SINH-Acceleration Method Overview

• SINH-acceleration is a numerical strategy that uses the hyperbolic sine function to transform integrals, enhancing convergence by reshaping analytic domains and decay properties.
• It utilizes multiple formulations—conformal maps, contour deformations, and compact sinh windows—to boost performance in quadrature, Sinc discretizations, and nonuniform sampling reconstructions.
• The technique adapts to both known and unknown singularities via optimization and Sinc-Padé approximations, achieving near double-exponential convergence in applications like option pricing.

SINH-acceleration denotes a family of numerical acceleration techniques built around the hyperbolic sine function, used to improve the convergence of quadrature, Sinc discretizations, Fourier inversion, option-pricing transforms, and nonuniform sampling reconstructions. In one formulation, the method composes a canonical outer map with a polynomial adjustment to the $\sinh$ map so that the preimages of nearby singularities are placed on the boundary of the maximal strip of analyticity, which yields asymptotically optimal double-exponential convergence for the trapezoidal rule and Sinc numerical methods. In another formulation, the integration contour is deformed by $\xi=i\omega_1+b\sinh(i\omega+y)$ into a cone where the characteristic exponent damps the integrand, so that the simplified trapezoidal rule converges exponentially with explicit error control. A third formulation uses a compactly supported sinh-type window to regularize Lagrangian nonuniform sampling series, producing sharper exponential error decay than Gaussian regularization in the setting analyzed in the literature (Slevinsky et al., 2014, Boyarchenko et al., 2018, Boyarchenko et al., 2021, Jiang et al., 25 Jan 2026).

1. Analytic structure and the role of the sinh map

The common analytic mechanism is the conversion of an oscillatory, slowly decaying, or singularity-limited problem into one whose transformed integrand is analytic in a strip and decays rapidly along the real line. In the quadrature and Sinc setting, the relevant domain is the strip

$$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$

together with the weighted Hardy space

$$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$

The convergence results apply to transformed integrands $f\circ\phi\cdot\phi'$ when the weight $\omega$ exhibits single- or double-exponential decay on the real axis (Slevinsky et al., 2014).

For canonical finite, infinite, and semi-infinite domains, the outer transformation is chosen from standard maps such as $\psi(z)=\tanh z$ on $[-1,1]$, $\psi(z)=\sinh z$ on $(-\infty,\infty)$, and $\xi=i\omega_1+b\sinh(i\omega+y)$0 or $\xi=i\omega_1+b\sinh(i\omega+y)$1 on $\xi=i\omega_1+b\sinh(i\omega+y)$2. The essential construction is then

$\xi=i\omega_1+b\sinh(i\omega+y)$3

Because $\xi=i\omega_1+b\sinh(i\omega+y)$4 grows like $\xi=i\omega_1+b\sinh(i\omega+y)$5, the composition $\xi=i\omega_1+b\sinh(i\omega+y)$6 induces double-exponential decay near endpoints or boundaries in the target domain. In the Fourier-inversion literature, the corresponding contour deformation is

$\xi=i\omega_1+b\sinh(i\omega+y)$7

with Jacobian $\xi=i\omega_1+b\sinh(i\omega+y)$8. There the deformation is selected so that the contour remains inside a union of strips and open cones where the characteristic exponent is analytic and damping is strong (Boyarchenko et al., 2018, Boyarchenko et al., 2021).

This suggests a unifying perspective: the term “SINH-acceleration” is not restricted to a single algorithm, but denotes a class of sinh-based conformal or regularizing devices whose purpose is to enlarge the effective analyticity domain seen by the numerical scheme and to convert weak decay into rapid decay.

2. Polynomial-adjusted sinh maps for trapezoidal and Sinc convergence

In the conformal-map formulation, the central problem is the treatment of finitely many singularities near the contour of integration. If the original integrand has singularities at $\xi=i\omega_1+b\sinh(i\omega+y)$9 in the $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$0-plane, these are pulled back through the outer map: $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$1 The parameters in $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$2 are then chosen so that the preimages lie on the boundary of the maximal strip $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$3: $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$4 with unknown real abscissae $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$5. This places all singularity preimages on $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$6, preserving analyticity in the interior and restoring the maximal feasible strip width $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$7 (Slevinsky et al., 2014).

The parameter selection is posed as a nonlinear program. The constraints above provide $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$8 complex conditions for the unknowns $$\mathscr{D}_d=\{z\in\mathbb{C}:|\operatorname{Im}z|<d\},$$9 and $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$0, leaving one degree of freedom. The construction maximizes $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$1, because $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$2 is directly proportional to the double-exponential decay constant $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$3; larger $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$4 yields faster endpoint decay and smaller errors. The optimization problem is

$$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$5

together with a stabilizing ad hoc constraint such as $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$6 for $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$7 or $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$8 for $$H^\infty(\mathscr{D}_d,\omega)=\left\{f:\mathscr{D}_d\to\mathbb{C}\ \text{analytic},\ \|f\|:=\sup_{z\in\mathscr{D}_d}\left|\frac{f(z)}{\omega(z)}\right|<\infty\right\}.$$9, with $f\circ\phi\cdot\phi'$0 in the reported experiments (Slevinsky et al., 2014).

