RKRaymond Kan

Raymond Kan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Raymond Kan has a total of 33 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Conditional and Lattice Saddlepoint Expansions of Arbitrary Order

    by Simon A. Broda, Raymond Kan

    The double saddlepoint approximation of Skovgaard (1987) to the distribution function of one component of a sample mean given the others, and its multivariate form, are the first terms of an asymptotic expansion whose second term was given by Kolassa (1996). We give every term of every order, for the conditional distribution function of any number of components given any number of others, for the conditional density, and for lattice as well as continuous variables; with no conditioning variables the lattice case is the continuitycorrected expansion of any order of the distribution function and the probability mass function of a sum of integer-valued random vectors, whose first term is the approximation of Daniels (1987). The expansion is of Hauschildt-Robinson type, it has no singular terms, its relative error after the terms of order n-p/2 is O(n-(p+1)/2) uniformly over compact sets of saddlepoints, and all of its terms are generated by the two recursions of Broda and Kan (2026): one on the Taylor coefficients of the cumulant generating function at the saddlepoint, and one for the mixed derivatives of the normal orthant probability. When a single component is conditioned upon the others, the integrals involved are one-dimensional, whatever the number of conditioning variables. A change of coordinates that makes the conditioning variables uncorrelated with the others under the tilted distribution is shown to be essential in double precision: without it the terms cancel by up to sixteen orders of magnitude. The expansions are applied to exact conditional tests in logistic regression and to the distribution of a sum of independent ratios of quadratic forms, such as the locally best invariant statistic for sphericity in heterogeneous panels of Broda (2019), which is a conditional probability in N + 1 dimensions; there, with five units of sixteen to twenty observations, Skovgaard's approximation has relative errors of 0.3 to 1.3% at the 10% to 0.1% points and the expansion of order four is within 2.4 × 10-4 , while with six to ten observations per unit the expansion does not converge and no approximation is reliable.

  • Open Access

    Uniform Saddlepoint Expansions for Multivariate tail Probabilities

    by Simon A. Broda, Raymond Kan

    We extend the first-order saddlepoint approximation of Kolassa (2003) for the orthant probability P(X ≥ x) of the mean of n independent and identically distributed random vectors to an expansion of arbitrary order, and sharpen his error bound. The expansion is of Hauschildt-Robinson type: its leading term is a multivariate normal orthant probability times an exponential factor and its correction terms are derivatives of the normal orthant probability. The relative error after the terms of order n-p/2 is O(n-(p+1)/2), uniformly over compact sets of saddlepoints with nonnegative components, including those at which a component vanishes; saddlepoints with negative components are handled by an inclusion-exclusion identity as in Kolassa (2003), for which we give the closed form and conditions under which it preserves the error bound, and which we show cannot in general be dispensed with. The main contribution that makes the approximation usable at every order is that every term of every order is generated by two short recursions, one on the Taylor coefficients of the cumulant generating function at the saddlepoint and one for the derivatives of the normal orthant probability, in place of the enumeration of tensor invariants of the usual construction. The first recursion also gives the density expansion of any order, and the univariate expansions of Daniels and of Hauschildt and Robinson and the expansions for a vector of ratios are special cases. In bivariate and trivariate examples, the first-order approximations of Wang (1990), Kolassa (2003), and Kolassa and Li (2010) have relative errors of up to 7% in two dimensions and 27% in three; the terms of orders two to five reduce the error, at order five, to between 10-12 and 2 × 10-4 , except when n = 1.

  • Open Access

    Computation of the Exact Density Function of the Product of a Wishart Matrix and a Normal Vector

    by Taras Bodnar, Raymond Kan, Stepan Mazur

    This paper derives the density function of the product of a Wishart matrix and a normal vector that are independently distributed. Unlike existing results, we allow the covariance matrix of the Wishart distribution and that of the normal vector to differ, and we do not assume them to be proportional. In this general setting, the density is shown to admit a two-dimensional integral representation. We further obtain the density of linear combinations of the product. For a single linear combination, it is again expressed as a two-dimensional integral whose integrand involves the modified Bessel function of the second kind, and the special case in which the covariance matrix of the normal vector is proportional to the inverse of the covariance matrix of the Wishart distribution is treated explicitly. In addition, we show that the characteristic function of any number 𝑝 of linear combinations is a one-dimensional integral, from which the density of 𝑝 < 𝑚 linear combinations is recovered as a (𝑝 + 1)-dimensional integral; together with the two-dimensional representation of the full product, this makes the density of any 𝑝 linear combinations available from an integral of dimension at most min(𝑝 + 1, 𝑚-𝑝 + 2), where 𝑚 is the dimension of the product. All resulting expressions can be evaluated numerically, and their accuracy is illustrated by a comparison with simulation; the product defines a new family of matrix-variate mean-variance mixtures of multivariate normal distributions.

