Distribution Of Function Of Random Variable

Introduction

In probability theory, a random variable is a numerical outcome of a random experiment. Even so, when we apply a deterministic function to that variable—say, squaring it, taking its absolute value, or converting temperature from Celsius to Fahrenheit—we obtain a new quantity that is itself a random variable. In practice, the distribution of a function of a random variable describes how the probabilities of the original outcomes are transferred to the transformed outcomes. Understanding this transformation is essential for tasks ranging from statistical modeling and simulation to engineering reliability analysis and machine‑learning feature engineering.

This article explains the theory behind the distribution of a function of a random variable, walks through the mechanics of deriving that distribution, illustrates the ideas with concrete examples, highlights common pitfalls, and answers frequently asked questions. By the end, you should feel comfortable determining the probability law of any transformed variable, whether the function is monotonic, piecewise, or many‑to‑one But it adds up..

Detailed Explanation

What Does “Distribution of a Function” Mean?

Let (X) be a random variable with known probability distribution—either a probability mass function (PMF) (p_X(x)) for discrete (X) or a probability density function (PDF) (f_X(x)) for continuous (X). Consider a measurable function (g:\mathbb{R}\to\mathbb{R}). The function of a random variable is defined as

[ Y = g(X). ]

Because (X) is random, (Y) inherits randomness. The goal is to find the distribution of (Y): its PMF (p_Y(y)) if (Y) is discrete, or its PDF (f_Y(y)) if (Y) is continuous. In essence, we “push forward” the probability mass or density of (X) through the mapping (g).

Why Is This Important?

Many practical problems involve non‑linear transformations. For instance:

• In finance, the log‑return of a stock price is (Y = \ln(X)) where (X) is the price.
• In signal processing, the power of a signal is (Y = X^2).
• In reliability engineering, the time to failure of a system may be expressed as a function of component lifetimes.

Knowing the distribution of (Y) enables us to compute expectations, variances, quantiles, and tail probabilities without resorting to simulation every time.

General Approach

The derivation depends on whether (X) is discrete or continuous and whether (g) is invertible (monotonic) or not. The two main tools are:

1. Change‑of‑variables formula for continuous, monotonic (g).
2. Law of the unconscious statistician (LOTUS) and partitioning method for general (g) (including many‑to‑one mappings).

Both approaches ultimately sum or integrate the original probability over the set of (x) that map to a given (y) The details matter here. Surprisingly effective..

Step‑by‑Step or Concept Breakdown

Below is a step‑by‑step recipe that works for both discrete and continuous cases, with special notes for monotonic functions.

Step 1: Identify the Type of (X)

• Discrete – work with PMFs and sums.
• Continuous – work with PDFs and integrals.

Step 2: Write the Definition of the Target Variable

[ Y = g(X). ]

Step 3: Express the Event ({Y \le y}) (or ({Y = y}) for discrete) in Terms of (X)

• For continuous (Y):
  [ F_Y(y) = P(Y \le y) = P\bigl(g(X) \le y\bigr) = P\bigl(X \in g^{-1}((-\infty, y])\bigr). ]
• For discrete (Y):
  [ p_Y(y) = P(Y = y) = P\bigl(g(X) = y\bigr) = \sum_{x: g(x)=y} p_X(x). ]

Step 4: Invert the Inequality (If Possible)

If (g) is strictly monotonic (either increasing or decreasing) on the support of (X), the inverse function (g^{-1}) exists and is single‑valued. Then:

• For increasing (g):
  [ {g(X) \le y} = {X \le g^{-1}(y)}. ]
• For decreasing (g):
  [ {g(X) \le y} = {X \ge g^{-1}(y)}. ]

Plug this into the CDF or PDF expression.

Step 5: Differentiate (Continuous Case) or Sum (Discrete Case)

• Continuous, monotonic (g):
  [ f_Y(y) = f_X\bigl(g^{-1}(y)\bigr) \cdot \left|\frac{d}{dy} g^{-1}(y)\right|. ]
  This is the classic change‑of‑variables formula; the absolute value accounts for possible decreasing (g) The details matter here..

• Discrete or non‑monotonic (g):
  Discrete: sum over all pre‑images:
  [ p_Y(y) = \sum_{x \in g^{-1}(y)} p_X(x). ]
  Continuous, general (g): partition the domain into intervals where (g) is monotonic, apply the change‑of‑variables formula on each piece, and add the contributions:
  [ f_Y(y) = \sum_{i} f_X\bigl(g_i^{-1}(y)\bigr) \left|\frac{d}{dy} g_i^{-1}(y)\right|, ]
  where each (g_i) is a monotonic branch of (g) That's the whole idea..

Step 6: Validate the Result

• Check that (f_Y(y) \ge 0) for all (y).
• Verify that (\int_{-\infty}^{\infty} f_Y(y),dy = 1) (continuous) or (\sum_y p_Y(y)=1) (discrete).
• Optionally compute (E[Y]) via LOTUS: (E[Y] = \sum g(x)p_X(x)) or (\int g(x)f_X(x)dx) to confirm consistency.

