Written and maintained by CASRAI Editorial Board
Published
On this page
The Gauss-Markov theorem says that, when a linear regression model satisfies a short list of assumptions about its errors, the ordinary least squares (OLS) estimator is the best linear unbiased estimator (BLUE) of the coefficients. “Best” means smallest variance among all estimators that are both linear in the observed outcomes and unbiased. The theorem does not require normally distributed errors, and it says nothing about estimators that are allowed to be biased. This page states the theorem and its assumptions, explains every symbol, works a five-observation example by hand, and shows what each assumption does and does not protect.
What the Gauss-Markov theorem says (OLS is BLUE)
Write the model as y = Xβ + ε. Statlect’s statement is that the OLS estimator “has the smallest variance among those that are unbiased and linear in the observed output variables”. Duke’s POLSCI 630 lecture notes put it the same way: given the necessary assumptions, OLS has the smallest variance among linear, unbiased estimators. Each word in “BLUE” is a separate restriction:
- Best: smallest variance. For the whole coefficient vector, this means that for any fixed 1×K row vector a, the estimator a·β̂ has variance no larger than a·β̃ for every competing linear unbiased estimator β̃; equivalently, Var(β̃) − Var(β̂) is positive semi-definite (Statlect).
- Linear: the estimator is a fixed linear combination of the observed outcomes, β̂ = (X′X)⁻¹X′y.
- Unbiased: its expected value, conditional on X, equals the true β.
- Estimator: a rule for computing a coefficient from data, so the comparison is between rules, not between single samples.
The statement is conditional on the observed values of the regressors. The RUC econometrics lecture on the theorem stresses that both the unbiasedness and the variance in the criterion are conditional on the X values.
Gauss-Markov theorem assumptions
The five assumptions below are the version listed in the Duke lecture notes; Statlect states the same conditions in matrix form (full-rank X, zero conditional mean of the errors, and Var(ε|X) = σ²I with σ² a positive constant).
- Linearity in the parameters. The outcome is a linear function of the coefficients plus an error. The regressors themselves may be transformed (squares, logs) as long as the coefficients enter linearly.
- Random sample. Observations are independent and identically distributed draws.
- No perfect collinearity. X has full column rank, so X′X is invertible and (X′X)⁻¹ exists.
- Exogeneity. E(ε|X) = 0: the errors have zero mean given the regressors.
- Spherical errors. Var(ε|X) = σ²I: every error has the same variance (homoskedasticity) and the errors are uncorrelated with each other.
The RUC lecture states the simple-regression version with four conditions: the linear model Yi = α + βXi + Ui, E(Ui|X1,…,Xn) = 0, E(Ui²|X1,…,Xn) = σ² for all i (homoskedasticity), and E(UiUj|X1,…,Xn) = 0 for i ≠ j. It concludes that, conditionally on the X values, the OLS estimators are BLUE.
Gauss-Markov theorem assumptions: what is not on the list
None of the three sources lists normally distributed errors as an assumption. Normality is a separate condition that matters for exact t and F distributions in small samples, not for the BLUE property. For checks on the remaining assumptions in practice, see assumptions of simple regression and residual checks.
Formula and symbols
With the model y = Xβ + ε, the OLS estimator and its conditional variance under the assumptions above are:
- β̂ = (X′X)⁻¹X′y
- β̂ = β + (X′X)⁻¹X′ε, so E(β̂|X) = β when E(ε|X) = 0
- Var(β̂|X) = σ²(X′X)⁻¹
Symbols: y is the N×1 vector of outcomes; X is the N×K matrix of regressors (including a column of ones when there is an intercept); β is the K×1 vector of true coefficients; ε is the N×1 vector of errors; σ² is the common error variance; X′ is the transpose of X; (X′X)⁻¹ is the inverse of X′X; I is the identity matrix; β̂ is the OLS estimator and β̃ is any rival estimator.
