The Standard Error of the Mean

Derivation

The Standard Error of the Mean

D-112 Home PU-106 Threads chance Depends on Variance of Linear Combinations and Covariance

Statement

For N independent measurements x₁, x₂, …, x_N drawn from a common distribution of mean μ and variance σ², the sample mean x̄ = (1/N)∑_i x_i is unbiased, E[x̄] = μ, and has variance Var(x̄) = σ²/N. Its standard deviation — the standard error of the mean — is therefore SE = σ/√N, so the uncertainty on the mean shrinks as N^−1/2.

Why it matters

This single result is the reason averaging works. It is the quantitative statement that combining many noisy measurements yields an estimate more precise than any one of them, and it fixes the exchange rate: to halve your uncertainty you must quadruple your data. Every reported laboratory value of the form “value ± error” on a mean rests on it.

It also draws a sharp line between two different quantities that beginners routinely confuse. The spread of the raw data, σ, does not shrink as you take more measurements — the population is what it is. What shrinks is the spread of the mean, and it does so only because independent errors partially cancel. The 1/√N law is the fingerprint of that cancellation.

Assumptions

Each measurement has the same finite variance σ².If the variance is not identical across measurements, the simple σ²/N collapses to (1/N²)∑_i σ_i², and the 1/√N scaling holds only on average; if σ² is infinite (e.g. a Cauchy distribution) the mean has no smaller spread than a single sample.

The measurements are pairwise uncorrelated: Cov(x_i, x_j) = 0 for i ≠ j.If the errors are correlated, cross terms survive and add a term (1/N²)∑_i≠j Cov(x_i,x_j); positively correlated noise (e.g. common drift) can make the variance floor out at a nonzero value no matter how large N grows.

The variance of a sum of uncorrelated variables is the sum of their variances (the prior result invoked in Step 3).This is exactly where independence enters; without it the linearity-of-variance identity Var(∑a_ix_i) = ∑a_i²Var(x_i) is false and the whole derivation stalls at the cross terms.

The population mean μ and variance σ² exist and are the same for every draw (identically distributed).If the distribution changes between measurements, “the” mean being estimated is ill-defined and x̄ converges to a time-average that need not equal any single μ_i.

Derivation

x̄ = (1/N) ∑_i=1^N x_i

Definition of the sample mean; write it as a linear combination with equal weights a_i = 1/N so the prior result applies. A

E[x̄] = (1/N) ∑_i=1^N E[x_i] = (1/N)(Nμ) = μ

Linearity of expectation (no independence needed here); each E[x_i] = μ. This shows the estimator is unbiased, so its variance is a fair measure of its error. A

Var(x̄) = Var( (1/N) ∑_i x_i ) = (1/N)² ∑_i=1^N Var(x_i)

Apply the prior result, variance of a linear combination of uncorrelated variables: Var(∑ a_ix_i) = ∑ a_i² Var(x_i) with a_i = 1/N. The constant 1/N comes out squared. B

= (1/N²) ∑_i=1^N σ² = (1/N²)(Nσ²)

Each term is identical, Var(x_i) = σ² (identically distributed), so the sum is just N copies. This is the step that fails silently if the x_i have unequal variances. B

Var(x̄) = σ²/N

Cancel one factor of N. Symbols only — no numbers yet. A

SE ≡ √Var(x̄) = σ/√N

Take the positive square root to convert variance (units of the quantity squared) into a standard deviation with the same units as x. This is the standard error of the mean. A

Result

Var(x̄) = σ²/N ⇒ SE = σ/√N

Reading. Averaging N independent measurements does not change the typical size of a single error, σ, but it reduces the error on the average by a factor √N. The gain has diminishing returns: going from 1 to 4 measurements halves the error, but going from 100 to 400 also only halves it. The mean converges to the true value μ, but slowly.

Units check. σ carries the units of x (say metres). N is a pure count, dimensionless, so √N is dimensionless. Hence SE = σ/√N has the units of x (metres), exactly as an uncertainty on a mean must. The variance σ²/N carries metres², consistent with being a variance.

