gkwdist implements the Generalized Kumaraswamy (GKw) distribution family and its seven nested sub-models for bounded continuous data on $(0,1)$. All functions are implemented in C++ via RcppArmadillo for maximum computational efficiency.

Key Features: - Seven flexible distributions for proportions, rates, and bounded data - Standard R distribution API: d*, p*, q*, r* - Analytical log-likelihood, gradient, and Hessian functions - All functions implemented in C++ for optimal performance - 10-50× faster than equivalent R implementations - Numerically stable for extreme parameter values

# Install from GitHub
# install.packages("devtools")
devtools::install_github("evandeilton/gkwdist")

Distribution Functions (C++ implementation with R interface):

• d*() — Probability density function (PDF)
• p*() — Cumulative distribution function (CDF)
• q*() — Quantile function (inverse CDF)
• r*() — Random number generation

Analytical Functions for Maximum Likelihood (C++ implementation):

All analytical functions use the signature: function(par, data) where par is a numeric vector of parameters.

• ll*(par, data) — Negative log-likelihood:

$$-\ell(\boldsymbol{\theta}; \mathbf{x}) = -\sum_{i=1}^n \log f(x_i; \boldsymbol{\theta})$$

• gr*(par, data) — Negative gradient (negative score vector):

$$-\nabla_{\boldsymbol{\theta}} \ell(\boldsymbol{\theta}; \mathbf{x})$$

• hs*(par, data) — Negative Hessian matrix:

$$-\nabla^2_{\boldsymbol{\theta}} \ell(\boldsymbol{\theta}; \mathbf{x})$$

Note: These functions return negative values to facilitate direct use with optimization routines like optim(), which perform minimization by default.

Notation: All parameters are strictly positive ($\alpha, \beta, \gamma, \delta, \lambda > 0$); support $x \in (0, 1)$. The beta function is $B(a,b) = \Gamma(a)\Gamma(b)/\Gamma(a+b)$, and the regularized incomplete beta function is $I_z(a,b) = B_z(a,b)/B(a,b)$ where $B_z(a,b) = \int_0^z t^{a-1}(1-t)^{b-1}dt$.

Parameters: $\alpha, \beta, \gamma, \delta, \lambda > 0$

PDF:

$$f_{\text{GKw}}(x; \alpha, \beta, \gamma, \delta, \lambda) = \frac{\lambda \alpha \beta}{B(\gamma, \delta)} x^{\alpha-1} (1-x^\alpha)^{\beta-1} [1-(1-x^\alpha)^\beta]^{\gamma\lambda-1} \left{1-[1-(1-x^\alpha)^\beta]^\lambda\right}^{\delta-1}$$

CDF:

$$F_{\text{GKw}}(x; \alpha, \beta, \gamma, \delta, \lambda) = I_{[1-(1-x^\alpha)^\beta]^\lambda}(\gamma, \delta)$$

Quantile: Numerical inversion of the CDF via root-finding algorithms.

Moments: Analytical expressions not available in closed form. Numerical integration or simulation methods required.

Relationship: Special case of GKw with $\lambda = 1$

PDF:

$$f_{\text{BKw}}(x; \alpha, \beta, \gamma, \delta) = \frac{\alpha \beta}{B(\gamma, \delta)} x^{\alpha-1} (1-x^\alpha)^{\beta-1} [1-(1-x^\alpha)^\beta]^{\gamma-1} \left{1-[1-(1-x^\alpha)^\beta]\right}^{\delta-1}$$

Simplifying:

$$f_{\text{BKw}}(x; \alpha, \beta, \gamma, \delta) = \frac{\alpha \beta}{B(\gamma, \delta)} x^{\alpha-1} (1-x^\alpha)^{\beta-1} [1-(1-x^\alpha)^\beta]^{\gamma-1} (1-x^\alpha)^{\beta(\delta-1)}$$

$$= \frac{\alpha \beta}{B(\gamma, \delta)} x^{\alpha-1} (1-x^\alpha)^{\beta\delta-1} [1-(1-x^\alpha)^\beta]^{\gamma-1}$$