Why no linear unbiased rival can beat OLS
Statlect’s proof has a short core. Any linear estimator can be written β̃ = Cy. Let D = C − (X′X)⁻¹X′, so that β̃ is OLS plus a correction Dy. Unbiasedness for every possible β forces DX = 0. Then Var(β̃|X) − Var(β̂|X) = σ²DD′, which is positive semi-definite, so no rival has a smaller variance and equality holds only when D = 0, that is, when β̃ is OLS itself.
For one slope, the RUC lecture gives the scalar version. Any linear estimator is β̃ = Σ ciYi with ci depending only on the X values. Unbiasedness for every α and β requires Σci = 0 and ΣciXi = 1. Under the assumptions, Cov(β̃, β̂) = Var(β̂), and therefore Var(β̃ − β̂) = Var(β̃) − Var(β̂) ≥ 0. The next section uses exactly these two conditions.
Worked example with real numbers
Take five observations with x = 1, 2, 3, 4, 5 and y = 2, 4, 5, 4, 6. The mean of x is 3 and the mean of y is 4.2.
Step 1: OLS slope. The sum of squared deviations of x is Sxx = (−2)² + (−1)² + 0² + 1² + 2² = 10. The OLS slope is a linear combination of y with weights wi = (xi − 3)/10 = −0.2, −0.1, 0, 0.1, 0.2. So β̂ = −0.2(2) − 0.1(4) + 0(5) + 0.1(4) + 0.2(6) = −0.4 − 0.4 + 0 + 0.4 + 1.2 = 0.8. The intercept is 4.2 − 0.8 × 3 = 1.8.
Step 2: a rival linear unbiased estimator. Use only the two end points: β̃ = (y5 − y1)/(x5 − x1) = (6 − 2)/4 = 1.0. This is the same device the RUC lecture uses with the first two observations. Its weights are ci = −0.25, 0, 0, 0, 0.25. Check the unbiasedness conditions: Σci = 0, and ΣciXi = −0.25(1) + 0.25(5) = 1. So it is linear and unbiased, and nothing in the theorem is violated by its giving a different number (1.0 versus 0.8) in this one sample.
Step 3: compare variances. For any linear estimator Σ ciYi with ΣciXi = 1 and Σci = 0, the conditional variance under homoskedastic, uncorrelated errors is σ² Σci². For OLS: Σwi² = 0.04 + 0.01 + 0 + 0.01 + 0.04 = 0.10, which equals 1/Sxx. For the end-point estimator: Σci² = 0.0625 + 0.0625 = 0.125. The ratio is 0.125 / 0.10 = 1.25, so the end-point rival has a variance 25% larger than OLS, whatever σ² is. The covariance check from the proof also holds: Σciwi = (−0.25)(−0.2) + (0.25)(0.2) = 0.10 = Var(β̂)/σ².
Step 4: more rivals. Using only x = 2 and x = 4 gives weights −0.5, 0.5 and a variance multiplier of 0.5. Using only the first two observations gives weights −1, 1 and a multiplier of 2, twenty times the OLS value. The data table below lists all four estimators. Every one is linear and unbiased; OLS has the smallest multiplier, as the theorem requires.
Step 5: estimate σ² and the standard error. The fitted values are 2.6, 3.4, 4.2, 5.0, 5.8, so the residuals are −0.6, 0.6, 0.8, −1.0, 0.2. Their squares sum to 0.36 + 0.36 + 0.64 + 1.00 + 0.04 = 2.40. With n − 2 = 3 degrees of freedom, s² = 2.40/3 = 0.80, and the standard error of the slope is √(0.80/10) = √0.08 ≈ 0.283.
An important caution: the 1.0 versus 0.8 difference in one sample does not show that OLS is “closer to the truth” here. The theorem is about the spread of each estimator across repeated samples with the same X values, not about any single sample.
