Distributions and Statistical Inference Formula Sheet and Short Notes
One-page CSIR NET revision: distribution table, chi-square/t/F, order statistics, UMVUE and CRLB, NP and LR tests, chi-square, rank tests, Bayes updates.
- 9 min
On this page · 12 sections
Every formula, theorem and trap from Standard Discrete and Continuous Distributions, Sampling Distributions, Standard Errors and Order Statistics, Point Estimation and Confidence Intervals, Hypothesis Testing: MP, UMP and Likelihood Ratio Tests, Chi-Square Goodness of Fit and Large-Sample Tests and Nonparametric Tests, Rank Correlation and Bayesian Inference, on one page.
The paper in one table
Per the NTA information bulletin for December 2026 (Mathematical Sciences, code 704):
| Part | Given | Attempt | Correct | Wrong | Maximum |
|---|---|---|---|---|---|
| B | 40 | 25 | +3 | −0.75 | 75 |
| C | 60 | 20 | +4.75 | 0 | 95 |
Part C: one or more options may be correct; marks only if you identify all correct options and mark no wrong one. No partial credit, no negative marking. Details on the Mathematical Sciences overview.
Distributions
| Law | Mean | Variance | MGF |
|---|---|---|---|
| Binomial\( (n, p) \) | \( np \) | \( npq \) | \( (q + pe^t)^n \) |
| Poisson\( (\lambda) \) | \( \lambda \) | \( \lambda \) | \( e^{\lambda(e^t - 1)} \) |
| Geometric (failures) | \( q/p \) | \( q/p^2 \) | \( \frac{p}{1 - qe^t} \) |
| Negative binomial (failures) | \( rq/p \) | \( rq/p^2 \) | \( \left(\frac{p}{1 - qe^t}\right)^r \) |
| Hypergeometric | \( nM/N \) | \( n\frac{M}{N}(1 - \frac{M}{N})\frac{N - n}{N - 1} \) | |
| Uniform\( (a, b) \) | \( \frac{a+b}{2} \) | \( \frac{(b-a)^2}{12} \) | \( \frac{e^{tb} - e^{ta}}{t(b-a)} \) |
| Normal | \( \mu \) | \( \sigma^2 \) | \( e^{\mu t + \sigma^2t^2/2} \) |
| Exponential (rate \( \lambda \)) | \( 1/\lambda \) | \( 1/\lambda^2 \) | \( \frac{\lambda}{\lambda - t} \) |
| Gamma\( (\alpha, \lambda) \) | \( \alpha/\lambda \) | \( \alpha/\lambda^2 \) | \( (1 - t/\lambda)^{-\alpha} \) |
| Beta\( (a, b) \) | \( \frac{a}{a+b} \) | \( \frac{ab}{(a+b)^2(a+b+1)} \) | |
| Cauchy | none | none | none; \( \varphi = e^{-\lvert t \rvert} \) |
| Lognormal | \( e^{\mu + \sigma^2/2} \) | \( (e^{\sigma^2} - 1)e^{2\mu + \sigma^2} \) | none |
- Binomial \( \to \) Poisson (\( np = \lambda \)); hypergeometric \( \to \) binomial; sums: binomial (same \( p \)), Poisson, gamma (same rate), \( \chi^2 \), normal.
- Memoryless: exponential and geometric only. Min of exponentials: Exp\( (\sum \lambda_i) \); max is not exponential.
- \( F(X) \sim U(0, 1) \); \( -\log U \sim \) Exp(1); \( -2\log U \sim \chi^2_2 \).
- Poisson given total: binomial; thinning: Poisson\( (\lambda p) \). Gamma ratio \( \frac{X}{X+Y} \sim \) Beta.
- Normal: \( \mu_4 = 3\sigma^4 \), MAD \( \sigma\sqrt{2/\pi} \); \( X + Y \perp X - Y \) for i.i.d. normals.
- Cauchy: ratio of independent \( N(0,1) \); \( \bar{X} \) is again Cauchy.