CDF:

$$F_{\text{BKw}}(x; \alpha, \beta, \gamma, \delta) = I_{1-(1-x^\alpha)^\beta}(\gamma, \delta)$$

Quantile: Numerical inversion via root-finding. For the inverse:

$$u = I_y(\gamma, \delta) \quad \text{where} \quad y = 1-(1-x^\alpha)^\beta$$

Solving for $x$:

$$x = \left[1-\left(1-I_u^{-1}(\gamma, \delta)\right)^{1/\beta}\right]^{1/\alpha}$$

Relationship: Special case of GKw with $\gamma = 1$

PDF:

$$f_{\text{KKw}}(x; \alpha, \beta, \delta, \lambda) = \alpha \beta \delta \lambda , x^{\alpha-1} (1-x^\alpha)^{\beta-1} [1-(1-x^\alpha)^\beta]^{\lambda-1} \left{1-[1-(1-x^\alpha)^\beta]^\lambda\right}^{\delta-1}$$

CDF:

$$F_{\text{KKw}}(x; \alpha, \beta, \delta, \lambda) = 1 - \left{1-[1-(1-x^\alpha)^\beta]^\lambda\right}^\delta$$

Quantile (closed-form):

$$Q_{\text{KKw}}(p; \alpha, \beta, \delta, \lambda) = \left[1 - \left(1 - \left[1-(1-p)^{1/\delta}\right]^{1/\lambda}\right)^{1/\beta}\right]^{1/\alpha}$$

Moments: Analytical expressions not available in closed form.

Relationship: Special case of GKw with $\gamma = \delta = 1$

PDF:

$$f_{\text{EKw}}(x; \alpha, \beta, \lambda) = \lambda \alpha \beta , x^{\alpha-1} (1-x^\alpha)^{\beta-1} [1-(1-x^\alpha)^\beta]^{\lambda-1}$$

CDF:

$$F_{\text{EKw}}(x; \alpha, \beta, \lambda) = [1-(1-x^\alpha)^\beta]^\lambda$$

Quantile (closed-form):

$$Q_{\text{EKw}}(p; \alpha, \beta, \lambda) = \left[1-\left(1-p^{1/\lambda}\right)^{1/\beta}\right]^{1/\alpha}$$

Moments:

$$\mathbb{E}(X^r) = \lambda \sum_{k=0}^{\infty} \frac{(-1)^k \binom{\lambda}{k+1}}{k+1} \cdot \beta B\left(1 + \frac{r}{\alpha}, (k+1)\beta\right)$$

where the binomial coefficient is generalized: $\binom{\lambda}{k+1} = \frac{\lambda(\lambda-1)\cdots(\lambda-k)}{(k+1)!}$.

Relationship: Special case of GKw with $\alpha = \beta = 1$

PDF:

$$f_{\text{MC}}(x; \gamma, \delta, \lambda) = \frac{\lambda}{B(\gamma, \delta)} x^{\gamma\lambda-1} (1-x^\lambda)^{\delta-1}$$

CDF:

$$F_{\text{MC}}(x; \gamma, \delta, \lambda) = I_{x^\lambda}(\gamma, \delta)$$

Quantile:

$$Q_{\text{MC}}(p; \gamma, \delta, \lambda) = [I_p^{-1}(\gamma, \delta)]^{1/\lambda}$$

where $I_p^{-1}(\gamma, \delta)$ is the inverse regularized incomplete beta function (quantile function of the Beta distribution).

Moments:

$$\mathbb{E}(X^r) = \frac{B(\gamma + r/\lambda, \delta)}{B(\gamma, \delta)}$$

which is valid for $r/\lambda > -\gamma$.