Interpretation: what the result means
The table gives reading guidance for each part of the result. Every row restates what the sources say about that piece of the statement.
| Part of the result | What it tells you | What it does not tell you |
|---|---|---|
| Linear | The comparison is limited to estimators of the form Cy (Statlect). | Nonlinear estimators are outside the claim. |
| Unbiased | The comparison is limited to estimators whose conditional expectation is β (Statlect, RUC). | A biased estimator with less variance may sometimes be preferred (Duke). |
| Best (smallest variance) | No linear unbiased rival has a smaller variance; the gap Var(β̃) − Var(β̂) is positive semi-definite (Statlect). | Nothing about a single sample’s error, or about rivals under other error structures. |
| Conditional on X | The criterion holds for the observed regressor values (RUC). | It is not a claim over random redraws of X. |
| Holds only if the assumptions hold | Under heteroskedastic errors, other linear unbiased estimators with lower variance exist (Duke). | It does not say OLS is then useless, only that it is no longer guaranteed best. |
Assumptions and common mistakes
Mistake 1: reading “best” as “best among all estimators”
The Duke notes state that OLS is best only among unbiased linear estimators, and that a biased estimator with lower variance may sometimes be preferable. Penalised methods such as ridge or the lasso accept bias to reduce variance, which is exactly the trade the theorem leaves open. See elastic net regression for one such method.
Mistake 2: treating normality as part of the theorem
The assumption lists in all three sources contain no distributional form for the errors. Only the first two moments are used: zero mean and constant variance with zero covariance.
Mistake 3: assuming heteroskedasticity makes OLS biased
Unbiasedness in the proof rests on E(ε|X) = 0, via β̂ = β + (X′X)⁻¹X′ε. Non-constant variance leaves that step intact, but it breaks the variance formula σ²(X′X)⁻¹ and the “best” claim: with heteroskedastic errors, other linear unbiased estimators have lower variance (Duke). The usual standard errors are then unreliable; see robust (heteroscedasticity-consistent) standard errors and how to read a residual plot.
Mistake 4: ignoring correlated errors
Spherical errors also require zero covariance between errors. Time series and clustered data commonly break this; see the Durbin-Watson test for autocorrelated residuals and clustered standard errors.
Mistake 5: missing the rank condition
If one regressor is an exact linear combination of the others, X′X is not invertible and the OLS formula does not exist; the theorem’s full-rank assumption excludes this. Near-collinearity is a different problem, covered in multicollinearity and VIF in regression.
Which assumption protects which property
| Assumption | Role in the proof |
|---|---|
| Full rank of X | Makes β̂ = (X′X)⁻¹X′y exist (Statlect). |
| E(ε|X) = 0 | Gives E(β̂|X) = β, the unbiasedness step (Statlect, RUC). |
| Var(ε|X) = σ²I | Gives Var(β̂|X) = σ²(X′X)⁻¹ and Var(β̃) − Var(β̂) = σ²DD′ (Statlect). |
| Linearity in parameters | Lets the model be written y = Xβ + ε (Duke, Statlect). |
How to report it (APA example)
The theorem is a property of the method, so you do not report it as a result; you report that the assumptions were examined. A sentence in APA 7 style might read: “We estimated the model by ordinary least squares. Residual plots showed no pattern of changing spread, and the Durbin-Watson statistic did not indicate autocorrelation; the Gauss-Markov conditions for OLS were therefore treated as reasonable.” Report the coefficient with its standard error and interval in the usual form, for example “b = 0.80, SE = 0.28, 95% CI [−0.10, 1.70]”, and state the robust standard error type if you used one. The interval in this example uses t(3) = 3.182: 0.80 ± 3.182 × 0.283 = [−0.10, 1.70]. Report the decision, not just the test; see how to read a regression table for the other columns.
How to calculate it in Excel, R or Python
There is no single “Gauss-Markov statistic” to compute. What you compute is the OLS fit and its diagnostics. The snippets below reproduce the worked example.
