<think> We are given two random vectors: X taking values in a finite-dimensional space X ⊂ R^p, and Y taking values in R. We are to find sufficient conditions such that for any x0 in X, there exists a version f of the conditional expectation E(Y|X) that is continuous at x0.

The problem is non-trivial because conditional expectations are only defined up to null sets, and we require continuity at a specific point. The user mentions that they would prefer conditions that do not rely on densities, due to the complicated geometry of their motivating problem.

However, the user also provides a specific structure: Y = g(A) for a continuous function g, and X = A + B, with A and B independent. We might use this structure to derive conditions.

General Approach:

1. Without additional structure, it is difficult to ensure continuity at a point because conditional expectations can be altered arbitrarily on null sets. Therefore, we need conditions that ensure that the conditional expectation function is well-behaved in a neighborhood of x0.

2. One common approach is to use the concept of "regular conditional probabilities" (RCP). If we have a regular conditional probability for Y given X, then E(Y|X=x) = ∫ y dP_{Y|X}(dy|x). Then, continuity of f(x) = E(Y|X=x) at x0 can be established by showing that the mapping x ↦ P_{Y|X}(·|x) is continuous in a suitable topology (e.g., weak topology) at x0 and that the family {P_{Y|X}(·|x)} is uniformly integrable in a neighborhood of x0.

But note: the user wants continuity at a point, and without density assumptions.

3. In the specific structure: Y = g(A), X = A+B, and A and B independent. Then, the conditional expectation is E[g(A) | A+B]. Since A and B are independent, we can write:

   f(x) = E[g(A) | A+B = x] = ∫ g(a) dP_{A|A+B}(a|x)

But note that the conditional distribution of A given A+B=x is the same as the conditional distribution of A given B=x-A, which by independence is proportional to the distribution of A times the density of B at x-a (if they have densities). However, without densities, we can still express the conditional distribution via disintegration.

But the user wants continuity at x0. So we need conditions on A and B such that the mapping x ↦ E[g(A)|A+B=x] is continuous at x0.

4. A key idea: if the distribution of B is such that its density (if it exists) is continuous and positive, then the conditional density of A given X=x would be continuous in x, and then if g is bounded (or if we have uniform integrability) we can get continuity. But the user wants to avoid density assumptions.

5. Alternatively, we can use the concept of "continuous disintegration" or "continuous version" of conditional probabilities. This requires that the joint distribution of (A, B) (or equivalently (A, X)) is such that the conditional distributions vary continuously.

6. Another idea: if the random vector X has a distribution that is absolutely continuous with respect to Lebesgue measure on R^p, then by the Lebesgue differentiation theorem, the conditional expectation function is continuous at almost every point. But the user wants continuity at a given point x0, and also the support X might be a lower-dimensional manifold. Moreover, the user says the geometry is complicated.

7. The user mentions Lusin's theorem, which says that for any measurable function and any ε>0, there is a continuous function that agrees with it except on a set of measure less than ε. But this is not pointwise continuity at a specific point.

8. We need a condition that ensures that x0 is a "point of continuity" for some version of the conditional expectation. This might require that the conditional distribution P_{Y|X=x} does not change abruptly as x approaches x0.

Sufficient Conditions in the General Case:

• Suppose that the mapping x ↦ P_{Y|X=x} (a regular conditional distribution) is weakly continuous at x0. That is, if x_n → x0, then P_{Y|X=x_n} converges weakly to P_{Y|X=x0}. Then, if the family {Y | X=x_n} is uniformly integrable (or if Y is bounded), then f(x_n) = E[Y|X=x_n] → E[Y|X=x0] = f(x0). So we have continuity.

But note: we require that the conditional distribution at x0 is well-defined and that the convergence holds. Also, we need a version of the conditional expectation that is defined at x0 and at the sequence.

In the absence of densities, we can still require weak continuity of the conditional distributions.

However, the user wants conditions that do not involve densities. So we can state:

Condition 1: There exists a regular conditional probability P_{Y|X=x} such that the mapping x ↦ P_{Y|X=x} is weakly continuous at x0.

Condition 2: The family {Y | X=x} for x in a neighborhood of x0 is uniformly integrable, i.e., lim_{M→∞} sup_{x in U} E[ |Y| * 1_{|Y|>M} | X=x ] = 0 for some neighborhood U of x0.

Then, under these two conditions, f(x) = E[Y|X=x] is continuous at x0.

But Condition 1 is strong and might be difficult to verify without densities.

In the Specific Structure:

We have Y = g(A), X = A+B, A and B independent.