Relationship: Special case of GKw with $\gamma = \delta = \lambda = 1$

PDF:

$$f_{\text{Kw}}(x; \alpha, \beta) = \alpha \beta , x^{\alpha-1} (1-x^\alpha)^{\beta-1}$$

CDF:

$$F_{\text{Kw}}(x; \alpha, \beta) = 1 - (1-x^\alpha)^\beta$$

Quantile (closed-form):

$$Q_{\text{Kw}}(p; \alpha, \beta) = [1-(1-p)^{1/\beta}]^{1/\alpha}$$

Moments:

$$\mathbb{E}(X^r) = \beta B\left(1 + \frac{r}{\alpha}, \beta\right) = \frac{\beta , \Gamma(1+r/\alpha) , \Gamma(\beta)}{\Gamma(1+r/\alpha+\beta)}$$

which is valid for $r/\alpha > -1$.

Special Cases:

$$\mathbb{E}(X) = \frac{\beta , \Gamma(1+1/\alpha) , \Gamma(\beta)}{\Gamma(1+1/\alpha+\beta)}$$

$$\mathbb{E}(X^2) = \frac{\beta , \Gamma(1+2/\alpha) , \Gamma(\beta)}{\Gamma(1+2/\alpha+\beta)}$$

$$\text{Var}(X) = \mathbb{E}(X^2) - [\mathbb{E}(X)]^2$$

Relationship: Special case of GKw with $\alpha = \beta = \lambda = 1$

PDF:

$$f_{\text{Beta}}(x; \gamma, \delta) = \frac{1}{B(\gamma, \delta)} x^{\gamma-1} (1-x)^{\delta-1}$$

CDF:

$$F_{\text{Beta}}(x; \gamma, \delta) = I_x(\gamma, \delta)$$

Quantile:

$$Q_{\text{Beta}}(p; \gamma, \delta) = I_p^{-1}(\gamma, \delta)$$

Moments:

$$\mathbb{E}(X^r) = \frac{B(\gamma+r, \delta)}{B(\gamma, \delta)} = \frac{\Gamma(\gamma+r) , \Gamma(\delta) , \Gamma(\gamma+\delta)}{\Gamma(\gamma) , \Gamma(\gamma+\delta+r) , \Gamma(\delta)}$$

$$\mathbb{E}(X) = \frac{\gamma}{\gamma+\delta}$$

$$\text{Var}(X) = \frac{\gamma\delta}{(\gamma+\delta)^2(\gamma+\delta+1)}$$

                              GKw(α, β, γ, δ, λ)
                              /               \
                           λ = 1             γ = 1
                            /                    \
                   BKw(α, β, γ, δ)         KKw(α, β, δ, λ)
                         |                          |
                     α = β = 1                    δ = 1
                         |                          |
                    MC(γ, δ, λ)              EKw(α, β, λ)
                         |                          |
                      λ = 1                    γ = δ = 1
                         |                          |
                    Beta(γ, δ)                   Kw(α, β)

Note: The Beta distribution is obtained from MC by setting $\lambda = 1$, or from GKw by setting $\alpha = \beta = \lambda = 1$. The Kumaraswamy distribution is obtained from EKw by setting $\lambda = 1$, or from GKw by setting $\gamma = \delta = \lambda = 1$.

library(gkwdist)

# Set parameters
alpha <- 2; beta <- 3; gamma <- 1.5; delta <- 2; lambda <- 1.2
x <- seq(0.01, 0.99, length.out = 100)

# Density
dens <- dgkw(x, alpha, beta, gamma, delta, lambda)

# CDF
cdf <- pgkw(x, alpha, beta, gamma, delta, lambda)

# Quantiles
q <- qgkw(c(0.25, 0.5, 0.75), alpha, beta, gamma, delta, lambda)
print(q)

# Random generation
set.seed(123)
random_sample <- rgkw(1000, alpha, beta, gamma, delta, lambda)

# Visualization
par(mfrow = c(1, 2))
plot(x, dens, type = "l", lwd = 2, col = "blue",
     main = "GKw PDF", xlab = "x", ylab = "Density")
grid()

plot(x, cdf, type = "l", lwd = 2, col = "red",
     main = "GKw CDF", xlab = "x", ylab = "F(x)")
grid()

library(gkwdist)

par(mfrow = c(1,1), mar = c(3,3,2,2))

x <- seq(0.001, 0.999, length.out = 500)