  • Open Access

    Approximations Of Arbitrary Order To The Moments Of Nonlinear Functions Of Random Variables

    by Simon A. Broda, Raymond Kan

    The expectation of a smooth function of a random vector whose cumulants are all of the same large order has an asymptotic expansion in inverse powers of that order. We obtain it by regrouping the Taylor series of the function about the mean according to the number of cumulant factors in each term, give one short recursion that generates its terms in any dimension and to any order, and bound the remainder under explicit conditions. The expansion contains the saddlepoint expansion of Shi, Wang and Reid as a special case and, unlike the degree-truncated Taylor series of Hurt, adds one complete order at each step. The Laplace approximation of Lieberman for moments of ratios of quadratic forms is extended to any order and several denominators, and is shown to be the Taylor expansion of the denominators alone. Variances and mean squared errors follow by combining terms of equal order of the first two moments; this also corrects the leading term of Lieberman's variance approximation. Applications include the slope, its t-ratio and the coefficient of determination in a regression with a stochastic regressor.

  • Open Access

    Inversion Formulas and Saddlepoint Expansions for Studentized Means

    by Simon A. Broda, Raymond Kan, Yong Bao

    Let (X i , Y i), i = 1,. .. , n, be independent and identically distributed pairs with Y i > 0, and let Q = X/ √ Ȳ : Student's t, the Studentized mean, the self-normalized sum and the sample Sharpe ratio are of this form or monotone transformations of it. We start from exact inversion formulas for the density and the distribution function of Q as double integrals of the joint moment generating function of (X, Y), from a companion paper, add one for the partial expectation, and derive from them saddlepoint expansions of arbitrary order. Every term of the density expansion is generated by two recursions, and the distribution function expansion is of Hauschildt-Robinson type, without singular terms. The saddlepoint lies beyond a singularity of the integrand whenever Q is in its upper tail, which leaves the expansions unaffected but determines the form the inversion formulas must take. For linear and quadratic forms of a Gaussian vector the joint cumulant generating function is explicit and quadratic in one argument, and the inversion formulas reduce to single integrals, which give the exact distribution of the sample Sharpe ratio of serially correlated returns and of the t-statistic with a kernel variance, in finite samples and in the fixed-bandwidth limit, with the expansions within 10-7 of it. The formulas are also applied to the finite-sample distribution of the Dickey-Fuller t-ratio, which a conditioning on the residual sum of squares and on the regression coefficients of the innovations on the deterministic terms turns into a Studentized mean: for the models with and without constant and trend, with a fixed or a random start, the expansion reproduces the exact distribution function to 10-5 at the critical values for 25 observations, in seconds.

  • Open Access

    Uniform Saddlepoint Expansions for Multivariate Tail Probabilities

    by Simon Broda, Raymond Kan

    We extend the first-order saddlepoint approximation of Kolassa (2003) for the orthant probability P(X ≥ x) of the mean of n independent and identically distributed random vectors to an expansion of arbitrary order, and sharpen his error bound. The expansion is of Hauschildt-Robinson type: its leading term is a multivariate normal orthant probability times an exponential factor and its correction terms are derivatives of the normal orthant probability. The relative error after the terms of order n-p/2 is O(n-(p+1)/2), uniformly over compact sets of saddlepoints with nonnegative components, including those at which a component vanishes; saddlepoints with negative components are handled by an inclusion-exclusion identity as in Kolassa (2003), for which we give the closed form and conditions under which it preserves the error bound, and which we show cannot in general be dispensed with. The main contribution that makes the approximation usable at every order is that every term of every order is generated by two short recursions, one on the Taylor coefficients of the cumulant generating function at the saddlepoint and one for the derivatives of the normal orthant probability, in place of the enumeration of tensor invariants of the usual construction. The first recursion also gives the density expansion of any order, and the univariate expansions of Daniels and of Hauschildt and Robinson and the expansions for a vector of ratios are special cases. In bivariate and trivariate examples, the first-order approximations of Wang (1990), Kolassa (2003), and Kolassa and Li (2010) have relative errors of up to 7% in two dimensions and 27% in three; the terms of orders two to five reduce the error, at order five, to between 10-12 and 2 × 10-4 , except when n = 1.