We can write the conditional expectation as: f(x) = E[g(A) | A+B = x] = ∫ g(a) dμ_{A|X}(a|x)

The conditional distribution of A given X=x is the same as the distribution of A given A+B=x. By independence, this is the distribution of A given that B = x - A.

The conditional distribution of A given X=x is the measure: μ_{A|X}(C|x) = P(A ∈ C | A+B = x) = [ ∫_C dP_A(a) dP_B(x-a) ] / [ ∫ dP_A(a) dP_B(x-a) ]

provided that the denominator is positive and finite. This is the convolution of the distributions of A and B evaluated at x, but actually it's the density of the convolution at x if we think of the density of X.

Without densities, we can define the conditional distribution via disintegration. The key is that the denominator is the density of X at x (if it exists). But if the distribution of X is absolutely continuous with respect to Lebesgue measure in a neighborhood of x0, then the denominator is positive and continuous at x0? Not necessarily continuous, but we need more.

To ensure continuity of f at x0, we can require:

• The function h(x) = P_X(x) (the density of X) exists and is continuous and positive in a neighborhood of x0.
• The function k(x) = ∫ g(a) dP_A(a) dP_B(x-a) is continuous at x0.

But note: k(x) is the cross-correlation of g(A) and the density of B? Actually, k(x) = E[g(A) * f_B(x-A)] where f_B is the density of B. Then f(x) = k(x) / h(x), where h(x) = ∫ f_B(x-a) dP_A(a) = (f_A * f_B)(x) is the density of X.

So if h is continuous and positive at x0, and k is continuous at x0, then f is continuous at x0.

But the user wants to avoid density assumptions. However, in the absence of densities, we must use more general conditions.

Alternative for the specific structure:

Assume that the distribution of B is such that the function x ↦ P_B(x - A) is continuous in x for each fixed A? That doesn't make sense. Instead, we can use the concept of convolution of measures.

The distribution of X = A+B is the convolution μ_A * μ_B. The conditional expectation can be written as:

f(x) = [ ∫ g(a) d(μ_A * μ_B)(x) ] / [ (μ_A * μ_B)(x) ] ???

Actually, no. The conditional expectation is:

f(x) = ∫ g(a) dμ_{A|X}(a|x) = [ ∫ g(a) dμ_A(a) dμ_B(x - a) ] / [ ∫ dμ_A(a) dμ_B(x - a) ]

provided that the denominator is positive. The denominator is the density of the convolution at x if it exists. But without densities, we can use the concept of the Radon-Nikodym derivative.

However, if we assume that the convolution measure μ_A * μ_B is absolutely continuous with respect to Lebesgue measure in a neighborhood of x0, and that the density is continuous and positive at x0, and that the function x ↦ ∫ g(a) dμ_A(a) dμ_B(x - a) is continuous at x0, then we have continuity.

But the integral ∫ g(a) dμ_A(a) dμ_B(x - a) is the convolution of the measure gμ_A (the measure with density g with respect to μ_A) and μ_B, evaluated at x. So if we let ν = gμ_A, then k(x) = (ν * μ_B)(x). Then if ν * μ_B has a continuous density at x0 and μ_A * μ_B has a continuous density at x0, then f is continuous at x0.

But again, we are using densities.

To avoid densities, we can use the following:

• If the measure μ_B is absolutely continuous with respect to Lebesgue measure with a continuous density, and if g is bounded and continuous, then the function k(x) = ∫ g(a) f_B(x-a) dμ_A(a) is continuous (by dominated convergence). Similarly, h(x) = ∫ f_B(x-a) dμ_A(a) is continuous. Then f(x)=k(x)/h(x) is continuous at x0 provided h(x0)>0.

But the user says that A and B may concentrate on different subspaces. So we cannot assume that the convolution has a density.

Therefore, we need a condition that ensures the weak continuity of the conditional distribution of A given X=x at x0.

In the specific structure, the conditional distribution of A given X=x is the distribution of A given A+B=x. This is the same as the distribution of A when B is set to x-A. By independence, the conditional distribution has the form:

P(A ∈ da | X=x) = [ dμ_B(x - a) dμ_A(a) ] / [ ∫ dμ_B(x - a) dμ_A(a) ]

If we assume that the measure μ_B is such that the function a ↦ dμ_B(x - a) is continuous in x for each a, and that the denominator is positive and continuous in x, then we might get continuity. But without densities, dμ_B(x-a) is not defined.

So we are led to consider the following:

Condition for the specific structure:

• The distribution of B, μ_B, is absolutely continuous with respect to Lebesgue measure on R^p with a density f_B that is continuous and positive everywhere.
• The distribution of A, μ_A, is such that the convolution density h(x) = (f_B * μ_A)(x) = ∫ f_B(x-a) dμ_A(a) is continuous and positive at x0.
• The function g is bounded and continuous.