Sampling distributions
| Law | Built from | Mean | Variance |
|---|---|---|---|
| \( \chi^2_n \) | \( \sum_{i=1}^n Z_i^2 \) = Gamma\( (\frac{n}{2}, \frac{1}{2}) \) | \( n \) | \( 2n \) |
| \( t_n \) | \( Z/\sqrt{V/n} \) | 0 (\( n > 1 \)) | \( \frac{n}{n-2} \) (\( n > 2 \)) |
| \( F_{m,n} \) | \( \frac{U/m}{V/n} \) | \( \frac{n}{n-2} \) (\( n > 2 \)) |
\( t_1 = \) Cauchy; \( t_n^2 = F_{1,n} \); \( 1/F_{m,n} = F_{n,m} \); \( F_{m,n;1-\alpha} = 1/F_{n,m;\alpha} \).
Normal sample: \( \bar{X} \sim N(\mu, \sigma^2/n) \), \( (n-1)S^2/\sigma^2 \sim \chi^2_{n-1} \), independent; \( \sqrt{n}(\bar{X} - \mu)/S \sim t_{n-1} \); \( \operatorname{Var}(S^2) = \frac{2\sigma^4}{n-1} \). Two samples: \( \frac{S_1^2/\sigma_1^2}{S_2^2/\sigma_2^2} \sim F_{n_1-1,n_2-1} \); pooled \( t_{n_1+n_2-2} \).
Standard errors: mean \( \sigma/\sqrt{n} \) (\( \times\sqrt{\frac{N-n}{N-1}} \) without replacement); proportion \( \sqrt{pq/n} \); median (normal) \( \approx 1.2533\sigma/\sqrt{n} \); s.d. \( \approx \sigma/\sqrt{2n} \); \( r \): \( \approx (1 - \rho^2)/\sqrt{n} \).
Asymptotics: CLT; Slutsky (replace \( \sigma \) by \( S \)); delta method \( \operatorname{Var} \times g'(\theta)^2 \); stabilisers \( \sqrt{\bar{X}} \) (Poisson, variance \( \frac{1}{4n} \)), \( \sin^{-1}\sqrt{\hat{p}} \) (\( \frac{1}{4n} \)), Fisher \( z \) (\( \frac{1}{n-3} \)); sample quantile variance \( \frac{p(1-p)}{nf(\xi_p)^2} \), median \( \frac{1}{4nf(m)^2} \).
Order statistics
- \( F_{(n)} = F^n \), \( F_{(1)} = 1 - (1 - F)^n \).
- \( f_{(r)}(x) = \frac{n!}{(r-1)!(n-r)!}F^{r-1}(1-F)^{n-r}f \) (“below, at, above”).
- Min and max jointly: \( n(n-1)[F(y) - F(x)]^{n-2}f(x)f(y) \), \( x < y \).
- Uniform: \( X_{(r)} \sim \) Beta\( (r, n-r+1) \), mean \( \frac{r}{n+1} \); range Beta\( (n-1, 2) \), mean \( \frac{n-1}{n+1} \).
- Exponential: \( (n-i+1)(X_{(i)} - X_{(i-1)}) \) i.i.d. Exp\( (\lambda) \); \( E\,X_{(n)} = \frac{1}{\lambda}\sum_{k=1}^n \frac{1}{k} \).
- \( P(X_{(r)} < m < X_{(s)}) = \sum_{k=r}^{s-1}\binom{n}{k}2^{-n} \); \( P(X_{(1)} < m < X_{(n)}) = 1 - 2^{1-n} \).
Estimation
- MSE \( = \) variance \( + \) bias\( ^2 \). Consistent if bias \( \to 0 \) and variance \( \to 0 \).
- Factorisation \( f = g(T; \theta)h(x) \). Exponential family: \( \sum t(X_i) \) complete sufficient.
- Rao–Blackwell: \( E(U \mid T) \). Lehmann–Scheffé: unbiased function of complete sufficient \( T \) = unique UMVUE.
