Probability Interview Prep
Probability interview prep — random variables, distributions, expectation, conditional probability, Bayes, CLT, MGFs, and convergence theorems for engineering candidates.
Quick answer
Probability is the mathematical theory of uncertainty, formalized through sample spaces, events, and probability measures (Kolmogorov axioms).
Probability questions test three layers at once: mechanical fluency (compute the expectation, conditional probability, or moment), conceptual understanding (state when CLT applies, distinguish independence from uncorrelatedness, articulate the tower property), and modeling judgment (which distribution captures this scenario, which assumption is being violated, what is the failure mode of the approximation).
Editorial review
Written by
CompoundLearn editorial team
Wireless / RF / hardware engineering
Reviewed by
CompoundLearn editorial team
Wireless / RF / hardware engineering
Last reviewed
Built from curated topic maps, editorial validation, and subject-matter review so the page stays aligned with the interview intent and the current content pipeline.
What it is
Probability is the mathematical theory of uncertainty, formalized through sample spaces, events, and probability measures (Kolmogorov axioms). Standard topics include random variables (discrete and continuous), probability mass and density functions, cumulative distribution functions, expectation and variance, joint distributions and marginals, conditional probability and independence, Bayes' theorem and Bayesian reasoning, common families (Bernoulli, Binomial, Poisson, Geometric, Uniform, Exponential, Gaussian, Beta, Gamma), the Law of Large Numbers, the Central Limit Theorem and its variants, moment generating functions and characteristic functions, transformations of random variables (change-of-variables formula), conditional expectation and the tower property, covariance and correlation, multivariate Gaussian and its closure properties, Markov chains and stationary distributions, Poisson processes and renewal theory, and convergence concepts (almost sure, in probability, in distribution, L^p). For engineering candidates, probability is the foundation of communications (BER analysis is integral of conditional Gaussian tails), reliability engineering (survival analysis), machine learning (every model is a probability distribution over outcomes), quantitative finance (option pricing as discounted expected payoff), signal processing (random processes, noise models), and any system that operates under uncertainty. Strong probabilistic intuition — knowing when independence applies, when CLT kicks in, when Markov approximations break — separates serious candidates from formula-pattern-matchers.
Why interviewers ask
Probability questions test three layers at once: mechanical fluency (compute the expectation, conditional probability, or moment), conceptual understanding (state when CLT applies, distinguish independence from uncorrelatedness, articulate the tower property), and modeling judgment (which distribution captures this scenario, which assumption is being violated, what is the failure mode of the approximation). ML interviews lean heavily on probability — every loss function is a likelihood, every regularizer is a prior, every inference algorithm is approximate Bayesian computation. Communications interviews use probability for BER analysis, channel-capacity arguments, and noise modeling — the canonical probe is "what is the BER of BPSK in AWGN" which decomposes into a tail-integral of a Gaussian. Quantitative-research interviews test probability heavily because trading strategies are bets on distributions and risk management is tail estimation. The strongest candidates fluently switch between formula and intuition: they can compute Bayes' rule mechanically, sketch the prior-and-likelihood-to-posterior reasoning geometrically, and quote canonical results (mean and variance of common distributions, MGFs of independent sums) without hesitation. Interviewers also use probability to test whether candidates have internalized the difference between probability theory and statistics — probability assumes the distribution and computes consequences; statistics observes data and infers the distribution. Confusing the two is a senior-level red flag.
Common mistakes
The most common mistake is treating PDFs as if they were probabilities — candidates substitute a value into a Gaussian PDF and report it as P(X = x), which is always 0 for continuous variables; the PDF value is a density, and probabilities require integration. A second mistake is conflating independence with uncorrelatedness, especially for non-Gaussian random variables; candidates "prove independence" by showing zero correlation, which is insufficient. A third mistake is base-rate neglect on Bayes problems — candidates apply the formula but ignore the prior, producing answers that are off by orders of magnitude. Fourth, candidates often forget the |dy/dx| Jacobian in change-of-variables for continuous distributions, leading to densities that do not integrate to 1. Fifth, on conditional expectation, weak candidates compute E[X | Y = y] for a specific y instead of the random variable E[X | Y], which is a function of Y; the tower property and martingale arguments require the latter. Sixth, on CLT, candidates apply the theorem to non-IID samples or to distributions with infinite variance (heavy tails) and report Gaussian convergence that does not hold. Seventh, on Markov chains, candidates blur two distinct conditions: irreducibility (plus positive recurrence) is what guarantees a unique stationary distribution, whereas aperiodicity is the extra condition needed for the chain to actually converge to it from any start — a periodic but irreducible finite chain still has a unique stationary distribution, it just oscillates around it instead of settling. Eighth, on multivariate Gaussians, candidates do not exploit closure properties — linear combinations of jointly Gaussians are Gaussian, conditional Gaussians are Gaussian — and reach for general integration when a closed form exists. Finally, candidates confuse "almost surely", "in probability", and "in distribution" convergence, which matters in advanced probability and statistical theory and is exactly what senior interviewers probe to distinguish coursework-only candidates from those who have internalized the modes of convergence.