Limiting cases

N = 1: SE = σ. A single measurement’s uncertainty is the population spread — no averaging, no reduction.
N → ∞: SE → 0. The sample mean converges (in mean square) to μ; this is the weak law of large numbers made quantitative.
Quadrupling data: N → 4N gives SE → SE/2. Fixed factor-of-2 improvement per factor-of-4 in effort.
σ → 0: SE = 0 for any N. Noiseless measurements already agree; averaging adds nothing.

Breaks when

Correlated errors. If measurements share a common fluctuation (thermal drift, a mis-calibrated instrument, 1/f noise), Cov(x_i,x_j) ≠ 0. For uniform positive correlation ρ, Var(x̄) = σ²[1/N + ρ(N−1)/N] → ρσ² as N→∞: the error floors out at σ√ρ and averaging stops helping. This is the single most common way the law fails in real labs.
Heavy-tailed / infinite-variance distributions. For a Cauchy (Lorentzian) distribution σ² does not exist, and remarkably the sample mean of N Cauchy variables is itself Cauchy with the same width — averaging gives no improvement at all. The derivation breaks at Step 4 where Var(x_i) = σ² is assumed finite.
Systematic error. The result governs only random (statistical) error. A constant bias b in the instrument survives averaging untouched: E[x̄] = μ + b. Driving SE to zero with huge N makes a biased result look arbitrarily precise while remaining wrong.

Failure modes

Reporting σ as the error on the mean. Quoting the sample standard deviation instead of σ/√N overstates the uncertainty on x̄ by a factor √N. The two answer different questions: “how spread is the data?” vs “how well do I know the mean?”
Dividing by N instead of √N. Forgetting the square root when converting from variance to standard error. Variance goes as 1/N; the error goes as 1/√N.
Believing more data always helps. Ignoring correlation or a systematic floor, then trusting an SE that has stopped shrinking in reality.
Using N instead of N−1 in the estimate of σ. When σ is unknown you estimate it from the same data with Bessel’s correction s² = (1/(N−1))∑(x_i−x̄)². The 1/N in the SE and the 1/(N−1) in s² are separate; conflating them is a classic slip.
Treating N as the number of readings when they are not independent. Sampling faster than the noise correlation time inflates N without adding information; the “effective” N is smaller.

Discussion

The deep content of the derivation lives entirely in Step 3, and it is worth naming exactly what does the work. When you add N random quantities, their means add linearly (Nμ) but their spreads add in quadrature (√N σ). Dividing the sum by N to form the mean scales the signal by 1/N and the noise by 1/N × √N = 1/√N. The whole 1/√N law is nothing but the mismatch between how signals (linearly) and independent noises (in quadrature) combine.

Physically, the √N in quadrature is the signature of a random walk. The sum ∑(x_i−μ) of independent zero-mean deviations is a random walk of N steps, whose typical displacement grows as √N — the same √N that governs Brownian diffusion, shot noise, and Johnson–Nyquist noise. Standard error, diffusion length, and photon-counting statistics are one piece of mathematics wearing three costumes. In particular, for counting N independent events (Poisson statistics) σ = √N, so the fractional error SE/N = 1/√N — the familiar “count longer to beat down the noise” of every detector.

There is a subtlety in which convergence the result guarantees. Var(x̄) = σ²/N → 0 is convergence in mean square, which implies convergence in probability — the weak law of large numbers. It says nothing on its own about the shape of the distribution of x̄. That is supplied separately by the central limit theorem, which states that (x̄−μ)/(σ/√N) tends to a standard normal for large N, regardless of the parent distribution (given finite variance). The two results are independent: the SEM fixes the width of the sampling distribution, the CLT fixes its shape. Only together do they license the ubiquitous “μ = x̄ ± σ/√N at 68% confidence”.

Common misconceptions. (i) “More measurements make the data less spread out.” No — σ is a property of the population and is unchanged; only the spread of the mean shrinks. (ii) “The SEM tells you the true value is within σ/√N.” It gives a confidence interval, not a guarantee, and only for random error — a systematic bias is invisible to it. (iii) “Averaging can beat any noise.” Only uncorrelated, finite-variance noise; correlated or heavy-tailed noise defeats it.

Worked examples

Example 1 — Timing a pendulum. A student measures the period of a pendulum N = 25 times. The readings have a sample standard deviation s = 0.20 s (taken as an estimate of σ). Find the standard error on the mean period.

