| Aspect | Discrete | Continuous |
|---|---|---|
| Defining condition | smallest m with P(X ≤ m) ≥ 0.5 and P(X ≥ m) ≥ 0.5 | unique m where F(m) = 0.5 |
| Uniqueness | may be non-unique when CDF jumps past 0.5 | unique when CDF is strictly increasing |
| CDF shape | step function with jumps at support values | smooth, continuous curve |
| Computation | summation: ∑ P(X = k) ≥ 0.5 | integration: ∫ᵐ f(x) dx = 0.5 |
| Solving method | evaluate CDF at each support value | solve F(m) = 0.5 or use m = F⁻¹(0.5) |
| Closed form availability | rare (geometric is a notable exception) | often available (normal, exponential, Weibull, ...) |
| Visual identification | step function crosses horizontal line at 0.5 | vertical line splitting density area into equal halves |
| Result is | a value or sometimes an interval | always a single point (non-pathological) |
| Property | Statement | Notes / example |
|---|---|---|
| Minimizes absolute deviation | median(X) = arg minₘ E[|X − m|] | mean minimizes squared deviation instead |
| Linear transformation | median(aX + b) = a · median(X) + b, a ≠ 0 | X has median 4 ⟹ 3X − 2 has median 10 |
| Monotonic invariance | median(g(X)) = g(median(X)) for strictly monotonic g | fails for the mean; holds for mode |
| Robustness to outliers | values beyond the 50% threshold have zero impact | ideal for contaminated data, income, housing |
| Uniqueness (continuous) | unique when CDF is strictly increasing | may be an interval if CDF has flat regions |
| Symmetric distributions | median = mean = center of symmetry | normal: all three equal μ |
| Independent of spread | two distributions can share a median yet differ in variance | N(0, 1) and N(0, 100) both have median 0 |
| No additivity | median(X + Y) ≠ median(X) + median(Y) in general | limits use in sums; mean is additive |
| Distribution | Type | Median | Notes |
|---|---|---|---|
| Discrete Uniform on {a, ..., b} | discrete | (a + b)/2 when a + b is even | two valid medians when a + b is odd |
| Binomial(n, p) | discrete | ≈ np (no closed form) | numerical CDF lookup for exact value |
| Geometric(p) | discrete | ⌈−ln(2) / ln(1 − p)⌉ | rare closed form among discrete distributions |
| Negative Binomial(r, p) | discrete | no closed form | numerical CDF required |
| Hypergeometric(N, K, n) | discrete | no closed form | numerical evaluation required |
| Poisson(λ) | discrete | ≈ λ + 1/3 − 0.02/λ (large λ) | numerical for small λ |
| Continuous Uniform on [a, b] | continuous | (a + b)/2 | midpoint of the interval; = mean |
| Normal(μ, σ) | continuous | μ | = mean = mode by symmetry |
| Exponential(λ) | continuous | ln(2)/λ ≈ 0.693/λ | strictly below the mean 1/λ |
| Beta(α, β) | continuous | 0.5 when α = β; else no closed form | numerical for general α, β |
| Gamma(k, θ) | continuous | ln(2)/θ when k = 1; else numerical | reduces to exponential at k = 1 |
| Weibull(k, λ) | continuous | λ (ln 2) | closed form for any shape k |
| Cauchy(x₀, γ) | continuous | x₀ | mean does not exist; median still well-defined |
| Lognormal(μ, σ) | continuous | e | strictly below the mean e |
| Edge case | When it occurs | Example |
|---|---|---|
| Median = mean | distribution is symmetric around a point | normal: median = mean = μ |
| Median ≪ mean | strong right skew or heavy right tail | lognormal: median e ≪ mean e |
| Non-unique discrete median | CDF lands exactly on 0.5 at a support value | discrete uniform on {1, 2, 3, 4}: medians at 2 and 3 |
| Median exists, mean doesn't | heavy-tailed distribution with undefined first moment | Cauchy: median = x₀; mean undefined |
| Median shifts under truncation | support is restricted, removing probability mass from one side | normal truncated to [0, ∞): median > μ |
| Median far from peaks | mixture of well-separated components | 0.5 · N(0, 1) + 0.5 · N(10, 1): median ≈ 5, between the two modes |
| Infinitely many medians | CDF is flat over an interval at level 0.5 | pathological constructed densities only |
| Mistake | Why it's wrong | Correct view |
|---|---|---|
| Treating median as the average | median is a percentile rank, not a weighted sum | use mean for averages, median for the 50% split |
| Assuming median = mean always | only true for symmetric distributions | check symmetry; skew separates them substantially |
| Checking only one inequality (discrete) | CDF can jump past 0.5 without landing on it | verify both P(X ≤ m) ≥ 0.5 and P(X ≥ m) ≥ 0.5 |
| Assuming uniqueness | discrete distributions can have multiple medians | inspect the CDF behavior near 0.5 |
| Reading median off the PMF peak | median depends on cumulative probability, not individual values | use the CDF; the peak is the mode |
| Expecting median to be in the dataset | median is a calculated position, not necessarily an observed value | for even n, sample median averages the two middle values |
| Confusing median with mode | median = probability split; mode = peak | exponential has mode 0 but median ln(2)/λ |
| Treating sample median as population median | small samples vary substantially | use confidence intervals; larger samples are more reliable |
| Aspect | Statement | Example / note |
|---|---|---|
| What it is | the value that splits total probability into two equal halves | the 50th percentile of the distribution |
| Discrete case | smallest m with P(X ≤ m) ≥ 0.5 and P(X ≥ m) ≥ 0.5 | may be non-unique when the CDF lands on 0.5 |
| Continuous case | solve F(m) = 0.5, or use m = F⁻¹(0.5) | exponential: m = ln(2)/λ |
| Vs other measures | a rank-based center, not a weighted balance point or peak | right-skewed: mode < median < mean |
| Symmetric vs skewed | equals mean and mode under symmetry; sits between them under skew | normal: median = μ; exponential: median < mean |
| Defining property | minimizes E[|X − m|]; invariant under monotonic transformations | median(g(X)) = g(median(X)) for strictly monotonic g |
| When it shines | skewed data, outlier resistance, ordinal data | income, housing prices, satisfaction ratings |
| When it's weak | no additivity; reveals nothing about spread; algebra is awkward | median(X + Y) ≠ median(X) + median(Y) |