Saddlepoint Approximation for Tweedie Densities

Why Approximate the Positive-Response Density?

For Tweedie models with power $1<p<2$, a response can be zero or positive. The positive part has a continuous density, but evaluating it exactly can require an infinite series. A saddlepoint approximation replaces that density with a formula built from the cumulant function and its local curvature. It is useful for calculation and for understanding how the model's mean, dispersion, and power shape the density.

This approximation concerns the density at $y>0$; it does not approximate the probability mass at zero by evaluating a continuous density there. For $1<p<2$, the model has a separate atom at zero. The distinction matters both mathematically and in likelihood calculations.

Derivation from the Cumulant Function

Write $Y$ for a Tweedie response with mean $\mu>0$, dispersion $\phi>0$, and power $1<p<2$. Its variance function is $V(\mu)=\mu^p$. In canonical form, define

\theta(\mu)=\frac{\mu^{1-p}}{1-p}, \qquad b(\theta(\mu))=\frac{\mu^{2-p}}{2-p}.

The cumulant generating function of $Y$, where it is finite, is

K(t)=\log E[e^{tY}] =\frac{b(\theta+\phi t)-b(\theta)}{\phi}.

For a target value $y>0$, the saddlepoint $t_y$ solves $K'(t_y)=y$. Since $b'(\theta(\mu))=\mu$, this means $\theta+\phi t_y=\theta(y)$, or $t_y=(\theta(y)-\theta(\mu))/\phi$. Also, $K''(t_y)=\phi b''(\theta(y))=\phi y^p$.

Definition: The Tweedie unit deviance between $y>0$ and $\mu>0$, for $1<p<2$, is $$ d(y,\mu)=2\left\{ \frac{y^{2-p}}{(1-p)(2-p)} -\frac{y\mu^{1-p}}{1-p} +\frac{\mu^{2-p}}{2-p} \right\}. $$ The positive-response saddlepoint approximation is $$ \widetilde f(y;\mu,\phi) =\frac{1}{\sqrt{2\pi\phi y^p}} \exp\left\{-\frac{d(y,\mu)}{2\phi}\right\}, \qquad y>0. $$

Here the tilde signals an approximation, not an exact density. To see why this expression arises, the saddlepoint exponent is

K(t_y)-t_y y =\frac{b(\theta(y))-b(\theta(\mu)) -y\{\theta(y)-\theta(\mu)\}}{\phi} =-\frac{d(y,\mu)}{2\phi}.

The last equality follows by substituting the definitions of $\theta$ and $b$. The local curvature term in the saddlepoint formula is $\{2\pi K''(t_y)\}^{-1/2}$, which equals $\{2\pi\phi y^p\}^{-1/2}$. Together, exponent and curvature give the displayed approximation.

What the Approximation Claims

The saddlepoint method is a local approximation to a density, obtained by evaluating an inverse transform near the point that contributes most to the integral. A precise version requires a regular setting: the cumulant generating function must be finite and differentiable near the saddlepoint, $K''(t_y)>0$, and the positive-component density must admit an inversion integral whose contribution away from the saddlepoint is negligible in the asymptotic regime considered. Under these conditions, a quadratic expansion of the exponent about the saddlepoint gives the leading term above. It does not assert that the omitted terms are zero.

Result: Under those inversion and local-asymptotic conditions, the leading saddlepoint term for the positive density at $y>0$ is $$ \frac{\exp\{K(t_y)-t_y y\}}{\sqrt{2\pi K''(t_y)}}. $$ For the Tweedie cumulant function, this is exactly $\widetilde f(y;\mu,\phi)$. In particular, the deviance supplies the exponential term, and the variance function evaluated at the observed response supplies the curvature term.

For completeness, the local calculation is as follows. In the inversion integral, tilt by $t_y$, so the exponential part at $y$ is $\exp\{K(t_y)-t_y y\}$. Write the integration variable as $t_y+iu$, where $i^2=-1$. A Taylor expansion gives

K(t_y+iu)-K(t_y)-iuy =-\frac{K''(t_y)u^2}{2}+O(|u|^3),

because $K'(t_y)=y$. The linear term cancels. In the local region that dominates the integral, the quadratic term yields the Gaussian integral $(2\pi K''(t_y))^{-1/2}$. The stated condition that the remaining part of the inversion integral is negligible is what permits this local contribution to give the leading approximation. This argument explains both the formula and why its accuracy depends on the asymptotic regime and regularity; the Taylor expansion alone is not an exact-density identity. $\square$

A separate check is that $d(\mu,\mu)=0$. More generally, $d(y,\mu)\geq 0$: its derivative with respect to $y$ is $2\{\theta(y)-\theta(\mu)\}$, which is negative for $y<\mu$ and positive for $y>\mu$, since $\theta$ increases with its argument. Thus its minimum is zero at $y=\mu$, as required for a deviance.

Worked Calculations

Worked Example: A Response Above the Mean

Take $p=1.5$, $\mu=1$, $y=2$, and $\phi=0.1$. For this power, the deviance simplifies to $d(y,1)=4(\sqrt y-1)^2$. Therefore,

d(2,1)=4(\sqrt 2-1)^2 =4(3-2\sqrt 2) \approx 0.6862915.