# Compute densities for all families
d_gkw  <- dgkw(x, 2, 3, 1.5, 2, 1.2)
d_bkw  <- dbkw(x, 2, 3, 1.5, 2)
d_kkw  <- dkkw(x, 2, 3, 2, 1.2)
d_ekw  <- dekw(x, 2, 3, 1.5)
d_mc   <- dmc(x, 2, 3, 1.2)
d_kw   <- dkw(x, 2, 5)
d_beta <- dbeta_(x, 2, 3)

# Plot comparison
plot(x, d_gkw, type = "l", lwd = 2, col = "black",
     ylim = c(0, max(d_gkw, d_bkw, d_kkw, d_ekw, d_mc, d_kw, d_beta)),
     main = "Distribution Family Comparison",
     xlab = "x", ylab = "Density")
lines(x, d_bkw, lwd = 2, col = "red")
lines(x, d_kkw, lwd = 2, col = "blue")
lines(x, d_ekw, lwd = 2, col = "green")
lines(x, d_mc,  lwd = 2, col = "purple")
lines(x, d_kw,  lwd = 2, col = "orange")
lines(x, d_beta, lwd = 2, col = "brown")
legend("topright",
       legend = c("GKw", "BKw", "KKw", "EKw", "MC", "Kw", "Beta"),
       col = c("black", "red", "blue", "green", "purple", "orange", "brown"),
       lwd = 2, cex = 0.85)
grid()

library(gkwdist)

# Generate synthetic data from Kumaraswamy distribution
set.seed(2024)

n <- 2000
true_alpha <- 2.5
true_beta  <- 3.5
data <- rkw(n, true_alpha, true_beta)

# Get starting values
par_ini <- gkwgetstartvalues(data, family = "kw", n_starts = 2)

# Optimization using BFGS with analytical gradient
fit <- optim(
  par = par_ini,
  fn = llkw,  # Negative log-likelihood function
  gr = grkw,  # Negative gradient (negative score function)
  data = data,
  method = "BFGS",
  hessian = TRUE
)

# Standard errors from observed information matrix
se <- sqrt(diag(solve(fit$hessian)))

# 95% Confidence intervals
ci <- cbind(
  Lower    = fit$par - 1.96 * se,
  Estimate = fit$par,
  Upper    = fit$par + 1.96 * se
)
rownames(ci) <- c("alpha", "beta")

cat("True parameters:    ", true_alpha, true_beta, "\n")
cat("Estimated parameters:", fit$par, "\n\n")
print(ci)

library(gkwdist)

# Using fitted model from Example 3
x_grid <- seq(0.001, 0.999, length.out = 200)

# Fitted density
fitted_dens <- dkw(x_grid, fit$par[1], fit$par[2])

# True density (for comparison)
true_dens <- dkw(x_grid, true_alpha, true_beta)

# Diagnostic plot
hist(data, breaks = 30, probability = TRUE,
     col = "lightgray", border = "white",
     main = "Kumaraswamy Distribution Fit",
     xlab = "Data", ylab = "Density")
lines(x_grid, fitted_dens, col = "red", lwd = 2, lty = 1)
lines(x_grid, true_dens, col = "blue", lwd = 2, lty = 2)
legend("topright",
       legend = c("Observed Data", "Fitted Model", "True Model"),
       col = c("gray", "red", "blue"),
       lwd = c(10, 2, 2), lty = c(1, 1, 2))
grid()

library(gkwdist)

# Generate data from Exponentiated Kumaraswamy
set.seed(456)
n <- 1500
data <- rekw(n, alpha = 2, beta = 3, lambda = 1.5)

# Define competing models
models <- list(
  Beta = list(
    nll = function(par) llbeta(par, data),
    gr = function(par) grbeta(par, data),
    start = gkwgetstartvalues(data, family = "beta"),
    npar = 2
  ),
  Kw = list(
    nll = function(par) llkw(par, data),
    gr = function(par) grkw(par, data),
    start = gkwgetstartvalues(data, family = "kw"),
    npar = 2
  ),
  EKw = list(
    nll = function(par) llekw(par, data),
    gr = function(par) grekw(par, data),
    start = gkwgetstartvalues(data, family = "ekw"),
    npar = 3
  ),
  MC = list(
    nll = function(par) llmc(par, data),
    gr = function(par) grmc(par, data),
    start = gkwgetstartvalues(data, family = "mc"),
    npar = 3
  )
)

