Lesson overview | Previous part | Lesson overview

Radon-Nikodym Theorem: Part 4: ML Applications to References

4. ML Applications

ML Applications develops the part of radon-nikodym theorem specified by the approved Chapter 24 table of contents. The treatment is measure-theoretic and AI-facing: every concept is tied to probability, expectation, density, or learning systems.

4.1 Importance sampling weights

Importance sampling weights belongs to the canonical scope of Radon-Nikodym Theorem. Here the point is not to repeat introductory probability, but to expose the measurable structure that makes the probability statement valid.

Working scope for this subsection: absolute continuity, singularity, Radon-Nikodym derivatives, change of measure, Lebesgue decomposition, likelihood ratios, and ML density ratios. The mathematical habit is to name the space, the sigma algebra, the measure, and the map before writing probabilities or expectations.

$$D_{\mathrm{KL}}(P\Vert Q)=\int \log\left(\frac{dP}{dQ}\right)dP\quad\text{when }P\ll Q.$$

Operational definition.

Change of measure rewrites an integral under one measure as a weighted integral under another measure.

Worked reading.

When $P\ll Q$, $\mathbb{E}_P[f]=\mathbb{E}_Q[f(dP/dQ)]$. Importance sampling is this identity estimated by samples from $Q$.

Three examples of importance sampling weights:

1. Importance-weighted validation under distribution shift.
2. KL divergence via log density ratio.
3. Off-policy policy-gradient correction.

Two non-examples clarify the boundary:

1. Using weights where the proposal misses target support.
2. Taking a likelihood ratio without naming both measures.

Proof or verification habit for importance sampling weights:

First prove the identity for indicators, extend to simple functions, then use monotone and signed integration.

set question        -> is the subset measurable?
function question   -> are inverse images measurable?
integral question   -> is the function measurable and integrable?
density question    -> is absolute continuity satisfied?
ML question         -> which measure defines the population claim?

In AI systems, importance sampling weights matters because probability language is constantly compressed into informal notation. Measure theory expands the notation so support, observability, null sets, and convergence assumptions are visible.

Density-ratio methods are everywhere in modern ML: VI, RLHF corrections, domain adaptation, off-policy evaluation, and calibration.

Practical checklist:

• Name the measurable space before naming the probability.
• Identify whether the object is a set, function, measure, distribution, or derivative of measures.
• Check whether equality is pointwise, almost everywhere, or distributional.
• Check whether limits are moved through integrals and which theorem justifies the move.
• For density ratios, check support and absolute continuity before dividing.
• For ML claims, distinguish population measure, empirical measure, model measure, and proposal measure.

Local diagnostic: State the target measure, proposal measure, and derivative.

The notebook version of this subsection uses finite spaces, step functions, empirical measures, or simple density ratios. These toy cases keep the objects visible while preserving the exact logic used in continuous ML models.

The learner should leave this subsection able to translate between the compact ML notation and the full measure-theoretic statement.

A useful way to study this subsection is to keep three layers separate:

1. Semantic layer: what real-world question is being asked?
2. Measurable layer: which event, function, or measure represents that question?
3. Computational layer: which sum, integral, sample average, or ratio estimates it?

For example, the semantic question may be whether a guardrail fails on a class of prompts. The measurable layer is an event in the prompt space. The computational layer is an empirical estimate under a validation or red-team distribution. Mixing these layers is how many probability arguments become ambiguous.

The same discipline applies to generative models. A generator is a measurable transformation of latent randomness. The generated distribution is the pushforward measure. A likelihood, density, or divergence is only meaningful after the target space, base measure, and support relation are clear.

When reading ML papers, silently expand phrases like "sample from the model," "take expectation over data," and "density ratio" into this measure-theoretic checklist. This turns informal notation into a statement that can be checked.

This is the level of precision needed for high-stakes evaluation, off-policy learning, variational inference, and theoretical generalization arguments.

A final question to ask is whether the claim would still be meaningful if the dataset were infinite, the model output lived in a function space, or the event being queried were defined by a limiting process. Measure theory is what keeps the answer honest.

4.2 KL divergence as $\int \log\frac{dP}{dQ}\,dP$

KL divergence as $\int \log\frac{dP}{dQ}\,dP$ belongs to the canonical scope of Radon-Nikodym Theorem. Here the point is not to repeat introductory probability, but to expose the measurable structure that makes the probability statement valid.

