同自由度独立缩放非中心χ²变量之和的分析与近似方法咨询

Hey there! Let's break down how to analyze and approximate the distribution of (X = X_1 + aX_2), where (X_1 \sim \chi^2(k,\lambda_1)) and (X_2 \sim \chi^2(k,\lambda_2)) are independent non-central chi-squared variables. You mentioned you've started with characteristic functions—excellent starting point, so let's expand on that and cover practical, actionable methods below.

First, let's formalize the characteristic function (CF) of (X) since you're already working with it. For independent variables, the CF of the sum is the product of individual CFs. The CF of a non-central chi-squared variable (\chi^2(k,\lambda)) is:

M_{χ²(k,λ)}(t) = (1-2t)^{-k/2} \exp\left( \frac{\lambda t}{1-2t} \right)

So for (X), this becomes:

M_X(t) = (1-2t)^{-k/2} \exp\left( \frac{\lambda_1 t}{1-2t} \right) \cdot (1-2at)^{-k/2} \exp\left( \frac{\lambda_2 a t}{1-2at} \right)

Moment matching is a straightforward, widely used approximation technique here:

• Calculate the first few moments of (X) using the properties of chi-squared variables:
  • Mean: (\mathbb{E}[X] = (k + \lambda_1) + a(k + \lambda_2))
  • Variance: (\text{Var}(X) = 2(k + 2\lambda_1) + a^2 \cdot 2(k + 2\lambda_2))
  • Higher-order moments (skewness, kurtosis) can be derived via CF derivatives or direct chi-squared moment formulas.
• Match these moments to a target distribution (Gamma or generalized non-central chi-squared):
  • Gamma Distribution Fit: For (\text{Gamma}(\alpha, \beta)) (shape (\alpha), rate (\beta)), solve:
    (\alpha = \frac{(\mathbb{E}[X])^2}{\text{Var}(X)}), (\beta = \frac{\mathbb{E}[X]}{\text{Var}(X)})
    For better precision, match skewness (Gamma's skewness is (2/\sqrt{\alpha})) to the theoretical or sample skewness of (X).
  • Generalized Non-Central χ² Fit: For (\chi^2(\nu, \lambda)) (allowing non-integer (\nu)), solve the system:
    (\nu + \lambda = \mathbb{E}[X]), (2(\nu + 2\lambda) = \text{Var}(X))
    This gives (\nu = 2\mathbb{E}[X] - \frac{\text{Var}(X)}{2}) and (\lambda = \frac{\text{Var}(X)}{2} - \mathbb{E}[X]).

If you need better accuracy—especially for tail probabilities—saddlepoint approximation is a powerful tool built on the cumulative generating function (CGF, the log of the CF):

• First, compute the CGF of (X):

  K_X(t) = -\frac{k}{2}\log(1-2t) + \frac{\lambda_1 t}{1-2t} -\frac{k}{2}\log(1-2at) + \frac{\lambda_2 a t}{1-2at}
  

• Find the saddlepoint (t_0) by solving (K_X'(t_0) = x) (for a target value (x)) or (K_X'(t_0) = \mathbb{E}[X]) (for pdf approximation). This usually requires numerical methods like Newton-Raphson.
• Use the saddlepoint pdf approximation formula:

  f_X(x) \approx \frac{1}{\sqrt{2\pi K_X''(t_0)}} \exp\left( K_X(t_0) - t_0 x \right)
  

This method outperforms moment matching in most cases, particularly when (a) is far from 1 or non-central parameters (\lambda_1, \lambda_2) are very different.

Since you're already running simulations, leverage that data to refine your approximations:

• Generate a large sample of (X) values, then compute sample moments or empirical pdfs.
• Adjust parameters of your target distribution (Gamma/non-central χ²) to minimize the KL divergence or mean squared error between the empirical and approximate pdfs.
• For tail risk analysis, compare simulated quantiles to those from your approximate distribution and tweak parameters until errors are within your acceptable threshold.

Don't overlook edge cases where exact or simpler approximations work:

• When (a = 1): (X) is exactly a non-central chi-squared variable with (2k) degrees of freedom and non-central parameter (\lambda_1 + \lambda_2). The CF simplifies to ((1-2t)^{-k} \exp\left( \frac{(\lambda_1+\lambda_2)t}{1-2t} \right)), which matches (\chi^2(2k, \lambda_1+\lambda_2)).
• When (a) is very small: (X \approx X_1), so approximate with (\chi^2(k, \lambda_1)).
• When (a) is very large: (X \approx aX_2), so approximate with a scaled non-central chi-squared (or a Gamma distribution, since scaled chi-squared is a special case of Gamma).

内容的提问来源于stack exchange，提问作者Sabrina Cantu