Real Examples

Example 1: Squaring a Standard Normal Variable

Let (X \sim N(0,1)) (standard normal). Define (Y = X^2).

• (g(x)=x^2) is not monotonic; it is decreasing on ((-\infty,0]) and increasing on ([0,\infty)).
• The support of (X) splits into two monotonic branches: (g_1(x)=x^2) for (x\le 0) (decreasing) and (g_2(x)=x^2) for (x\ge 0) (increasing).

For (g).

Using the change‑of‑variables formula on each branch:

[ f_Y(y) = f_X(-\sqrt{y})\left|\frac{d}{dy}(-\sqrt{y})\right| + f

X(\sqrt{y})\left|\frac{d}{dy}(\sqrt{y})\right|]

Given $f_X(x) = \frac{1}{\sqrt{2\pi}} e^{-\frac{x^2}{2}}$, we calculate the derivatives:

1. 2. Practically speaking, for the branch $x = -\sqrt{y}$: $\frac{dx}{dy} = -\frac{1}{2\sqrt{y}}$. For the branch $x = \sqrt{y}$: $\frac{dx}{dy} = \frac{1}{2\sqrt{y}}$.

Substituting these into the formula: [ f_Y(y) = \frac{1}{\sqrt{2\pi}} e^{-\frac{(-\sqrt{y})^2}{2}} \left| -\frac{1}{2\sqrt{y}} \right| + \frac{1}{\sqrt{2\pi}} e^{-\frac{(\sqrt{y})^2}{2}} \left| \frac{1}{2\sqrt{y}} \right| ] [ f_Y(y) = \frac{1}{\sqrt{2\pi}} e^{-\frac{y}{2}} \left( \frac{1}{2\sqrt{y}} + \frac{1}{2\sqrt{y}} \right) = \frac{1}{\sqrt{2\pi y}} e^{-\frac{y}{2}} ]

This result is the PDF of a Chi-squared distribution with one degree of freedom ($\chi^2_1$), which matches theoretical expectations.

Example 2: Linear Transformation of a Uniform Variable

Let $X \sim \text{Uniform}(a, b)$. Define $Y = mX + c$.

• Since $g(x) = mx + c$ is a linear function, it is monotonic (assuming $m \neq 0$).
• The inverse function is $x = g^{-1}(y) = \frac{y - c}{m}$.
• The derivative is $\frac{dx}{dy} = \frac{1}{m}$.

Using the change-of-variables formula: [ f_Y(y) = f_X\left(\frac{y-c}{m}\right) \cdot \left| \frac{1}{m} \right| ]

Since $f_X(x) = \frac{1}{b-a}$ for $x \in [a, b]$, we have: [ f_Y(y) = \frac{1}{b-a} \cdot \frac{1}{|m|} = \frac{1}{|m|(b-a)} ] The support of $Y$ is the interval between $ma+c$ and $mb+c$ (depending on the sign of $m$). This confirms that a linear transformation of a uniform distribution remains a uniform distribution Not complicated — just consistent. And it works..

Conclusion

Transforming random variables is a fundamental skill in probability theory, essential for everything from statistical inference to machine learning. The method used depends heavily on the nature of the transformation $g(X)$:

1. Monotonic Transformations: These are the simplest to handle via the direct change-of-variables formula, as the mapping between $X$ and $Y$ is one-to-one.
2. Non-Monotonic Transformations: These require partitioning the domain into monotonic segments, calculating the density for each segment, and summing the results.
3. Discrete Transformations: These rely on summing the probabilities of all $x$ values that map to a specific $y$.

By mastering these techniques—specifically the ability to identify monotonic branches and correctly apply the Jacobian (the absolute derivative)—you can derive the distributions of complex functions of simpler, well-known random variables.

The process of transforming random variables reveals the elegant interplay between calculus and probability theory. Because of that, when dealing with monotonic functions, the change-of-variables formula provides a straightforward path to the transformed distribution, requiring only careful application of the inverse function and its derivative. Even so, the true power of this technique emerges when handling more complex scenarios Easy to understand, harder to ignore. Still holds up..

Non-monotonic transformations, such as squaring a normal variable, demonstrate how multiple branches must be considered simultaneously. Each monotonic segment contributes to the final density, and the absolute value in the Jacobian ensures proper accounting for orientation changes. This approach naturally leads to important distributions like the chi-squared, showing how fundamental distributions arise from simple transformations of others.

For discrete random variables, the methodology shifts from integration to summation, reflecting the countable nature of possible outcomes. The key insight remains consistent: we must account for all ways the original variable can produce a given transformed value.

In practical applications, these techniques enable statisticians and data scientists to:

• Derive sampling distributions for test statistics
• Model complex phenomena through transformations of simpler processes
• Understand how uncertainty propagates through mathematical operations
• Generate new distributions for specialized modeling needs

Mastering variable transformation not only provides computational tools but also deepens understanding of distributional relationships. It's a foundational skill that bridges theoretical probability with real-world applications, making it indispensable for anyone working with statistical models or probabilistic systems.