Null and Alternate Hypotheses

This concept is the foundation of hypothesis testing, which you'll use to evaluate whether patterns you see in data are real or just noise. It unlocks: Significance Levels, Hypothesis Testing, ANOVA, Cross-Validation, and beyond. In AI practice, you'll use it to decide whether a new model actually outperforms the old one, whether a feature matters, or whether an A/B test shows a genuine improvement.

Suppose you run a coffee shop and you're thinking of switching to a new espresso blend. You don't know if the new blend is better, worse, or the same. To find out, you'd compare: the new blend has some effect (better or worse) or no effect (same as before). The 'no effect' statement is called the null hypothesis — it's the default, the boring assumption that nothing changed. The 'some effect' statement is the alternative hypothesis. In any test, you start by assuming the null is true, then you look at the data. If the data is so unlikely under the null, you reject the null and accept the alternative. It's like a courtroom: the null is 'innocent until proven guilty,' and you need strong evidence to convict (reject). The key is that you never 'prove' the null; you either reject it or fail to reject it.

Imagine you go to a doctor with a rare disease. The doctor runs a test. The null hypothesis is that you are healthy—no disease. The alternative is that you have the disease. The test is designed to detect the disease. If the test comes back positive, you might reject the null and conclude you have the disease (but there's a chance of a false positive). If the test comes back negative, you fail to reject the null—you're assumed healthy, but it's not proof you're disease-free (the test might miss it). This analogy works because the null is the safe default, and we only abandon it with strong evidence. But it stops working: in science, we never accept the null as true; we simply say we lack evidence against it. Also, the 'evidence' here is statistical, not absolute.

The null hypothesis is a default statement of 'no effect' or 'no difference', and the alternative hypothesis is the statement you accept if the data provides enough evidence against the null.

You haven't studied other concepts yet, but this is the entry point to statistical inference. It builds on everyday ideas of probability (we'll touch on that) and arithmetic. It leads directly to Significance Levels (how much evidence is 'enough') and Hypothesis Testing (the full procedure). It also connects to ANOVA (testing multiple groups) and Cross-Validation (estimating performance), which you'll see later.