Working scope for this subsection: absolute continuity, singularity, Radon-Nikodym derivatives, change of measure, Lebesgue decomposition, likelihood ratios, and ML density ratios. The mathematical habit is to name the space, the sigma algebra, the measure, and the map before writing probabilities or expectations.

$$P\ll Q\quad\Longleftrightarrow\quad Q(A)=0\Rightarrow P(A)=0.$$

Operational definition.

Change of measure rewrites an integral under one measure as a weighted integral under another measure.

Worked reading.

When $P\ll Q$, $\mathbb{E}_P[f]=\mathbb{E}_Q[f(dP/dQ)]$. Importance sampling is this identity estimated by samples from $Q$.

Three examples of kl divergence as $\int \log\frac{dp}{dq}\,dp$:

1. Importance-weighted validation under distribution shift.
2. KL divergence via log density ratio.
3. Off-policy policy-gradient correction.

Two non-examples clarify the boundary:

1. Using weights where the proposal misses target support.
2. Taking a likelihood ratio without naming both measures.

Proof or verification habit for kl divergence as $\int \log\frac{dp}{dq}\,dp$:

First prove the identity for indicators, extend to simple functions, then use monotone and signed integration.

set question        -> is the subset measurable?
function question   -> are inverse images measurable?
integral question   -> is the function measurable and integrable?
density question    -> is absolute continuity satisfied?
ML question         -> which measure defines the population claim?

In AI systems, kl divergence as $\int \log\frac{dp}{dq}\,dp$ matters because probability language is constantly compressed into informal notation. Measure theory expands the notation so support, observability, null sets, and convergence assumptions are visible.

Density-ratio methods are everywhere in modern ML: VI, RLHF corrections, domain adaptation, off-policy evaluation, and calibration.

Practical checklist:

• Name the measurable space before naming the probability.
• Identify whether the object is a set, function, measure, distribution, or derivative of measures.
• Check whether equality is pointwise, almost everywhere, or distributional.
• Check whether limits are moved through integrals and which theorem justifies the move.
• For density ratios, check support and absolute continuity before dividing.
• For ML claims, distinguish population measure, empirical measure, model measure, and proposal measure.

Local diagnostic: State the target measure, proposal measure, and derivative.

The notebook version of this subsection uses finite spaces, step functions, empirical measures, or simple density ratios. These toy cases keep the objects visible while preserving the exact logic used in continuous ML models.

The learner should leave this subsection able to translate between the compact ML notation and the full measure-theoretic statement.

A useful way to study this subsection is to keep three layers separate:

1. Semantic layer: what real-world question is being asked?
2. Measurable layer: which event, function, or measure represents that question?
3. Computational layer: which sum, integral, sample average, or ratio estimates it?

For example, the semantic question may be whether a guardrail fails on a class of prompts. The measurable layer is an event in the prompt space. The computational layer is an empirical estimate under a validation or red-team distribution. Mixing these layers is how many probability arguments become ambiguous.

The same discipline applies to generative models. A generator is a measurable transformation of latent randomness. The generated distribution is the pushforward measure. A likelihood, density, or divergence is only meaningful after the target space, base measure, and support relation are clear.

When reading ML papers, silently expand phrases like "sample from the model," "take expectation over data," and "density ratio" into this measure-theoretic checklist. This turns informal notation into a statement that can be checked.

This is the level of precision needed for high-stakes evaluation, off-policy learning, variational inference, and theoretical generalization arguments.

A final question to ask is whether the claim would still be meaningful if the dataset were infinite, the model output lived in a function space, or the event being queried were defined by a limiting process. Measure theory is what keeps the answer honest.

4.3 Likelihood ratios in classification and density-ratio estimation

Likelihood ratios in classification and density-ratio estimation belongs to the canonical scope of Radon-Nikodym Theorem. Here the point is not to repeat introductory probability, but to expose the measurable structure that makes the probability statement valid.

Working scope for this subsection: absolute continuity, singularity, Radon-Nikodym derivatives, change of measure, Lebesgue decomposition, likelihood ratios, and ML density ratios. The mathematical habit is to name the space, the sigma algebra, the measure, and the map before writing probabilities or expectations.

$$P(A)=\int_A \frac{dP}{dQ}\,dQ.$$

Operational definition.