# Fit all models using optim with analytical gradients
fits <- lapply(models, function(m) {
  optim(par = m$start, fn = m$nll, gr = m$gr, method = "BFGS")
})

# Extract log-likelihoods and compute information criteria
loglik <- sapply(fits, function(f) -f$value)
k <- sapply(models, `[[`, "npar")

results <- data.frame(
  Model  = names(models),
  Coefs  = sapply(fits, function(f) paste0(round(f$par, 3), collapse = "|")),
  LogLik = round(loglik, 2),
  nPar   = k,
  AIC    = round(-2 * loglik + 2 * k, 2),
  BIC    = round(-2 * loglik + k * log(n), 2)
)

# Sort by AIC
results <- results[order(results$AIC), ]
print(results, row.names = FALSE)
cat("\nBest model by AIC:", results$Model[1], "\n")

library(gkwdist)

# Generate data from EKw
set.seed(789)
n <- 1000
data <- rekw(n, alpha = 2, beta = 3, lambda = 1.5)
params <- c(2, 3, 1.5)

# Compute analytical functions at true parameters
nll <- llekw(params, data)
neg_score <- grekw(params, data)
neg_hess <- hsekw(params, data)

# Fisher information matrix (negative of negative Hessian = Hessian)
fisher <- -neg_hess

# Asymptotic standard errors
se <- sqrt(diag(solve(fisher)))
names(se) <- c("alpha", "beta", "lambda")

cat("Negative log-likelihood:", nll, "\n")
cat("\nNegative score vector (should be close to zero at true params):\n")
print(neg_score)
cat("\nNegative Hessian matrix:\n")
print(neg_hess)
cat("\nFisher Information matrix:\n")
print(fisher)
cat("\nAsymptotic standard errors:\n")
print(se)

library(gkwdist)

# Generate data and fit model
set.seed(101)
n <- 2000
data <- rkw(n, alpha = 2, beta = 3)

# Fit using optim
fit <- optim(
  par = c(1, 1),
  fn = function(par) llkw(par, data),
  gr = function(par) grkw(par, data),
  method = "BFGS"
)

# Theoretical quantiles
p <- ppoints(n)
theoretical_q <- qkw(p, fit$par[1], fit$par[2])

# Empirical quantiles
empirical_q <- sort(data)

# Q-Q plot
plot(theoretical_q, empirical_q,
     xlab = "Theoretical Quantiles",
     ylab = "Empirical Quantiles",
     main = "Q-Q Plot: Kumaraswamy Distribution",
     pch = 19, col = rgb(0, 0, 1, 0.5))
abline(0, 1, col = "red", lwd = 2, lty = 2)
grid()

library(gkwdist)

# Generate data from GKw
set.seed(2025)
n <- 10000
true_params <- c(alpha = 2, beta = 2.5, gamma = 1.5, delta = 2, lambda = 1.3)
data <- rgkw(n, true_params[1], true_params[2], true_params[3],
             true_params[4], true_params[5])

# Get starting values
par_ini <- gkwgetstartvalues(data, family = "gkw", n_starts = 5)

# Fit using optim with analytical gradient
fit_gkw <- optim(
  par = par_ini, 
  fn = llgkw,    # Negative log-likelihood
  gr = grgkw,    # Negative gradient
  data = data,
  method = "BFGS",
  hessian = TRUE,
  control = list(maxit = 1000)
)

# Results
se <- sqrt(diag(solve(fit_gkw$hessian)))
estimates <- data.frame(
  Parameter = names(true_params),
  True = true_params,
  Estimate = fit_gkw$par,
  SE = se
)
print(estimates, digits = 4)

library(gkwdist)

