The recurring move goes like this: you run the numbers, Pearson's r comes back at 0.02, and you conclude the two variables aren't related — maybe you even drop one as a predictor or call them "independent." It's a tempting jump, because a single number sitting near zero looks like a clean verdict on the whole relationship.
But it isn't. A Pearson correlation near zero means there's no straight-line trend between your two variables. It does not mean they're unrelated, and it does not mean they're independent. Pearson's r only measures the linear part of a relationship. A strong curved pattern — a U-shape, a parabola, a sine wave — can produce r ≈ 0 because its rising half and falling half cancel each other out. Before you drop a predictor because "the correlation was basically zero," plot the data. The plot will often show a relationship the single number can't.
Why r = 0 feels like "no relationship"
The trap is built into the vocabulary. We say "no correlation" out loud, and in everyday English "correlation" just means "connection." So r = 0 gets quietly translated to "no connection," and "no connection" gets translated to "these variables have nothing to do with each other."
It doesn't help that Pearson's r is usually the only association number a beginner has been handed. It's one tidy value between -1 and 1, it ships in every spreadsheet and every df.corr() call, and most courses introduce it without ever stressing the word doing all the work: linear. This is one of a family of single-number summaries that quietly throw away information — the same trap shows up in other beginner traps with probability and descriptive statistics. So when the number comes back as 0.02, it feels like a verdict on the whole relationship rather than a verdict on one specific shape — the straight line.
This is the same lesson Anscombe's quartet teaches: four datasets with nearly identical summary statistics, including correlation, that look completely different once you plot them. The number is not the data. It's a one-dimensional shadow of it.
The worked example: a parabola with r = 0.00
Here is a case where x determines y perfectly and Pearson's r is still exactly zero.
Take eleven points, x from -5 to 5, and let y = x²:
| x | -5 | -4 | -3 | -2 | -1 | 0 | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| y | 25 | 16 | 9 | 4 | 1 | 0 | 1 | 4 | 9 | 16 | 25 |
There is no noise here. Plot it and you get a clean, symmetric U. Knowing x tells you y with certainty.
Now compute the correlation. Pearson's r is built from the covariance numerator:
$$\sum (x_i - \bar{x})(y_i - \bar{y})$$
The mean of x is 0 (the values are symmetric around zero), so $x_i - \bar{x}$ is just $x_i$ itself: -5, -4, ..., 4, 5.
The mean of y is (25+16+9+4+1+0+1+4+9+16+25) / 11 = 110 / 11 = 10. So $y_i - \bar{y}$ runs 15, 6, -1, -6, -9, -10, -9, -6, -1, 6, 15.
Multiply each pair $(x_i)(y_i - \bar{y})$ and add them up:
- (-5)(15) = -75
- (-4)(6) = -24
- (-3)(-1) = 3
- (-2)(-6) = 12
- (-1)(-9) = 9
- (0)(-10) = 0
- (1)(-9) = -9
- (2)(-6) = -12
- (3)(-1) = -3
- (4)(6) = 24
- (5)(15) = 75
Sum: (-75 - 24 + 3 + 12 + 9 + 0 - 9 - 12 - 3 + 24 + 75) = 0.
The numerator is exactly zero, so r = 0.00. The left half of the parabola has a falling trend, the right half has a rising trend, and because the data are symmetric about x = 0 they cancel term for term. The straight line of best fit through this U is perfectly flat — slope zero — and Pearson's r is the standardized version of that slope. A flat line gives r = 0, no matter how tightly the points hug the curve.
So you have r = 0.00 sitting next to a scatterplot that screams "obvious relationship." Both readings are correct. They're answering different questions.
What actually captures the relationship here
Pearson missed it. So would Spearman's rank correlation, which is also about 0 for this dataset — Spearman detects monotonic relationships (always increasing or always decreasing), and a U-shape isn't monotonic. It goes down, then up. Both linear and monotonic measures are blind to this pattern for the same reason: the two halves disagree about direction.
Two things see it clearly:
Add the curve to your model. Regress y on both x and x². With the squared term included, the fit is exact — R² = 1.0. You recover the fact that y is fully determined by x, because now you're fitting the right shape. The relationship was never weak; you were measuring it with a straight ruler.
Use a dependence measure that doesn't assume a shape. Distance correlation is one. It's zero only when two variables are genuinely independent, and for the parabola above it comes back high. That's the precise statement r = 0 was never making: r = 0 says "no linear trend," while distance correlation = 0 says "no relationship of any kind."
That distinction is the whole point. r = 0 is a claim about straight lines. Independence is a claim about everything — and independence is a stricter, more slippery condition than it looks, the same way a positive test result doesn't give you the probability of disease just because the two quantities sit next to each other.
How to not get fooled
Plot first, always. A scatterplot takes one line of code and catches every shape a correlation coefficient throws away. If you only ever look at the number, you are trusting that the relationship — if there is one — happens to be a straight line.
When r comes back near zero, ask what shape you'd expect on physical grounds. Dose and effect that helps in small amounts and harms in large ones, error that's lowest at some optimal setting and rises on both sides, anything with a sweet spot — these are U-shaped or inverted-U by nature, and Pearson will report them as "nothing."
If the plot shows structure but r is small, reach for the tool that matches the structure: a polynomial or spline term in a regression, Spearman for monotonic-but-curved, distance correlation or mutual information for "is there any dependence at all." Don't drop the predictor on the strength of one number that was only ever testing for a line.
Check yourself
You compute Pearson's r between two variables and get r = 0.01. What can you correctly conclude?
A) The two variables are independent. B) The two variables have no relationship. C) There is no strong linear trend; a curved relationship may still exist. D) You made a calculation error, because related variables always have nonzero r.
Correct answer: C.
r ≈ 0 tells you the best straight-line fit is essentially flat. That's all. The variables could be independent, or they could be locked in a perfect U-shape like y = x². A rules out a relationship the number can't rule out. B makes the same leap from "no line" to "no relationship." D is backwards — a strong non-monotonic relationship reliably produces r near 0, as the parabola shows. Only a scatterplot, or a shape-agnostic measure, can separate these cases.
Close the gap
The slip here isn't bad arithmetic — the r = 0.00 is correct. It's reading a number as the answer to a bigger question than it asked. Knowing which relationship a statistic can and can't see, and reaching for the right one when the default comes up empty, is the kind of judgment that's easy to state and hard to build alone. That's what Gradual Learning is designed to teach: not just the formula, but where it quietly stops applying.