Absolute continuity $P\ll Q$ means $Q$-null sets are also $P$-null. Under sigma-finiteness, Radon-Nikodym gives a density $dP/dQ$.

Worked reading.

If $Q$ is a proposal distribution and $P$ is a target distribution, then $dP/dQ$ is the exact importance weight when $P\ll Q$.

Three examples of likelihood ratios in classification and density-ratio estimation:

1. Gaussian density with respect to Lebesgue measure.
2. Categorical probabilities with respect to counting measure.
3. Policy likelihood ratio in off-policy evaluation.

Two non-examples clarify the boundary:

1. A point mass treated as having Lebesgue density.
2. A target distribution with support outside the proposal support.

Proof or verification habit for likelihood ratios in classification and density-ratio estimation:

The theorem is an existence result for a measurable derivative that reconstructs one measure by integration against another.

set question        -> is the subset measurable?
function question   -> are inverse images measurable?
integral question   -> is the function measurable and integrable?
density question    -> is absolute continuity satisfied?
ML question         -> which measure defines the population claim?

In AI systems, likelihood ratios in classification and density-ratio estimation matters because probability language is constantly compressed into informal notation. Measure theory expands the notation so support, observability, null sets, and convergence assumptions are visible.

This is the rigorous foundation for densities, likelihood ratios, importance sampling, and KL divergence.

Practical checklist:

• Name the measurable space before naming the probability.
• Identify whether the object is a set, function, measure, distribution, or derivative of measures.
• Check whether equality is pointwise, almost everywhere, or distributional.
• Check whether limits are moved through integrals and which theorem justifies the move.
• For density ratios, check support and absolute continuity before dividing.
• For ML claims, distinguish population measure, empirical measure, model measure, and proposal measure.

Local diagnostic: Before dividing densities, verify the denominator measure dominates the numerator measure.

The notebook version of this subsection uses finite spaces, step functions, empirical measures, or simple density ratios. These toy cases keep the objects visible while preserving the exact logic used in continuous ML models.

The learner should leave this subsection able to translate between the compact ML notation and the full measure-theoretic statement.

A useful way to study this subsection is to keep three layers separate:

1. Semantic layer: what real-world question is being asked?
2. Measurable layer: which event, function, or measure represents that question?
3. Computational layer: which sum, integral, sample average, or ratio estimates it?

For example, the semantic question may be whether a guardrail fails on a class of prompts. The measurable layer is an event in the prompt space. The computational layer is an empirical estimate under a validation or red-team distribution. Mixing these layers is how many probability arguments become ambiguous.

The same discipline applies to generative models. A generator is a measurable transformation of latent randomness. The generated distribution is the pushforward measure. A likelihood, density, or divergence is only meaningful after the target space, base measure, and support relation are clear.

When reading ML papers, silently expand phrases like "sample from the model," "take expectation over data," and "density ratio" into this measure-theoretic checklist. This turns informal notation into a statement that can be checked.

This is the level of precision needed for high-stakes evaluation, off-policy learning, variational inference, and theoretical generalization arguments.

A final question to ask is whether the claim would still be meaningful if the dataset were infinite, the model output lived in a function space, or the event being queried were defined by a limiting process. Measure theory is what keeps the answer honest.

4.4 Variational inference and ELBO

Variational inference and ELBO belongs to the canonical scope of Radon-Nikodym Theorem. Here the point is not to repeat introductory probability, but to expose the measurable structure that makes the probability statement valid.

Working scope for this subsection: absolute continuity, singularity, Radon-Nikodym derivatives, change of measure, Lebesgue decomposition, likelihood ratios, and ML density ratios. The mathematical habit is to name the space, the sigma algebra, the measure, and the map before writing probabilities or expectations.

$$\int f\,dP=\int f\frac{dP}{dQ}\,dQ.$$

Operational definition.

Change of measure rewrites an integral under one measure as a weighted integral under another measure.

Worked reading.

When $P\ll Q$, $\mathbb{E}_P[f]=\mathbb{E}_Q[f(dP/dQ)]$. Importance sampling is this identity estimated by samples from $Q$.

Three examples of variational inference and elbo:

1. Importance-weighted validation under distribution shift.
2. KL divergence via log density ratio.
3. Off-policy policy-gradient correction.

