 497-518) ^{[Minlos 1959]}, written in the orbit of Gel'fand's seminar. The theorem became the centrepiece of Volume 4 of Gel'fand and Vilenkin's Generalized Functions (Academic Press, 1964), which framed it as the natural marriage of distribution theory and probability: generalised functions are precisely the right sample space for generalised random processes.

The philosophical payoff is the dissolution of a paradox that had troubled the heuristic path-integral. Physicists wrote $\int e^{- S [ϕ]} D ϕ$ as if a uniform "flat" measure $D ϕ$ existed on a space of fields; no such measure exists, because translation-invariant measures on infinite-dimensional spaces cannot be countably additive. Bochner-Minlos replaces the illusion with a precise object: the Gaussian free-field measure on $S^{'}$ , defined not by a density against a non-existent flat measure but directly through its characteristic functional. Edward Nelson, James Glimm, and Arthur Jaffe ^{[Glimm-Jaffe]} made this the foundation of constructive quantum field theory, and Takeyuki Hida ^{[Hida 1980]} turned the white-noise instance into an autonomous calculus. The lesson the theorem teaches is that the home of a random field is never the space of test-functions but its dual — fields are distributions — and that nuclearity is the exact geometric condition under which infinitely many independent degrees of freedom can still be assembled into a single probability law.

Bibliography Master

@book{Bochner1932Fourier,
  author    = {Bochner, Salomon},
  title     = {Vorlesungen {\"u}ber Fouriersche Integrale},
  publisher = {Akademische Verlagsgesellschaft},
  address   = {Leipzig},
  year      = {1932}
}

@article{Minlos1959Generalized,
  author  = {Minlos, R. A.},
  title   = {Generalized random processes and their extension to a measure},
  journal = {Trudy Moskov. Mat. Obshch.},
  volume  = {8},
  pages   = {497--518},
  year    = {1959}
}

@book{GelfandVilenkinVol4,
  author    = {Gel'fand, I. M. and Vilenkin, N. Ya.},
  title     = {Generalized Functions, Vol. 4: Applications of Harmonic Analysis},
  publisher = {Academic Press},
  address   = {New York},
  year      = {1964}
}

@book{Hida1980Brownian,
  author    = {Hida, Takeyuki},
  title     = {Brownian Motion},
  publisher = {Springer-Verlag},
  series    = {Applications of Mathematics},
  volume    = {11},
  year      = {1980}
}

@book{GlimmJaffe1987,
  author    = {Glimm, James and Jaffe, Arthur},
  title     = {Quantum Physics: A Functional Integral Point of View},
  publisher = {Springer-Verlag},
  edition   = {2},
  year      = {1987}
}

@book{ReedSimonII,
  author    = {Reed, Michael and Simon, Barry},
  title     = {Methods of Modern Mathematical Physics, Vol. II: Fourier Analysis, Self-Adjointness},
  publisher = {Academic Press},
  address   = {New York},
  year      = {1975}
}

@book{Simon1974Pphi2,
  author    = {Simon, Barry},
  title     = {The $P(\phi)_2$ Euclidean (Quantum) Field Theory},
  publisher = {Princeton University Press},
  series    = {Princeton Series in Physics},
  year      = {1974}
}

Prerequisites

02.10.04
02.14.04
02.07.07
08.14.01

Tier anchors

beginner: the idea that a probability law is completely pinned down by one function — its Fourier transform, the 'characteristic function' — and that this still works when the random object is a whole field rather than a single number, provided the test-space is small enough
intermediate: Gel'fand-Vilenkin *Generalized Functions* Vol. 4 (Academic Press, 1964) Ch. IV (positive-definite functionals, the Bochner-Minlos theorem, Gaussian measures on $\mathcal{S}'$); Reed-Simon *Methods of Modern Mathematical Physics* Vol. II (Academic Press, 1975) §IX.5 (the free Euclidean field as a measure on $\mathcal{S}'$)
master: Gel'fand-Vilenkin *Generalized Functions* Vol. 4 §IV.3-4 (Minlos' theorem, cylinder measures, nuclearity); Hida *Brownian Motion* (Springer, 1980) Ch. 3 (white noise as a measure on $\mathcal{S}'$); Glimm-Jaffe *Quantum Physics: A Functional Integral Point of View* (Springer, 2nd ed. 1987) §6.1, §6.6 (Minlos and the free field); Simon *The $P(\phi)_2$ Euclidean (Quantum) Field Theory* (Princeton, 1974) §I

References

Gel'fand-Vilenkin — Generalized Functions, Vol. 4: Applications of Harmonic Analysis · Academic Press 1964; Ch. IV §3 (positive-definite functionals on nuclear spaces, the Minlos theorem), §4 (Gaussian measures, the construction of measures on $\mathcal{S}'$)
Minlos — Generalized random processes and their extension to a measure · Trudy Moskov. Mat. Obšč. 8 (1959) 497-518; the original statement and proof of the nuclear-space extension of Bochner's theorem
Bochner — Vorlesungen über Fouriersche Integrale · Akademische Verlagsgesellschaft, Leipzig 1932; the theorem characterising continuous positive-definite functions on $\mathbb{R}^n$ as Fourier transforms of finite positive measures
Hida — Brownian Motion · Springer 1980; Ch. 3 (white noise as the standard Gaussian measure on $\mathcal{S}'(\mathbb{R})$, the characteristic functional $\exp(-\tfrac12\|f\|_2^2)$)
Glimm-Jaffe — Quantum Physics: A Functional Integral Point of View · Springer, 2nd ed. 1987; §6.1, §6.6 (the Minlos theorem and the free Euclidean field measure on $\mathcal{S}'$)
Reed-Simon — Methods of Modern Mathematical Physics, Vol. II: Fourier Analysis, Self-Adjointness · Academic Press 1975; §IX.5 (Bochner-Minlos and the free Markov field)

Estimated time

beginner: 18m
intermediate: 46m
master: 82m