SE = σ/√N

Start from the result; keep symbols. A

SE = s/√N = (0.20 s)/√25

Substitute the estimate s for the unknown σ, and N = 25. A

= (0.20 s)/5 = 0.040 s

√25 = 5; divide. A

SE = 0.040 s

Reading. Although any single timing scatters by ~0.20 s, the mean period is known to about 0.04 s — five times better — because 25 independent errors partly cancel. To halve this to 0.02 s the student would need N = 100 readings.

Example 2 — Counting radioactive decays (Poisson). A detector records N = 10 000 counts in a fixed interval. For Poisson counting the variance of the count equals the mean, so σ = √N. Find the absolute and fractional uncertainty on the count.

σ = √N

Poisson property: for counting statistics the standard deviation of a single measured count of N events is √N. (Here the “measurement” is the total count itself.) B

σ = √10000 = 100 counts

Substitute the number. A

σ/N = √N / N = 1/√N = 1/100 = 0.01

Form the fractional error; it falls as 1/√N — the same law. A

N = 10 000 ± 100 counts (1% relative)

Reading. To reach 0.1% precision (fractional error 1/√N = 0.001) you need N = 10⁶ counts — a hundredfold longer count for a tenfold better precision. This is the practical face of 1/√N in every counting experiment.

Problems

A voltmeter reading has population standard deviation σ = 3.0 mV. How many independent readings must be averaged so that the standard error on the mean is at most 0.5 mV?
Solution
Require σ/√N ≤ 0.5 mV, i.e. √N ≥ σ/0.5 = 3.0/0.5 = 6, so N ≥ 36. You need at least 36 readings. (Note the quadratic cost: to reach 0.25 mV would take 144 readings.)
Two students each measure a length 100 times and quote the mean. Student A reports the standard error as 0.5 cm; student B reports 0.5 cm as the “error” but has actually quoted the sample standard deviation. What is B’s true standard error, and by what factor did B overstate the uncertainty on the mean?
Solution
B’s s = 0.5 cm is the population spread. The true SE is s/√N = 0.5/√100 = 0.5/10 = 0.05 cm. B overstated the uncertainty on the mean by a factor √N = √100 = 10. A’s figure is the correctly computed standard error.
Starting from Var(x̄) = (1/N²)∑_i,j Cov(x_i,x_j), suppose every pair has equal correlation Cov(x_i,x_j) = ρσ² for i≠j and Var(x_i) = σ². Derive Var(x̄) and find its limit as N→∞.
Solution
The double sum has N diagonal terms each σ² and N(N−1) off-diagonal terms each ρσ². So Var(x̄) = (1/N²)[Nσ² + N(N−1)ρσ²] = σ²[1/N + ρ(N−1)/N]. As N→∞, 1/N→0 and (N−1)/N→1, giving Var(x̄) → ρσ². The error floors at σ√ρ — averaging cannot beat correlated noise. Setting ρ=0 recovers σ²/N.
A signal V = 2.0 mV sits on noise of standard deviation σ = 8.0 mV, so a single sample has signal-to-noise ratio V/σ = 0.25. How many independent samples must be averaged for the SNR of the mean, V/(σ/√N), to reach 5?
Solution
SNR of the mean is V√N/σ = (V/σ)√N = 0.25√N. Require 0.25√N = 5, so √N = 20 and N = 400. You need 400 samples. SNR grows as √N because the signal averages coherently while the noise averages incoherently.
A measurement has random error with σ = 0.10 units and an unknown constant systematic bias b = 0.30 units. You average N = 400 readings. Give (a) the standard error and (b) the actual total offset of the reported mean from the true value. Comment on what averaging did and did not achieve.
Solution
(a) SE = σ/√N = 0.10/√400 = 0.10/20 = 0.005 units. (b) The random part is beaten down to 0.005, but the systematic bias is unaffected by averaging: E[x̄] = μ + 0.30, so the mean is offset from the truth by 0.30 units — sixty times the quoted SE. Averaging shrank the statistical error but left the bias entirely intact; the result now looks extremely precise (±0.005) while being badly inaccurate. No amount of N fixes a systematic error — only calibration does.