Two non-examples clarify the boundary:

1. Using weights where the proposal misses target support.
2. Taking a likelihood ratio without naming both measures.

Proof or verification habit for variational inference and elbo:

First prove the identity for indicators, extend to simple functions, then use monotone and signed integration.

set question        -> is the subset measurable?
function question   -> are inverse images measurable?
integral question   -> is the function measurable and integrable?
density question    -> is absolute continuity satisfied?
ML question         -> which measure defines the population claim?

In AI systems, variational inference and elbo matters because probability language is constantly compressed into informal notation. Measure theory expands the notation so support, observability, null sets, and convergence assumptions are visible.

Density-ratio methods are everywhere in modern ML: VI, RLHF corrections, domain adaptation, off-policy evaluation, and calibration.

Practical checklist:

• Name the measurable space before naming the probability.
• Identify whether the object is a set, function, measure, distribution, or derivative of measures.
• Check whether equality is pointwise, almost everywhere, or distributional.
• Check whether limits are moved through integrals and which theorem justifies the move.
• For density ratios, check support and absolute continuity before dividing.
• For ML claims, distinguish population measure, empirical measure, model measure, and proposal measure.

Local diagnostic: State the target measure, proposal measure, and derivative.

The notebook version of this subsection uses finite spaces, step functions, empirical measures, or simple density ratios. These toy cases keep the objects visible while preserving the exact logic used in continuous ML models.

The learner should leave this subsection able to translate between the compact ML notation and the full measure-theoretic statement.

A useful way to study this subsection is to keep three layers separate:

1. Semantic layer: what real-world question is being asked?
2. Measurable layer: which event, function, or measure represents that question?
3. Computational layer: which sum, integral, sample average, or ratio estimates it?

For example, the semantic question may be whether a guardrail fails on a class of prompts. The measurable layer is an event in the prompt space. The computational layer is an empirical estimate under a validation or red-team distribution. Mixing these layers is how many probability arguments become ambiguous.

The same discipline applies to generative models. A generator is a measurable transformation of latent randomness. The generated distribution is the pushforward measure. A likelihood, density, or divergence is only meaningful after the target space, base measure, and support relation are clear.

When reading ML papers, silently expand phrases like "sample from the model," "take expectation over data," and "density ratio" into this measure-theoretic checklist. This turns informal notation into a statement that can be checked.

This is the level of precision needed for high-stakes evaluation, off-policy learning, variational inference, and theoretical generalization arguments.

A final question to ask is whether the claim would still be meaningful if the dataset were infinite, the model output lived in a function space, or the event being queried were defined by a limiting process. Measure theory is what keeps the answer honest.

4.5 Off-policy evaluation and policy-change corrections

Off-policy evaluation and policy-change corrections belongs to the canonical scope of Radon-Nikodym Theorem. Here the point is not to repeat introductory probability, but to expose the measurable structure that makes the probability statement valid.

Working scope for this subsection: absolute continuity, singularity, Radon-Nikodym derivatives, change of measure, Lebesgue decomposition, likelihood ratios, and ML density ratios. The mathematical habit is to name the space, the sigma algebra, the measure, and the map before writing probabilities or expectations.

$$D_{\mathrm{KL}}(P\Vert Q)=\int \log\left(\frac{dP}{dQ}\right)dP\quad\text{when }P\ll Q.$$

Operational definition.

Change of measure rewrites an integral under one measure as a weighted integral under another measure.

Worked reading.

When $P\ll Q$, $\mathbb{E}_P[f]=\mathbb{E}_Q[f(dP/dQ)]$. Importance sampling is this identity estimated by samples from $Q$.

Three examples of off-policy evaluation and policy-change corrections:

1. Importance-weighted validation under distribution shift.
2. KL divergence via log density ratio.
3. Off-policy policy-gradient correction.

Two non-examples clarify the boundary:

1. Using weights where the proposal misses target support.
2. Taking a likelihood ratio without naming both measures.

Proof or verification habit for off-policy evaluation and policy-change corrections:

First prove the identity for indicators, extend to simple functions, then use monotone and signed integration.

set question        -> is the subset measurable?
function question   -> are inverse images measurable?
integral question   -> is the function measurable and integrable?
density question    -> is absolute continuity satisfied?
ML question         -> which measure defines the population claim?

In AI systems, off-policy evaluation and policy-change corrections matters because probability language is constantly compressed into informal notation. Measure theory expands the notation so support, observability, null sets, and convergence assumptions are visible.