Excel
=SLOPE(B2:B6, A2:A6) returns 0.8 =INTERCEPT(B2:B6, A2:A6) returns 1.8 =INDEX(LINEST(B2:B6, A2:A6, TRUE, TRUE), 2, 1) returns the slope standard error, about 0.283
R
x <- 1:5 y <- c(2, 4, 5, 4, 6) fit <- lm(y ~ x) summary(fit)$coefficients # slope 0.8, SE about 0.283 sum(resid(fit)^2) # 2.4
Python
import numpy as np x = np.array([1, 2, 3, 4, 5.0]) y = np.array([2, 4, 5, 4, 6.0]) X = np.column_stack([np.ones_like(x), x]) beta = np.linalg.solve(X.T @ X, X.T @ y) # (X'X)^-1 X'y -> [1.8, 0.8] resid = y - X @ beta s2 = resid @ resid / (len(y) - 2) # 0.8 cov = s2 * np.linalg.inv(X.T @ X) # sigma^2 (X'X)^-1 print(beta, np.sqrt(np.diag(cov)))
To see the theorem in the arithmetic, compute the weights of any rival and its sum of squares, as in the worked example: for weights c, check that sum(c) is 0, that sum(c * x) is 1, and that sum(c**2) is at least 0.10 for these x values.
FAQ
What is the Gauss-Markov theorem in simple terms?
It states that, if the errors in a linear regression have zero mean given the regressors and equal, uncorrelated variances, OLS gives the most precise estimates of any method that is linear in the outcomes and unbiased.
What does BLUE stand for?
Best linear unbiased estimator: smallest variance among estimators that are linear in the observed outcomes and unbiased (Statlect, RUC).
What are the Gauss-Markov assumptions?
Linearity in the parameters, a random sample, no perfect collinearity, E(ε|X) = 0, and spherical errors, Var(ε|X) = σ²I (Duke).
Does the Gauss-Markov theorem require normal errors?
No. The assumption lists in the sources contain no distributional form for the errors.
What happens if homoskedasticity fails?
Other linear unbiased estimators then have lower variance than OLS (Duke), and the standard OLS variance formula no longer applies. Robust standard errors address inference.
Are Gauss-Markov, Markov-Gauss and the “Markov theorem” the same thing?
Searches for “gauss markov”, “gauss-markov”, “markov gauss theorem” and “markov theorem” in a regression context all refer to this result on OLS. It is unrelated to Markov chains.
Can a biased estimator beat OLS?
Yes, in mean squared error terms. The theorem only ranks unbiased linear estimators, and Duke notes that a biased estimator with less variance may sometimes be preferred.
Sources
- Statlect, “Gauss Markov theorem”, statlect.com/fundamentals-of-statistics/Gauss-Markov-theorem (read 2026-11-12).
- Duke University POLSCI 630, “OLS Assumptions & Properties”, lecture 3 (read 2026-11-12).
- Renmin University of China, Introductory Econometrics, “Lecture: Gauss-Markov theorem”, lecture notes PDF (read 2026-11-12).
Related reading: regression analysis: assumptions, interpretation and reporting.
Data: Weights and variance multipliers of four linear unbiased slope estimators (x = 1 to 5)
For the regressor values x = 1, 2, 3, 4, 5, each row gives a linear unbiased slope estimator as weights on y, checks the unbiasedness conditions (weights sum to 0, weights times x sum to 1) and reports the variance as a multiple of sigma squared (sum of squared weights) plus the estimate for y = 2, 4, 5, 4, 6. Compiled by the authors from the estimator form and conditions in the Statlect and RUC lecture treatments of the Gauss-Markov theorem.
| Estimator | Weights c1 to c5 on y1 to y5 | Sum of weights | Sum of weights times x | Variance multiplier (times sigma squared) | Slope estimate for y = 2, 4, 5, 4, 6 |
|---|---|---|---|---|---|
| OLS (all five points) | -0.2, -0.1, 0, 0.1, 0.2 | 0 | 1 | 0.1 | 0.8 |
| End points only (x=1, x=5) | -0.25, 0, 0, 0, 0.25 | 0 | 1 | 0.125 | 1 |
| Middle pair (x=2, x=4) | 0, -0.5, 0, 0.5, 0 | 0 | 1 | 0.5 | 0 |
| First two points (x=1, x=2) | -1, 1, 0, 0, 0 | 0 | 1 | 2 | 2 |
Sources
- Statlect, Gauss Markov theorem (accessed 2026-11-12)
- RUC Introductory Econometrics, Lecture on the Gauss-Markov theorem (accessed 2026-11-12)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







