Product Measures, Lᵖ Spaces, and the Bridge to Probability
Fubini for iterated integrals, the geometry of \(L^p\), and the punchline: probability IS measure theory.
Leads to: Phase 7 opens by declaring \((\Omega,\mathcal{F},\mathbb{P})\) a measure space; this lesson builds the bridge.
Learning Objectives
Click a status chip to cycle: Not started → In progress → Studied → Practiced → Needs review → Mastered.
- State the product measure construction and Fubini–Tonelli for iterated integrals.
- Define the \(L^p\) norm and prove/interpret Hölder and Minkowski inequalities.
- Explain why \(L^p\) is complete (Riesz–Fischer) and why \(L^2\) is special.
- Translate every measure-theoretic object into its probability counterpart.
- Compute an \(L^p\) norm and verify expectation equals an integral against \(\mathbb{P}\).
Key Vocabulary
- Product measure
- On \(X\times Y\), the measure with \((\mu\times\nu)(A\times B)=\mu(A)\nu(B)\); underlies joint distributions and independence.
- Fubini–Tonelli
- Iterated integrals equal the double integral: for \(f\ge0\) (Tonelli) or \(f\in L^1\) (Fubini), \(\int\!\!\int f\,d\mu\,d\nu=\int\!\!\int f\,d\nu\,d\mu\).
- L^p space
- Measurable \(f\) with \(\lVert f\rVert_p=(\int|f|^p d\mu)^{1/p}\lt \infty\), functions identified when equal a.e.
- Hölder's inequality
- \(\int|fg|\le\lVert f\rVert_p\lVert g\rVert_q\) for \(1/p+1/q=1\); source of Cauchy–Schwarz at \(p=q=2\).
- Riesz–Fischer
- \(L^p\) is a complete normed space (Banach); \(L^2\) is a Hilbert space with inner product \(\langle f,g\rangle=\int fg\).
- Probability space
- A measure space \((\Omega,\mathcal{F},\mathbb{P})\) with \(\mathbb{P}(\Omega)=1\); expectation is the integral against \(\mathbb{P}\).
Intuition & Motivation
Product measures and Fubini
In probability this is exactly why a joint density factorizes under independence and why \(\mathbb{E}[XY]\) can be computed by iterated integration.
The \(L^p\) spaces
At \(p=q=2\), Hölder becomes Cauchy–Schwarz \(|\langle f,g\rangle|\le\lVert f\rVert_2\lVert g\rVert_2\). By Riesz–Fischer, each \(L^p\) is complete; \(L^2\) additionally carries the inner product \(\langle f,g\rangle=\int fg\,d\mu\), making it a Hilbert space - the setting for projections, conditional expectation, and least squares.
The bridge: probability is measure theory
| Measure theory | Probability | Symbol |
|---|---|---|
| Measure space \((X,\mathcal{F},\mu)\) | Probability space | \((\Omega,\mathcal{F},\mathbb{P}),\ \mathbb{P}(\Omega)=1\) |
| Measurable function | Random variable | \(X:\Omega\to\mathbb{R}\) |
| Integral \(\int X\,d\mu\) | Expectation | \(\mathbb{E}[X]=\int_\Omega X\,d\mathbb{P}\) |
| \(L^2\) inner product | Covariance (centered) | \(\Cov(X,Y)=\langle X-\mathbb{E}X,\,Y-\mathbb{E}Y\rangle\) |
| Almost everywhere | Almost surely | a.e. \(\to\) a.s. |
| Product measure | Independence / joint law | \(\mathbb{P}_X\times\mathbb{P}_Y\) |
Interactive: \(L^p\) norm and expectation-as-integral
Compute an \(L^p\) norm, verify Cauchy–Schwarz, and confirm that discrete expectation equals the integral \(\sum x_i p_i\) against a probability measure.
- Swapping iterated integrals without checking Tonelli (\(f\ge0\)) or Fubini (\(f\in L^1\)) - order can matter otherwise.
- Treating \(L^p\) elements as functions rather than a.e.-equivalence classes.
- Using Hölder with non-conjugate exponents (\(1/p+1/q\ne1\)).
- Forgetting that \(\mathbb{E}\) requires \(X\in L^1\); a heavy-tailed \(X\) may have no finite mean.
- \(L^2\) is the quant’s home: covariance is an inner product, so uncorrelated \(=\) orthogonal and regression \(=\) projection.
- Verify integrability before invoking Fubini; Tonelli on \(|f|\) is the standard pre-check.
- Read every probability statement as a measure statement - it removes the mystique and imports all of Phase 6’s theorems.
- Jensen, Markov, and Chebyshev inequalities are all measure-theoretic; you now have the tools to prove them.
Knowledge Check
Practical Exercise
Let \(X\) and \(Y\) be random variables in \(L^2(\Omega,\mathcal{F},\mathbb{P})\). Using Cauchy–Schwarz, prove \(|\Cov(X,Y)|\le\sigma_X\sigma_Y\), the correlation bound.
Center: let \(\tilde X=X-\mathbb{E}X\), \(\tilde Y=Y-\mathbb{E}Y\). Then \(\Cov(X,Y)=\mathbb{E}[\tilde X\tilde Y]=\langle\tilde X,\tilde Y\rangle\) in \(L^2\).
Cauchy–Schwarz (Hölder at \(p=q=2\)) gives \(|\langle\tilde X,\tilde Y\rangle|\le\lVert\tilde X\rVert_2\lVert\tilde Y\rVert_2\).
But \(\lVert\tilde X\rVert_2=\sqrt{\mathbb{E}[\tilde X^2]}=\sigma_X\) and likewise \(\lVert\tilde Y\rVert_2=\sigma_Y\). Hence \(|\Cov(X,Y)|\le\sigma_X\sigma_Y\), so the correlation \(\rho=\Cov/(\sigma_X\sigma_Y)\in[-1,1]\).
Lesson Summary
Retrieval Practice
Close the lesson and answer from memory before checking. This is deliberate, effortful recall - the single highest-yield study action.
A: \(\mathbb{E}[X]=\int_\Omega X\,d\mathbb{P}\); for a discrete \(X=\sum x_i\mathbf{1}_{\{\omega_i\}}\) this simple-function integral equals \(\sum_i x_i\mathbb{P}(\{\omega_i\})=\sum_i x_i p_i\).
A: It carries an inner product \(\langle f,g\rangle=\int fg\), making it a Hilbert space; covariance is that inner product, so uncorrelated means orthogonal and regression is projection.
Completion Checklist
- I can explain the core ideas in my own words
- I worked the derivations/examples by hand
- I completed the interactive workbench(es)
- I passed the knowledge check
Source References
This lesson synthesizes and paraphrases concepts from the sources below. No copyrighted text, problem sets, or solutions are reproduced. Return to the originals for full depth.
- Measure, Integration & Real Analysis (Sheldon Axler, GTM 282) current - Ch. 5–8 - Ch. 5–8: product measures, Fubini, \(L^p\) spaces, Hilbert-space \(L^2\).
- Probability: Theory and Examples (Rick Durrett, 5th ed.) current - Ch. 1 - Ch. 1: probability spaces, random variables, and expectation as an integral.