What is the Mahalanobis distance and how is it calculated?

It is the distance of a point x from the centre μ of a dataset, measured in units that account for how spread out and correlated the features are. The formula is D² = (x − μ)ᵀ Σ⁻¹ (x − μ) and D = √(D²), where Σ is the covariance matrix and Σ⁻¹ its inverse. The inverse covariance shrinks high-variance directions and rotates out correlations, so every direction is weighted fairly.

How does Mahalanobis distance differ from Euclidean distance?

Euclidean distance, √(Σ (xᵢ − μᵢ)²), treats every feature equally and ignores correlation, so a wide-ranging feature dominates. Mahalanobis inserts Σ⁻¹ between the differences, which standardises each feature and removes correlation. When the features are uncorrelated and have unit variance the two match; otherwise Mahalanobis is the scale-invariant, correlation-aware version. This tool shows both side by side.

How do you use Mahalanobis distance to detect outliers?

Compute D² for the point, then compare it to a chi-square threshold. Under multivariate normality D² follows a χ² distribution with p degrees of freedom, where p is the number of features. If D² exceeds the critical value χ²(p, 1 − α) the point is flagged as an outlier at significance α. The calculator reports the critical value, the p-value, and a plain-language verdict.

What chi-square threshold flags a multivariate outlier?

It depends on the number of features p and the significance level α. For p = 2 features at α = 0.05 the threshold is χ²(2, 0.95) = 5.991, so a squared Mahalanobis distance above 5.991 is an outlier. For p = 1 it is 3.841, for p = 3 it is 7.815. The tool computes χ²(p, 1 − α) for any p and for α of 0.10, 0.05, or 0.01.

Why must the covariance matrix be invertible for Mahalanobis distance?

The formula uses Σ⁻¹ directly, so Σ must be non-singular. It becomes singular when two or more features are perfectly correlated (collinear), when a feature is constant, or when there are fewer observations than features (n ≤ p). In those cases the inverse does not exist and the distance is undefined. The calculator detects this and explains the cause instead of returning a misleading number.

Should I use the sample (n − 1) or population (n) covariance?

Use the sample estimator (dividing by n − 1) when your rows are a sample and you want an unbiased covariance — this matches numpy.cov. Use the population estimator (dividing by n) when the rows are the whole population or you want the maximum-likelihood estimate — this matches scikit-learn's EmpiricalCovariance. The Mahalanobis distances differ slightly between the two; the tool offers both and states which is active.

Does the Mahalanobis distance reduce to the z-score in one dimension?

Yes. With a single feature Σ is just the variance σ², so D = |x − μ| / σ — the absolute z-score. For the dataset [2, 4, 4, 4, 5, 5, 7, 9] with mean 5 and population σ = 2, the point 9 gives D = |9 − 5| / 2 = 2.0. This one-dimensional identity is a useful check that the covariance-aware maths is behaving.

Does this calculator send my data anywhere?

No. Parsing the dataset, estimating μ and Σ, inverting the covariance, and evaluating the quadratic form all run in your browser with plain JavaScript. Nothing is uploaded, logged, or stored, so you can paste real feature vectors or sensor readings without concern. The page keeps working offline once it has loaded.

AI · Machine learning

Mahalanobis Distance Calculator

Measure how far a point sits from a multivariate dataset, adjusted for the spread and correlation of its features. Get the Mahalanobis distance, the squared distance, a chi-square outlier verdict with p-value, and a Euclidean comparison — with every matrix shown. No signup, nothing uploaded.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 16, 2026

Mahalanobis distance calculator

Mode

Distance of one point x from the dataset's centre μ, with the χ² outlier test.

Dataset — one observation per line

Rows are observations, columns are features. Separate values with commas, spaces, or tabs. Needs at least 2 rows, and more rows than features (n > p).

Query point x

Same number of values as the dataset has columns.

Covariance estimator

Significance level α

Examples

Outlier at α = 0.05

D² = 6.2500 exceeds the χ²₂ critical value 5.9915 at 95% confidence. p-value = 0.0439 — the chance a genuine sample from this distribution sits at least this far from the centre.

Mahalanobis D

2.5000

2-D · 5 obs

Squared (D²)

6.2500

dᵀ Σ⁻¹ d

Euclidean

3.6056

Uncorrected ‖d‖₂

p-value

0.0439

vs α = 0.05

Decimals

Cross-check. The explicit-inverse quadratic form dᵀ Σ⁻¹ d gives D² = 6.2500; the independent route — solving Σ y = d and taking dᵀ y, the way SciPy computes it — gives 6.2500. They reconcile, as they must.

Mean, covariance and its inverse

Mean vector μ (2 features)

0.0000

Difference d = x − μ (2 features)

3.0000

2.0000

Covariance Σ (2 features)

f1	4.0000	0.0000
f2	0.0000	1.0000

Inverse covariance Σ⁻¹ (2 features)

f1	0.2500	0.0000
f2	0.0000	1.0000

Method: D² = (x − μ)ᵀ Σ⁻¹ (x − μ); μ and Σ from the dataset (n − 1 divisor); outlier test D² ~ χ² with p degrees of freedom — Mahalanobis (1936), scikit-learn, NIST e-Handbook. Nothing leaves this page.