# Generate data
set.seed(999)
n <- 10000
data <- rkw(n, alpha = 2, beta = 3)
par <- c(2, 3)

# Negative analytical gradient
neg_grad_analytical <- grkw(par, data)

# Numerical gradient (using finite differences)
neg_grad_numerical <- numDeriv::grad(
  func = function(p) llkw(p, data),
  x = par
)

# Compare
comparison <- data.frame(
  Parameter = c("alpha", "beta"),
  Analytical = neg_grad_analytical,
  Numerical = neg_grad_numerical,
  Difference = abs(neg_grad_analytical - neg_grad_numerical)
)
print(comparison, digits = 8)

The C++ implementation provides substantial performance gains:

library(microbenchmark)

# Generate large dataset
n <- 10000
data <- rkw(n, 2, 3)

# Compare log-likelihood computation
benchmark <- microbenchmark(
  R_sum_log_d = -sum(log(dkw(data, 2, 3))),  # Manual negative log-likelihood
  Cpp_ll = llkw(c(2, 3), data),               # C++ negative log-likelihood
  times = 100
)

print(benchmark)
plot(benchmark)
# Typical results: C++ implementation is 10-50× faster

Why C++ Implementation Matters:

• Speed: Critical for maximum likelihood estimation with large datasets
• Numerical Stability: Better handling of extreme parameter values and edge cases
• Memory Efficiency: Optimized memory allocation in tight loops
• Scalability: Linear scaling with sample size
• Precision: Analytical derivatives are exact (up to floating-point precision)

Model Selection Workflow:

1. Start with Beta and Kumaraswamy (2 parameters)
2. Check goodness-of-fit using Q-Q plots and formal tests
3. If inadequate, try EKw or MC (3 parameters)
4. For complex patterns, use BKw, KKw, or GKw (4-5 parameters)
5. Use AIC/BIC to balance fit quality and parsimony
6. Validate final model with residual diagnostics

• Cordeiro, G. M., & de Castro, M. (2011). A new family of generalized distributions. Journal of Statistical Computation and Simulation, 81(7), 883-898. doi:10.1080/00949650903530745

• Carrasco, J. M. F., Ferrari, S. L. P., & Cordeiro, G. M. (2010). A new generalized Kumaraswamy distribution. arXiv:1004.0911. arxiv.org/abs/1004.0911

• Kumaraswamy, P. (1980). A generalized probability density function for double-bounded random processes. Journal of Hydrology, 46(1-2), 79-88. doi:10.1016/0022-1694(80)90036-0

• Jones, M. C. (2009). Kumaraswamy’s distribution: A beta-type distribution with some tractability advantages. Statistical Methodology, 6(1), 70-81. doi:10.1016/j.stamet.2008.04.001

• Lemonte, A. J., & Cordeiro, G. M. (2013). An extended Lomax distribution. Statistics, 47(4), 800-816. doi:10.1080/02331888.2011.568119

• Cordeiro, G. M., & Lemonte, A. J. (2011). The β-Birnbaum–Saunders distribution: An improved distribution for fatigue life modeling. Computational Statistics & Data Analysis, 55(3), 1445-1461. doi:10.1016/j.csda.2010.10.007

• McDonald, J. B. (1984). Some generalized functions for the size distribution of income. Econometrica, 52(3), 647-663. doi:10.2307/1913469

• Cordeiro, G. M., & Brito, R. S. (2012). The beta power distribution. Brazilian Journal of Probability and Statistics, 26(1), 88-112. doi:10.1214/10-BJPS124

citation("gkwdist")

@Manual{gkwdist2025,
  title  = {gkwdist: Generalized Kumaraswamy Distribution Family},
  author = {J. E. Lopes},
  year   = {2025},
  note   = {R package},
  url    = {https://github.com/evandeilton/gkwdist}
}

J. E. Lopes
LEG - Laboratory of Statistics and Geoinformation
PPGMNE - Graduate Program in Numerical Methods in Engineering
Federal University of Paraná (UFPR), Brazil
Email: [email protected]

MIT License. See LICENSE file for details.