Density-ratio methods are everywhere in modern ML: VI, RLHF corrections, domain adaptation, off-policy evaluation, and calibration.

Practical checklist:

• Name the measurable space before naming the probability.
• Identify whether the object is a set, function, measure, distribution, or derivative of measures.
• Check whether equality is pointwise, almost everywhere, or distributional.
• Check whether limits are moved through integrals and which theorem justifies the move.
• For density ratios, check support and absolute continuity before dividing.
• For ML claims, distinguish population measure, empirical measure, model measure, and proposal measure.

Local diagnostic: State the target measure, proposal measure, and derivative.

The notebook version of this subsection uses finite spaces, step functions, empirical measures, or simple density ratios. These toy cases keep the objects visible while preserving the exact logic used in continuous ML models.

The learner should leave this subsection able to translate between the compact ML notation and the full measure-theoretic statement.

A useful way to study this subsection is to keep three layers separate:

1. Semantic layer: what real-world question is being asked?
2. Measurable layer: which event, function, or measure represents that question?
3. Computational layer: which sum, integral, sample average, or ratio estimates it?

For example, the semantic question may be whether a guardrail fails on a class of prompts. The measurable layer is an event in the prompt space. The computational layer is an empirical estimate under a validation or red-team distribution. Mixing these layers is how many probability arguments become ambiguous.

The same discipline applies to generative models. A generator is a measurable transformation of latent randomness. The generated distribution is the pushforward measure. A likelihood, density, or divergence is only meaningful after the target space, base measure, and support relation are clear.

When reading ML papers, silently expand phrases like "sample from the model," "take expectation over data," and "density ratio" into this measure-theoretic checklist. This turns informal notation into a statement that can be checked.

This is the level of precision needed for high-stakes evaluation, off-policy learning, variational inference, and theoretical generalization arguments.

A final question to ask is whether the claim would still be meaningful if the dataset were infinite, the model output lived in a function space, or the event being queried were defined by a limiting process. Measure theory is what keeps the answer honest.

5. Common Mistakes

6. Exercises

1. (*) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

2. (*) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

3. (*) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

4. (**) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

5. (**) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

6. (**) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

7. (***) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

8. (***) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

9. (***) Work through a measure-theory task for radon-nikodym theorem.

   • (a) State the measurable space and measure.
   • (b) Identify the relevant measurable set, function, integral, or density.
   • (c) Prove the required property or compute the finite example.
   • (d) Interpret the result for an ML, LLM, or evaluation setting.

10. (***) Work through a measure-theory task for radon-nikodym theorem.

• (a) State the measurable space and measure.
• (b) Identify the relevant measurable set, function, integral, or density.
• (c) Prove the required property or compute the finite example.
• (d) Interpret the result for an ML, LLM, or evaluation setting.

7. Why This Matters for AI

8. Conceptual Bridge

Radon-Nikodym Theorem sits after game theory because deployed AI systems are adaptive, but the probability statements used to evaluate those systems still need rigorous foundations. Strategic behavior changes which measure is relevant; measure theory explains what it means to integrate, compare, and transform those measures.

The backward bridge is probability and information theory. Earlier chapters used PMFs, PDFs, expectations, KL divergence, and likelihoods computationally. Chapter 24 explains the measurable spaces and domination assumptions behind those formulas.

The forward bridge is differential geometry. Once probability measures and density ratios are rigorous, later chapters can treat manifolds, Riemannian metrics, natural gradients, and optimization on curved parameter spaces with less handwaving.

+------------------------------------------------------------------+
| Chapter 23: adaptive agents and strategic pressure               |
| Chapter 24: measurable events, integrals, laws, and densities     |
| Chapter 25: manifolds, geometry, geodesics, and curved learning   |
+------------------------------------------------------------------+

References

• Stanford. Stats 310A Lecture Notes. https://web.stanford.edu/class/stats310a/lnotes.pdf
• UC Davis. Lecture Notes on Measure Theory. https://www.math.ucdavis.edu/~hunter/measure_theory/measure_theory.html
• Lawler. Notes on Probability. https://www.math.uchicago.edu/~lawler/probnotes.pdf
• Wolfram MathWorld. Radon-Nikodym Theorem. https://mathworld.wolfram.com/RadonNikodymTheorem.html