The curvature prefactor is

\frac{1}{\sqrt{2\pi(0.1)2^{1.5}}} =\frac{1}{\sqrt{1.777153}} \approx 0.7501.

The exponential factor is $\exp\{-0.6862915/0.2\}=\exp(-3.4314575)\approx 0.0323395$, which rounds to $0.0323$ to four decimal places. Multiplying the factors gives

\widetilde f(2;1,0.1) \approx 0.7501(0.0323395) \approx 0.02426.

This is an approximation to the positive density at $y=2$, not the probability of observing exactly 2 and not the probability of a positive response.

Worked Example: Response Size Changes the Curvature and Deviance

Keep $p=1.5$, $\mu=1$, and $\phi=0.1$, but compare $y=0.25$ and $y=4$. The deviance formula gives $d(y,1)=4(\sqrt y-1)^2$.

At $y=0.25$, $d=4(0.5-1)^2=1$ and $y^{1.5}=0.125$. Hence

\widetilde f(0.25;1,0.1) =\frac{e^{-1/0.2}}{\sqrt{2\pi(0.1)(0.125)}} =\frac{e^{-5}}{\sqrt{0.0785398}} \approx 3.5682(0.00673795) \approx 0.02404.

At $y=4$, $d=4(2-1)^2=4$ and $y^{1.5}=8$. Thus

\widetilde f(4;1,0.1) =\frac{e^{-4/0.2}}{\sqrt{2\pi(0.1)(8)}} =\frac{e^{-20}}{\sqrt{5.02655}} \approx 0.4459(2.06115\times10^{-9}) \approx 9.19\times10^{-10}.

The very small value at 4 reflects the large deviance penalty. These calculations show how the formula responds to distance from the mean, but they do not by themselves measure error against the exact density.

Worked Example: Changing Power and Dispersion

First set $\mu=1$, $y=2$, and $\phi=0.1$, and compare powers. At $p=1.25$, substitution in the deviance formula gives $d\approx0.72754$, while $2^p\approx2.37841$. Therefore,

\widetilde f(2;1,0.1) \approx \frac{\exp\{-0.72754/0.2\}} {\sqrt{2\pi(0.1)(2.37841)}} \approx \frac{0.0263}{1.2225} \approx 0.0215.

At $p=1.75$, the same substitution gives $d\approx0.64846$ and $2^p\approx3.36359$, so

\widetilde f(2;1,0.1) \approx \frac{\exp\{-0.64846/0.2\}} {\sqrt{2\pi(0.1)(3.36359)}} \approx \frac{0.0391}{1.4538} \approx 0.0269.

Now fix $p=1.5$, $\mu=1$, and $y=2$, and change dispersion from $0.1$ to $0.4$. The deviance remains $0.6862915$, while the prefactor becomes $1/\sqrt{2\pi(0.4)2^{1.5}}\approx0.3751$. The exponential factor is $\exp\{-0.6862915/0.8\}\approx0.4241$, giving

\widetilde f(2;1,0.4) \approx 0.3751(0.4241) \approx 0.1591.

These are saddlepoint values, not exact-density comparisons. They illustrate that changing power alters both the deviance and curvature, while increasing dispersion weakens the deviance penalty and changes the prefactor.

Assessing Accuracy and Avoiding Misinterpretation

A saddlepoint approximation is generally most credible when the local expansion is informative and the density is not dominated by boundary or highly discrete-like behavior. For Tweedie models, small dispersion often makes the deviance term sharply concentrated around the mean, supporting a local approximation in an appropriate interior regime. But this is a guide, not a universal error bound. Accuracy can vary with $y$, $\mu$, $p$, and $\phi$, and the approximation may be less reliable for responses near zero, where the positive component's boundary behavior and the atom at zero are especially important.

The power also affects local shape. For example, the standardized skewness of a Tweedie response is $p\sqrt{\phi}\,\mu^{p/2-1}$. Large skewness is a warning that a symmetric local Gaussian contribution may be a weaker description of the density's shape. It is not a direct formula for saddlepoint error. When accurate likelihood values are needed, compare the approximation with exact series evaluation over the response and parameter region of interest, rather than assuming a good fit at one response guarantees accuracy elsewhere.

Common pitfall: Treating $\widetilde f(0;\mu,\phi)$ as the probability of a zero response is incorrect. The formula was derived for $y>0$ and contains a factor $y^{-p/2}$; it is not an atom-probability formula. Use the separate Tweedie point mass at zero for $P(Y=0)$, and use the positive density or its exact series for $y>0$.

Another common mistake is to call the saddlepoint expression the exact Tweedie density simply because it has a closed form. The exact positive density and the approximation are different objects. The approximation supplies a tractable leading term; checking numerical accuracy requires comparison to the exact series or another justified benchmark.

Check Your Understanding

Use the definitions and derivation above to answer these questions.

For $p=1.5$ and $\mu=1$, simplify the unit deviance as a function of $y$.
Which part of the saddlepoint formula comes from $K''(t_y)$, and why is it evaluated at $y$ rather than $\mu$?
Why is the positive-response saddlepoint density not a formula for $P(Y=0)$?
With $p=1.5$, $\mu=1$, and $y=2$, explain qualitatively what increasing $\phi$ does to the deviance penalty and curvature prefactor.
What numerical comparison would you make before relying on the approximation across a broad range of responses?