How it works

The Mahalanobis distance, introduced by P. C. Mahalanobis in 1936, measures the distance from a point x to the centre μ of a distribution, weighted by the inverse covariance matrix so that spread and correlation are taken into account. Unlike Euclidean distance it is scale-invariant per feature and aware of how features move together, which is why it is the standard multivariate outlier and anomaly metric.

D²(x) = (x − μ)ᵀ Σ⁻¹ (x − μ), D = √(D²)

The calculator works in five steps:

Mean vector. Average each feature column across the n observations to get μ.
Covariance matrix. Build Σ_jk = (1/D) Σ (x_ij − μ_j)(x_ik − μ_k), using divisor D = n − 1 for the sample estimator (matches numpy.cov) or D = n for the population estimator (matches scikit-learn EmpiricalCovariance).
Invert the covariance. Compute Σ⁻¹ by Gauss–Jordan elimination with partial pivoting. To stay accurate when features have very different scales, the tool factors Σ into standard deviations and a correlation matrix, inverts the well-conditioned correlation matrix, then scales back. If the determinant is effectively zero — collinear features, a constant feature, or n ≤ p — the matrix is singular and the tool stops with an explanation.
Quadratic form. With d = x − μ, the squared distance is D² = dᵀ Σ⁻¹ d and D = √(D²). In point-to-point mode the same Σ⁻¹ is applied to d = a − b.
Outlier test. Under multivariate normality D² follows a chi-square distribution with p degrees of freedom (NIST e-Handbook), so the tool reports the critical value χ²(p, 1 − α) and the p-value P(χ²_p > D²), computed from the regularized incomplete gamma function, and flags the point when D² exceeds the critical value.

As a credibility check the calculator computes D² a second way — by solving the linear system Σ y = d and taking dᵀ y, the route SciPy prefers over forming the inverse explicitly — and confirms the two agree to floating-point precision. All arithmetic is double-precision and rounded only for display.

Worked examples

Point-to-distribution outlier test (2 features), sample covariance

Dataset (X, Y): (−2,−1) (−2,1) (0,0) (2,−1) (2,1), so μ = (0, 0)
Sample covariance Σ = [[4, 0], [0, 1]] — X spreads 4× wider, no correlation
Inverse Σ⁻¹ = [[0.25, 0], [0, 1]]; query x = (3, 2), so d = (3, 2)
D² = 0.25·3² + 1·2² = 2.25 + 4 = 6.25, D = √6.25 = 2.5
χ²(2, 0.95) = 5.991; 6.25 > 5.991 → outlier, p-value = e^(−3.125) = 0.0439
Euclidean for contrast = √(9 + 4) = 3.606 — no verdict, over-weights X

Point-to-point under the shared covariance

Points a = (2, 1), b = (−2, −1); difference d = (4, 2)
Same Σ⁻¹ = [[0.25, 0], [0, 1]] from the dataset above
D² = 0.25·4² + 1·2² = 4 + 4 = 8, D = √8 = 2.828
Euclidean = √(16 + 4) = √20 = 4.472
Mahalanobis is smaller — the gap lies along the wide X axis, down-weighted

One dimension — the z-score identity (edge case)

Dataset [2, 4, 4, 4, 5, 5, 7, 9], single feature, mean μ = 5
Sum of squared deviations = 9 + 3 + 0 + 4 + 16 = 32
Population σ² = 32 / 8 = 4, so σ = 2 (population estimator)
Query x = 9: D = |9 − 5| / 2 = 2.0 — exactly the absolute z-score
Sample estimator σ² = 32 / 7 gives σ = 2.138 and D = 1.871

Frequently asked questions

Sources & references

The formulas on this page were last cross-checked against these sources on 2026-07-16. Mahalanobis distance is a stable mathematical definition, so this tool needs no rate or schedule updates — only the worked examples are periodically re-reconciled against NumPy and SciPy.

Related tools

LiveAI

Top-k Accuracy Calculator

Compute classification top-1 through top-k accuracy from a pasted table of ranked predictions or raw class scores. Reports the exact hit count, a full top-1…top-k ladder with marginal gains, and the samples missed at k. Mirrors scikit-learn's top_k_accuracy_score and runs entirely in your browser.

Open tool

LiveAI

Balanced Accuracy Calc

Compute balanced accuracy (the macro-average of per-class recall) from a binary confusion matrix or per-class results, contrasted against plain accuracy, with sensitivity, specificity and the adjusted chance-corrected score. Matches scikit-learn, runs entirely in your browser.

Open tool

LiveAI

Mutual Information Calc

Compute the mutual information I(X;Y) between two discrete variables from a count or joint-probability table, in bits or nats, with the full information-theory breakdown — marginal entropies H(X), H(Y), joint entropy H(X,Y), conditional entropies, normalized MI, and every per-cell term. Runs in your browser, matched to scikit-learn.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want to suggest an improvement?

Email me at [email protected] — most fixes ship within 24 hours.