- \( I(\theta) = -E[\partial_\theta^2 \log f] \); CRLB \( \frac{\tau'(\theta)^2}{nI(\theta)} \); attained only in exponential families by the natural statistic.
| Model | \( I(\theta) \) (one obs.) | MLE | UMVUE |
|---|---|---|---|
| Bernoulli \( p \) | \( \frac{1}{pq} \) | \( \bar{X} \) | \( \bar{X} \); \( pq \): \( \frac{T(n-T)}{n(n-1)} \) |
| Poisson \( \lambda \) | \( \frac{1}{\lambda} \) | \( \bar{X} \) | \( \bar{X} \); \( e^{-\lambda} \): \( (1 - \frac{1}{n})^T \) |
| Normal \( \mu \); \( \sigma^2 \) | \( \frac{1}{\sigma^2} \); \( \frac{1}{2\sigma^4} \) | \( \bar{X} \); \( \frac{1}{n}\sum(X_i - \bar{X})^2 \) | \( \bar{X} \); \( S^2 \) (variance \( \frac{2\sigma^4}{n-1} \) > CRLB) |
| Exponential rate \( \lambda \) | \( \frac{1}{\lambda^2} \) | \( 1/\bar{X} \) | \( \frac{n-1}{T} \) |
| \( U(0, \theta) \) | (irregular) | \( X_{(n)} \) | \( \frac{n+1}{n}X_{(n)} \), variance \( \frac{\theta^2}{n(n+2)} \) |
- MoM for \( U(0, \theta) \): \( 2\bar{X} \), variance \( \frac{\theta^2}{3n} \).
- MLE: invariant; asymptotically \( N(\theta, \frac{1}{nI}) \); can be biased, non-unique (\( U(\theta - \frac{1}{2}, \theta + \frac{1}{2}) \)), at a boundary.
Confidence intervals: \( \bar{X} \pm z\sigma/\sqrt{n} \); \( \bar{X} \pm t_{n-1}S/\sqrt{n} \); \( \sigma^2 \in \left[\frac{(n-1)S^2}{\chi^2_{\text{upper}}}, \frac{(n-1)S^2}{\chi^2_{\text{lower}}}\right] \); exponential mean via \( \frac{2\sum X_i}{\theta} \sim \chi^2_{2n} \); \( U(0, \theta) \): \( [X_{(n)}, X_{(n)}\alpha^{-1/n}] \). Sample size \( n \ge (z\sigma/E)^2 \).
Testing
- Size \( = \sup_{\Theta_0}P(\text{reject}) \); power under \( H_1 \); \( \beta = 1 - \) power; p-value \( \le \alpha \iff \) reject.
- NP lemma: reject when \( f_1/f_0 > k \), randomise at equality; MP tests are unbiased.
- Karlin–Rubin: MLR in \( T \) and one-sided \( H_1 \) \( \Rightarrow \) UMP test “reject for large \( T \)”.
- Two-sided normal mean: no UMP (UMPU exists). \( U(0, \theta) \) two-sided: UMP exists.
- LRT: \( \Lambda = \sup_{\Theta_0}L/\sup_\Theta L \); \( -2\log\Lambda \to \chi^2_r \), \( r \) = number of restrictions. Normal LRTs: \( t \), \( \chi^2_{n-1} \), pooled \( t \), \( F \), ANOVA.
Large-sample and chi-square tests
| Test | Statistic |
|---|---|
| Mean | \( \frac{\bar{x} - \mu_0}{s/\sqrt{n}} \) |
| Two means | \( \frac{\bar{x}_1 - \bar{x}_2}{\sqrt{s_1^2/n_1 + s_2^2/n_2}} \) |
| Proportion | \( \frac{\hat{p} - p_0}{\sqrt{p_0q_0/n}} \) |
| Two proportions | pooled \( \hat{p} = \frac{x_1 + x_2}{n_1 + n_2} \) in the SE |
| Correlation | \( (\tanh^{-1}r - \tanh^{-1}\rho_0)\sqrt{n - 3} \) |
| Goodness of fit | \( \sum \frac{(O-E)^2}{E} = \sum \frac{O^2}{E} - N \), df \( k - 1 - s \) |
| \( r \times c \) table | \( E_{ij} = \frac{R_iC_j}{N} \), df \( (r-1)(c-1) \) |
| \( 2 \times 2 \) | \( \frac{N(ad - bc)^2}{(a+b)(c+d)(a+c)(b+d)} \), df 1 \( = Z^2 \) |
Critical values: \( z \): 1.645 / 1.96 (5% one/two-tailed), 2.326 / 2.576 (1%). \( \chi^2_1 \): 3.84 (5%), 6.63 (1%). Pool classes with small expected counts before counting \( k \).