  • Open Access

    Exact Finite-Sample Distribution and Moments of the Least Squares Estimator in First-Order Autoregressions: O(1) Algorithms

    by Raymond Kan, Yong Bao

    We derive the exact finite-sample distribution function, density and moments (when they exist) of the least squares estimator of the coefficient in a Gaussian firstorder autoregression. The intercept may be known or unknown and the initial observation fixed or stationary. With a fixed initial observation, the results cover every real coefficient, including the unit root and the explosive region. The joint characteristic function of the numerator and denominator of the estimation error follows from a bordered, rank-one-corrected tridiagonal Toeplitz determinant and a noncentral exponential factor. Five scalar primitives give this function, and all the derivatives that the inversion and moment formulas require, in closed form, at a cost independent of sample size. Chebyshev expansions resolve confluence near the unit root, accretivity settles the branch of the complex square root, and rearranged closed forms avoid cancellation in the explosive region. The same tools evaluate the local-to-unity limits of all four cases, including the Dickey-Fuller limits, and permit comparison with the closed-form explosive limit. In our benchmarks, exact p-values take milliseconds, and exact critical values and moments under a tenth of a second, at sample sizes from 100 to 1000000 .

  • Open Access

    Constructing Davis's Invariant Polynomials of Two Matrix Arguments by Wishart Averages

    by Grant Hillier, Raymond Kan, George Kirikos

    Davis's invariant polynomials of two symmetric matrix arguments extend the zonal polynomials and underlie much of noncentral matrix-variate distribution theory, but they are laborious to construct, and complete tables have appeared in print only up to total degree five. We give a construction by linear algebra alone. On the space of trace monomials of a given bidegree, three averaging operators commute: two replace one argument of the polynomial by a Wishart matrix and average, while the third transforms both arguments by the same Wishart matrix; their joint eigenspaces are exactly the families of Davis's polynomials, one-dimensional except where a Littlewood-Richardson multiplicity exceeds one. The eigenvalues are explicit integers, distinct for distinct partitions as polynomials in the degrees of freedom, so a separating choice exists at every degree. The operator matrices are built exactly, from a Wishart moment recursion of Hillier and Kan (2024), and every step is certified. Davis's published tables are confirmed entry by entry, his privately circulated degreesix tables are regenerated with four numerical errors corrected, and complete tables are produced through total degree fifteen. As an application, we obtain a finite closed form for the zonal polynomial moments of a noncentral Wishart matrix, in place of a known infinite series.

  • Open Access

    On the Joint Distribution of a Quadratic and a Linear Form in Normal Random Variables

    by Raymond Kan, Taras Bodnar, Stepan Mazur

    We derive computable representations of the joint density and the joint cumulative distribution function of a quadratic form U = x ′Ax and a linear form V = a ′x when x has a multivariate normal distribution with arbitrary mean vector and nonsingular covariance matrix. In contrast to the double Laguerre series expansion of Schöne and Schmid (2000), which requires a zero mean vector and a positive semidenite A with rank greater than two, our results allow for a nonzero mean and an indenite A, and they impose no rank condition; the singular case A ∝ aa′ is characterized and solved in closed form, and the explicit critical curve on which the joint density can be unbounded is identied. Our main representation expresses the joint distribution function as a single integral of elementary functions involving the error function with complex argument, which can be evaluated quickly and to high accuracy by numerical quadrature. Two alternative representations, one based on conditioning and one based on an innite mixture of chi-squared distributions, provide independent verication. We further obtain, under explicit nondegeneracy conditions, the joint density of two quadratic forms and a linear form and the joint density of two general second-degree polynomials in normal random variables, and simple closed-form expressions for the conditional cumulant generating function and its derivatives of all orders, which deliver an integration-free saddlepoint approximation to the joint density at any chosen order of the (asymptotic) Daniels expansion. In numerical experiments, our single-integral formula reproduces Table I of Schöne and Schmid (2000) to within the accuracy of their reported values, with the entire table computed in a fraction of a second.