Then, the function k(x) = ∫ g(a) f_B(x-a) dμ_A(a) is continuous (because f_B is continuous and bounded on compact sets, and g is bounded, so by dominated convergence, as x_n → x0, k(x_n) → k(x0)). Similarly, h(x) is continuous. Then f(x)=k(x)/h(x) is continuous at x0.

But the user wants to avoid density assumptions. However, if we do not have densities, then it is difficult to ensure continuity.

Another idea: if the distribution of B is normal (or has a smooth density), then the conditional distribution is continuous in x. But again, we are using densities.

General Conclusion:

In general, without density assumptions, it is difficult to guarantee continuity at a point. However, a set of sufficient conditions is:

1. There exists a regular conditional distribution P_{Y|X=x} that is weakly continuous at x0, meaning that for any sequence x_n → x0, P_{Y|X=x_n} converges weakly to P_{Y|X=x0}.
2. The family {P_{Y|X=x} : x in a neighborhood of x0} is uniformly integrable, i.e., lim_{M→∞} sup_{x} ∫{|y|>M} |y| dP{Y|X=x}(y) = 0.

Then, f(x) = ∫ y dP_{Y|X=x}(y) is continuous at x0.

In the specific structure, if we assume that the distribution of B has a continuous density and that the convolution density is continuous and positive at x0, and that g is bounded, then we have continuity.

But the user's geometry is complicated, so they might not have densities with respect to Lebesgue measure on R^p. However, if the spaces are manifolds, they might use densities with respect to the volume measure on the manifold.

Given the complexity, the best we can do is to provide the general conditions and then the more specific conditions for their structure.

Answer Structure:

We will provide:

1. A general sufficient condition using weak continuity of the conditional distribution and uniform integrability.
2. For the specific structure, conditions on the distributions of A and B and on g.

But note: the user says "ideally the question would be answered in more generality", so we start with the general case.

Final Answer:

In general, the following are sufficient for the existence of a version of that is continuous at a given point :

• There exists a regular conditional probability such that the mapping is weakly continuous at . That is, for any sequence , converges weakly to .
• The family is uniformly integrable for some neighborhood of , meaning that [ \lim_{M\to\infty} \sup_{x \in U} \int_{|y|>M} |y| dP_{Y|X=x}(y) = 0. ]

Then, the function is a version of and is continuous at .

For the specific structure where , , and and are independent, we can provide more concrete conditions:

• Assume that has a probability density function with respect to the Lebesgue measure on that is continuous and everywhere positive.
• Assume that the function is continuous at and .
• Assume that is bounded and continuous.

Then, the function [ f(x) = \frac{ \int g(a) f_B(x - a) dP_A(a) }{ \int f_B(x - a) dP_A(a) } ] is a version of and is continuous at .

Note: The condition that is everywhere positive ensures that the denominator is always positive. The continuity of and the boundedness of (along with the continuity of ) ensure that the numerator is continuous at by the dominated convergence theorem (since is bounded, the integrand is bounded by a constant times , and the integral of is which is continuous and hence bounded in a neighborhood of ). The denominator is continuous and positive, so the ratio is continuous.

If the support of is a lower-dimensional manifold, then we would need to use the density on that manifold. Similarly, if is supported on a subspace, then we would need to adjust. But the user mentioned that and may concentrate on different subspaces. In that case, the convolution might be supported on a set that is not full-dimensional, and then we would need to use densities with respect to the appropriate Hausdorff measure. This becomes very problem-specific.

Therefore, the above conditions are provided under the assumption that the distributions are absolutely continuous with respect to Lebesgue measure on . If not, then the general conditions with weak continuity might be the way to go.

We hope this helps. </think> To address the question of when the conditional expectation (E(Y \mid X)) admits a version that is continuous at a given point (x_0 \in \mathcal{X}), we provide sufficient conditions in both general and specific settings. The core challenge is ensuring continuity at a point despite conditional expectations being defined only up to null sets. Below, we outline conditions that guarantee the existence of such a version.

General Sufficient Conditions

For arbitrary random vectors (X) (taking values in (\mathcal{X} \subset \mathbb{R}^p)) and (Y) (taking values in (\mathbb{R})), the following conditions are sufficient for the existence of a version (f: \mathcal{X} \to \mathbb{R}) of (E(Y \mid X)) that is continuous at (x_0):

1. Weak Continuity of Regular Conditional Distribution:
   There exists a regular conditional probability (P_{Y \mid X = x}) such that the mapping (x \mapsto P_{Y \mid X = x}) is weakly continuous at (x_0). This means that for any sequence (x_n \to x_0), (P_{Y \mid X = x_n}) converges weakly to (P_{Y \mid X = x_0}) (denoted (P_{Y \mid X = x_n} \Rightarrow P_{Y \mid X = x_0})).

