Mathematics for Data Science
Data science turns measurements into decisions under uncertainty. The mathematics is not a single theorem but a stack: descriptive summaries, probability models, estimation, hypothesis testing, prediction, and experimental design—implemented at scale with linear algebra and optimization. This part emphasizes statistical foundations with CS systems awareness (pipelines, A/B tests, metrics).
The data science loop (mathematical view)
- Collect samples from a process (ideally i.i.d. or with known dependence)
- Describe with robust summaries and visualizations
- Model \(P(Y\mid X)\) or \(P(X)\) or causal \(P(Y\mid\mathrm{do}(X))\)
- Estimate parameters / predict with uncertainty
- Decide (ship feature? alert? allocate?)
- Validate on held-out data or sequential experiments
Skipping uncertainty quantification produces confident wrong products.
Core mathematical areas
Statistics and probability
- Descriptive statistics and exploratory data analysis (EDA)
- Estimation: bias, variance, MSE, consistency
- Confidence intervals and hypothesis tests
- Bayesian updating
- Bootstrap and resampling
Linear algebra for tabular / embedding data
- Data matrix \(X\in\mathbb{R}^{n\times d}\)
- PCA / SVD for dimensionality reduction
- Least squares and regularized regression
Optimization and calibration
- Fitting models by minimizing empirical risk
- Proper scoring rules (log loss, Brier)
- Threshold selection under class imbalance
Causality and design (often neglected)
- Randomized experiments (A/B tests)
- Confounding in observational data
- Potential outcomes sketch
Worked example 1 — metric vs estimator
“Average revenue per user” is a functional of a distribution. The sample mean estimates it; variance and dependence (users repeat) affect standard errors. Math separates parameter from estimator.
Worked example 2 — classification pipeline
Labels \(y\in\{0,1\}\), features \(x\). Model \(\hat p(x)\approx P(y=1\mid x)\). Threshold \(\hat p\ge t\) yields decisions. Choosing \(t\) is decision theory, not only “accuracy maximization.”
Descriptive → inferential bridge
Descriptive stats summarize the sample you have. Inferential stats make claims about the process that generated it, requiring assumptions (random sampling, model class, independence).
Worked example 3
Mean latency on yesterday’s logs describes yesterday. Inferring next week’s P99 needs a model of traffic mix and nonstationarity—often the hard part.
Roadmap
- Statistical foundations — summaries, estimation, testing, error types, power (detailed chapter)
- Cross-links: Probability advanced (
17), Linear algebra advanced (18), Optimization (11), Information theory (10)
Notation
- Sample \((x_i)_{i=1}^n\) or pairs \((x_i,y_i)\)
- Estimator \(\hat\theta_n\); true \(\theta\)
- Empirical distribution \(\hat P_n=\frac1n\sum\delta_{x_i}\)
- Loss \(\ell\); risk \(R\); empirical risk \(\hat R_n\)
Worked example 4 — empirical risk
\(\hat R_n(f)=\frac1n\sum_i \ell(y_i,f(x_i))\) is a Monte Carlo estimate of \(R(f)=\mathbb{E}[\ell(y,f(x))]\) when data are i.i.d.
Systems constraints
| Constraint | Math impact |
|---|---|
| Streaming data | Online estimators, sketching |
| Multiple testing | Family-wise error, FDR |
| Logging bias | Selection bias; counterfactual issues |
| Delayed outcomes | Censoring; off-policy evaluation |
| Privacy | Noise injection; bias-variance trade |
Worked example 5 — peeking in A/B tests
Repeatedly testing until \(p<0.05\) inflates Type I error. Sequential testing / always-valid p-values fix the math of early stopping.
Estimands, estimators, and error
An estimand is the target quantity (population mean conversion, ATE, \(P(Y=1\mid x)\)). An estimator is a function of the sample (sample mean, difference-in-means, logistic MLE). Quality metrics:
\[ \mathrm{Bias}(\hat\theta)=\mathbb{E}[\hat\theta]-\theta,\qquad \mathrm{Var}(\hat\theta)=\mathbb{E}[(\hat\theta-\mathbb{E}\hat\theta)^2],\qquad \mathrm{MSE}=\mathrm{Bias}^2+\mathrm{Var}. \]
Worked example 6 — biased but useful
Ridge coefficients are biased toward zero yet can lower MSE under collinearity. “Unbiased” is not always the right product goal.
Worked example 7 — standard error
For i.i.d. sample mean, \(\mathrm{se}=\sigma/\sqrt{n}\). Doubling precision (halving se) needs roughly \(4\times\) data—budgeting law of data science.
Supervised learning as statistics
Prediction models estimate functionals of \(P(Y\mid X)\). Classification thresholds implement decision theory; calibration asks whether \(\hat p(x)\approx P(Y=1\mid X=x)\). Cross-validation estimates risk of the selection procedure, not only a fixed model.
Worked example 8 — class imbalance
Accuracy can be \(99\%\) by always predicting “negative” when base rate is \(1\%\). Prefer precision/recall, PR-AUC, or expected cost under a loss matrix.
Causal vs predictive summary table
| Question | Tool |
|---|---|
| What will happen if we observe \(X=x\)? | Predictive model |
| What if we set \(T=1\)? | Causal estimand + design/ID |
| Which features associate with \(Y\)? | Regression / MI (not automatically causal) |
| Did the launch move the metric? | Experiment or quasi-experiment |
Worked example 9 — logging policy
Bandit logs of \(\pi(a\mid x)\) enable off-policy evaluation of a new policy \(\pi'\) via importance weights—if support conditions hold. Pure supervised fit on logged rewards without correction is biased.
Pitfalls
- p-hacking and metric shopping
- Train-test leakage in feature pipelines
- Simpson’s paradox aggregation
- Confusing prediction with causation
- Ignoring dependence (time series, users, graphs)
- Huge \(n\) ≠ no bias (systematic measurement error)
Checkpoint
- Separate description, inference, prediction, causation
- Write risk vs empirical risk
- Name bias/variance of an estimator
- Explain why A/B tests beat naive before/after
- List three pipeline leakages
Exercises
- Define bias and variance of \(\hat\theta\); give an example with nonzero bias.
- Why can accuracy be a bad metric at 1% positive class?
- Write an A/B test design: unit of randomization, primary metric, guardrails.
- Explain bootstrap SE for the median at a high level.
- Give a Simpson’s paradox contingency table (numbers).
- When is MAP estimation “like regularization”?
- PCA: what objective does the first PC optimize?
- Multiple metrics: how does multiple testing arise in dashboards?
- Streaming mean: write Welford update (or online mean formula).
- Causal: why does “users who clicked had higher revenue” not prove clicks cause revenue?
- Calibration: define ECE at a high level.
- Checkpoint project: pick a KPI at work/school; identify estimand, estimator, assumption most likely false.
Summary
Data science math is statistical decision-making with matrices and optimizers attached. Master estimands, estimators, uncertainty, and design—then scale with computation. The next chapter builds statistical foundations in depth.