Frequently asked questions
- What is the difference between PMF, PDF, and CDF?
- A probability mass function (PMF) gives P(X = x) for discrete random variables — values are probabilities and sum to 1. A probability density function (PDF) describes continuous random variables — its value at x is not a probability but a density, and probabilities are integrals over intervals. A cumulative distribution function (CDF) is F(x) = P(X ≤ x) for any random variable; it is monotone non-decreasing, right-continuous, and ranges from 0 to 1. CDFs work uniformly for discrete and continuous, which is why measure-theoretic probability uses them as the primary object.
- When does the Central Limit Theorem apply, and what does it actually say?
- The classical CLT says that the sum (or sample mean) of n independent identically-distributed random variables with finite variance, properly normalized, converges in distribution to a standard normal as n → ∞. The "properly normalized" matters: (S_n - n·μ) / (σ·√n) is what converges, not S_n itself. CLT generalizes to non-identically-distributed variables under Lyapunov or Lindeberg conditions, and to martingales and Markov chains under additional structure. The interview signal is whether candidates know which CLT applies, and what breaks (heavy tails: Cauchy distribution has no variance, sample means do not converge to normal).
- What is the difference between independence and uncorrelatedness?
- Independence means P(A ∩ B) = P(A)·P(B), or equivalently the joint distribution factors as the product of marginals. Uncorrelatedness means E[XY] = E[X]·E[Y]. Independence implies uncorrelatedness, but the converse fails — uncorrelated but dependent random variables are easy to construct (X uniform on [-1, 1], Y = X^2: E[XY] = E[X^3] = 0 but Y is determined by X). For jointly Gaussian variables, the two are equivalent, which is why "uncorrelated Gaussians are independent" is the rare exception that interview questions probe.
- How does Bayes' theorem work, and where do candidates trip up?
- Bayes: P(A | B) = P(B | A) · P(A) / P(B). The trap is base-rate neglect — candidates apply the formula correctly but underweight the prior P(A). Classic example: a 99%-accurate medical test for a disease with 0.1% prevalence. Naive intuition says a positive test means 99% chance of disease, but Bayes gives ~9% — because the false-positive rate dominates when the prior is small. Quantitative-research and ML interviews use base-rate problems specifically to test whether candidates compute correctly under cognitive bias.
- What is a moment generating function and why is it useful?
- The MGF of X is M(t) = E[e^(tX)], whenever the expectation exists in a neighborhood of 0. Three properties make it powerful: derivatives at 0 give moments (M^(n)(0) = E[X^n]); MGF of a sum of independent variables is the product of MGFs (which is how you derive the distribution of sums in closed form); MGF uniquely determines distribution when it exists in a neighborhood of 0. The characteristic function E[e^(itX)] is the MGF's always-defined cousin, used in proofs of CLT and inversion theorems.
- How do conditional expectation and the tower property work?
- Conditional expectation E[X | Y] is itself a random variable — a function of Y — defined as the best (in MSE) function of Y for predicting X. The tower property says E[E[X | Y]] = E[X]: averaging the conditional expectation over Y recovers the unconditional. This is the engine behind iterated expectations in finance (price as discounted expected payoff), in dynamic programming (value function as expected future value), and in EM algorithms (E-step computes the conditional expectation of latent variables given observed data).
Related topics
Essential AI-Native Skills for Probability
Modern engineering work increasingly uses AI tools for design and code review, debugging, documentation, test and testbench generation, and workflow automation. The goal is not to let AI replace engineering judgment — it is to move faster while keeping verification discipline.
- Use AI to explain unfamiliar code, logs, waveforms, datasheets, or test failures.
- Break large problems into small, reviewable steps you can verify independently.
- Ask AI for hypotheses, then validate them against tests, measurements, simulations, or lab data.
- Version-control your analysis scripts, testbenches, and configs — keep changes small and reviewable.
- Document your assumptions, design tradeoffs, and debugging decisions.
- Verify AI output before trusting it: run the checks that fit the domain — unit tests, linters, simulations, or bench/lab measurements.
- Review AI output for correctness, edge cases, and real-world consequences.
Probability — coming to the question bank
The adaptive practice engine is already live for core wireless, RF, and ML systems. Probability isn't covered in the question bank yet — get notified when it's added.