Nonparametric tests and rank correlation
| Test | Null facts | Assumes |
|---|---|---|
| Sign | \( S \sim \) Bin\( (n, \frac{1}{2}) \) | continuity |
| Wilcoxon signed-rank | \( E\,T^+ = \frac{n(n+1)}{4} \), \( \operatorname{Var} = \frac{n(n+1)(2n+1)}{24} \) | symmetry |
| Mann–Whitney | \( E\,U = \frac{n_1n_2}{2} \), \( \operatorname{Var} = \frac{n_1n_2(N+1)}{12} \); \( W = U + \frac{n_1(n_1+1)}{2} \) | independent samples |
| Runs | \( E\,R = 1 + \frac{2n_1n_2}{N} \) | |
| Kruskal–Wallis | \( H \approx \chi^2_{k-1} \) | \( k \) independent samples |
| Kolmogorov–Smirnov | \( \sup\lvert F_n - F_0 \rvert \) distribution-free | \( F_0 \) continuous, fully specified |
ARE vs \( t \) under normality: Wilcoxon / Mann–Whitney \( 3/\pi \approx 0.955 \), sign \( 2/\pi \approx 0.637 \).
Spearman \( r_s = 1 - \frac{6\sum d^2}{n(n^2-1)} \) (Pearson on ranks; null variance \( \frac{1}{n-1} \)). Kendall \( \tau = \frac{C - D}{n(n-1)/2} \). Both invariant under increasing transformations.
Bayesian inference
- Posterior \( \propto \) likelihood \( \times \) prior.
- Beta\( (a, b) \) + \( x \) successes in \( m \) \( \to \) Beta\( (a + x, b + m - x) \). Gamma\( (\alpha, \beta) \) + Poisson data \( \to \) Gamma\( (\alpha + \sum x, \beta + n) \). Gamma + exponential data \( \to \) Gamma\( (\alpha + n, \beta + \sum x) \). Normal mean: precisions add, mean is precision-weighted.
- Squared loss: posterior mean; absolute: median; 0–1: mode.
- Uniform prior on \( p \): posterior mean \( \frac{x+1}{n+2} \). Jeffreys \( \propto \sqrt{I(\theta)} \); binomial Beta\( (\frac{1}{2}, \frac{1}{2}) \).
- Credible interval: \( P(\theta \in C \mid x) = 1 - \alpha \); HPD is the shortest.
Counter-examples and traps table
| Claim | Status | Witness |
|---|---|---|
| Every distribution has a mean | false | Cauchy |
| \( \bar{X} \) is always consistent | false | Cauchy: \( \bar{X} \) is Cauchy for every \( n \) |
| The maximum of exponentials is exponential | false | \( P(\max \le x) = (1 - e^{-\lambda x})^n \) |
| \( \bar{X} \perp S^2 \) for any population | false | only for normal populations |
| An unbiased estimator’s square root is unbiased | false | \( E\,S < \sigma \) |
| The UMVUE attains the CRLB | false | \( S^2 \) for \( \sigma^2 \); \( \frac{n-1}{T} \) for an exponential rate |
| The CRLB applies to \( U(0, \theta) \) | false | UMVUE variance \( \frac{\theta^2}{n(n+2)} \) |
| The MLE is unique | false | \( U(\theta - \frac{1}{2}, \theta + \frac{1}{2}) \) |
| Minimal sufficient implies complete | false | \( U(\theta, \theta + 1) \) |
| A UMP two-sided test exists for a normal mean | false | MP tests reject in opposite tails |
| Every discrete MP test is non-randomised | false | Poisson, size 0.05 |
| \( \chi^2 \) df ignores estimated parameters | false | Poisson fit: \( k - 2 \) |
| Signed-rank test needs normality | false | needs symmetry only |
| A 95% CI contains \( \theta \) with probability 0.95 | false | coverage is a property of the procedure |
| A Bayes estimator is unbiased | false | \( \frac{n\bar{X}}{n+1} \) under a \( N(0, 1) \) prior |
Key Takeaways
What’s Next
- Back to the chapter overview or the Mathematical Sciences overview.
- The probability foundations start at Descriptive Statistics and Exploratory Data Analysis.
- Revisit Hypothesis Testing: MP, UMP and Likelihood Ratio Tests if the testing table above felt slow.