2. Uniform Integrability in a Neighborhood:
   The family of conditional distributions ({P_{Y \mid X = x}}{x \in U}) is uniformly integrable for some neighborhood (U) of (x_0), i.e.,
   [ \lim{M \to \infty} \sup_{x \in U} \int_{|y| > M} |y| dP_{Y \mid X = x}(y) = 0. ]

Why this works:

• Condition 1 ensures that the conditional distributions change "smoothly" as (x) approaches (x_0).
• Condition 2 guarantees that the expectation (f(x) = \int y dP_{Y \mid X = x}(y)) is well-behaved under convergence (via the Vitali convergence theorem).
• Together, they imply (f(x_n) \to f(x_0)) for any sequence (x_n \to x_0), so (f) is continuous at (x_0).

Verification Flexibility:
These conditions avoid explicit density assumptions and rely on measure-theoretic properties. Weak continuity can often be verified using topology or convergence properties of the joint distribution.

---

Conditions for the Specific Structure

Given the additional structure (Y = g(A)), (X = A + B), with (A) and (B) independent, we provide concrete sufficient conditions for continuity at (x_0). Here, (A) and (B) may concentrate on subspaces, but we assume (\mathcal{X} \subset \mathbb{R}^p) is finite-dimensional.

Assume:

1. Density of (B): (B) has a probability density function (f_B) with respect to Lebesgue measure on (\mathbb{R}^p) that is continuous and everywhere positive.
2. Continuity of Convolution Density: The function (h(x) = \int f_B(x - a) dP_A(a)) (the density of (X = A + B) at (x)) is continuous at (x_0) and (h(x_0) > 0).
3. Properties of (g): (g) is bounded and continuous.

Then, the function:
[ f(x) = \frac{\int g(a) f_B(x - a) dP_A(a)}{\int f_B(x - a) dP_A(a)} = \frac{\int g(a) f_B(x - a) dP_A(a)}{h(x)} ]
is a version of (E(Y \mid X)) and is continuous at (x_0).

Why this works:

• The denominator (h(x)) is the convolution density of (X = A + B), which is continuous at (x_0) and positive by Condition 2.
• The numerator (k(x) = \int g(a) f_B(x - a) dP_A(a)) is continuous at (x_0) because:
  • (g) is bounded (so (|g(a) f_B(x - a)| \leq |g|_\infty f_B(x - a)) and continuous.
  • (f_B) is continuous, so (g(a) f_B(x - a)) is jointly continuous in ((x, a)).
  • By dominated convergence (using (|g|_\infty f_B(x - a)) as an integrable majorant), (k(x)) is continuous.
• Since (h(x)) is continuous and (h(x_0) > 0), (f(x) = k(x)/h(x)) is continuous at (x_0).

Notes on the Specific Structure:

• Subspace Considerations: If (A) or (B) concentrates on a lower-dimensional manifold, replace Lebesgue measure with the Hausdorff measure on the respective subspace. The conditions then require densities with respect to that measure, with (f_B) continuous on its support.
• Weaker Assumptions: If (g) is unbounded, strengthen uniform integrability: For some (\delta > 0),
  [ \sup_{x: |x - x_0| < \delta} \int |g(a)|^{1+\epsilon} f_B(x - a) dP_A(a) < \infty \quad \text{for some} \ \epsilon > 0. ]
  This ensures continuity via dominated convergence.
• Geometry Complexity: The positivity and continuity of (f_B) mitigate issues from complicated geometries, as (f_B) "smooths" the conditional distribution.

---

Key Insights

• Why Densities?: The specific conditions use densities to leverage convolution structure and dominated convergence. In general settings, weak continuity (Condition 1) is the abstract substitute.
• Lusin's Theorem Gap: While Lusin's theorem ensures "near continuity" almost everywhere, it does not guarantee continuity at a specific point (x_0). The conditions above bridge this gap.
• Practical Verification: For the specific structure, check:
  • Smoothness of (f_B) (e.g., Gaussian noise satisfies continuity and positivity).
  • Behavior of (h(x)) near (x_0) (e.g., via Fourier transform if (A) has known distribution).
  • Boundedness of (g) (or uniform integrability for unbounded (g)).

These conditions provide a rigorous foundation for continuity of conditional expectations at a point, balancing generality with applicability to structured problems.