The theoretical basis is the strip-width constraint

$f\circ\phi\cdot\phi'$1

identified as necessary by a nonexistence theorem. The asymptotically optimal regime is therefore

$f\circ\phi\cdot\phi'$2

For the trapezoidal rule under double-exponential decay,

$f\circ\phi\cdot\phi'$3

When $f\circ\phi\cdot\phi'$4, the asymptotic rate becomes

$f\circ\phi\cdot\phi'$5

The corresponding Sinc approximation satisfies

$f\circ\phi\cdot\phi'$6

and under $f\circ\phi\cdot\phi'$7,

$f\circ\phi\cdot\phi'$8

Within this framework, the polynomial-adjusted sinh map achieves the maximal admissible strip and improves the secondary constants via $f\circ\phi\cdot\phi'$9, which explains the observed acceleration relative to classical sinh or unadapted double-exponential maps (Slevinsky et al., 2014).

3. Adaptive singularity recovery and algorithmic realization

When singularity locations are unknown, the conformal-map approach is combined with Sinc-Padé approximation. For Sinc points $\omega$0 and samples $\omega$1, the approximant

$\omega$2

is defined by solving

$\omega$3

using central Sinc samples. The poles of $\omega$4 are then treated as approximate singularity locations and are inserted into the same nonlinear program used in the known-singularity case (Slevinsky et al., 2014).

Because Sinc sampling is double-exponentially distributed, the associated linear system becomes ill-conditioned if too many far-out samples are used. The reported practical restrictions are therefore to use samples near the center only and to take $\omega$5, with the paper giving $\omega$6 and $\omega$7 as representative choices. The adaptive procedure begins with a naive DE approximation, doubles $\omega$8 until the relative error is below $\omega$9, then iterates pole estimation, nonlinear-map optimization, and geometric refinement until the target tolerance $\psi(z)=\tanh z$0 is reached. The paper states that this adaptive method has “almost the same convergence properties” as the non-adaptive method with known singularities (Slevinsky et al., 2014).

The end-to-end implementation in the quadrature/Sinc setting proceeds as follows. One first identifies the canonical domain and selects the outer map $\psi(z)=\tanh z$1. If singularities are known, they are pulled back by $\psi(z)=\tanh z$2 and the nonlinear program for $\psi(z)=\tanh z$3 is solved; an optional homotopy from an easy initial configuration may be used to generate initial guesses. If singularities are unknown, the adaptive Sinc-Padé stage supplies approximate pole data. The transformed mesh is then set with the optimal DE step, the mapped integrand $\psi(z)=\tanh z$4 is evaluated, and $\psi(z)=\tanh z$5 is increased geometrically until the desired tolerance is achieved. Each quadrature step uses $\psi(z)=\tanh z$6 function evaluations, the nonlinear program has $\psi(z)=\tanh z$7 variables and constraints and is reported as efficiently solvable with modern solvers such as Ipopt, and $\psi(z)=\tanh z$8 for mapped Sinc can be obtained by Newton iteration (Slevinsky et al., 2014).

A recurrent comparison in this literature is with exact Schwarz–Christoffel strip maps. The polynomial-adjusted sinh map is approximate rather than exact, but it is much cheaper to evaluate because it uses elementary functions rather than strip-map integrals, while still preserving the decisive strip-width property.

4. Fourier inversion, contour deformation, and option-pricing transforms

In Fourier-based probability and pricing, SINH-acceleration is a contour-deformation technique for integrals of the form

$\psi(z)=\tanh z$9

or, more generally, for Lévy, Heston, CIR, and subordinated models with characteristic exponent $[-1,1]$0. The core map is

$[-1,1]$1

chosen so that the deformed contour remains in the analyticity domain of $[-1,1]$2 and $[-1,1]$3, while the sign of $[-1,1]$4 determines whether the contour is tilted upward or downward to turn the oscillatory factor into exponential decay. The transformed integrand is analytic in a strip $[-1,1]$5 and is evaluated by the simplified trapezoidal rule, with discretization error bounded by Stenger’s estimate

$[-1,1]$6

in the practical parameter selection scheme reported in the option-pricing literature (Boyarchenko et al., 2018, Boyarchenko et al., 2021).

A central advantage over FFT-based schemes is the decoupling of discretization, truncation, and aliasing errors. In the B-spline projection method (PROJ), the coefficients

$[-1,1]$7

are computed pointwise on the deformed contour. For linear B-splines, the dual filter

$[-1,1]$8

has pole families at

$[-1,1]$9

and the contour may be shifted deliberately across these poles when beneficial, with residues computed explicitly and reused across grid points through residue series $\psi(z)=\sinh z$0. The deformed-contour strategy removes the periodic wrap-around intrinsic to FFT truncation and thereby eliminates aliasing on narrow supports or asymmetric tails (Boyarchenko et al., 2021).

The 2018 treatment develops the same contour principle for PDFs, CDFs, European options, Heston pricing, CIR bond options, CIR-subordinated Lévy models, calibration, and quantile computation via “conformal principal components.” The reusable nodes

$\psi(z)=\sinh z$1

allow repeated evaluation of $\psi(z)=\sinh z$2, $\psi(z)=\sinh z$3, and $\psi(z)=\sinh z$4 with $\psi(z)=\sinh z$5-dependence entering only through $\psi(z)=\sinh z$6, which is then used in Newton or bisection schemes for tail quantiles (Boyarchenko et al., 2018).

The reported performance is strongly application-dependent but consistently favorable in the regimes studied. For NTS PDFs at the peak, SINH achieves $\psi(z)=\sinh z$7 with $\psi(z)=\sinh z$8–$\psi(z)=\sinh z$9 in about $(-\infty,\infty)$0 average time; for Heston puts, $(-\infty,\infty)$1–$(-\infty,\infty)$2 yields $(-\infty,\infty)$3 in roughly $(-\infty,\infty)$4–$(-\infty,\infty)$5; and for CIR bond options the node counts are much smaller than those required by fractional-parabolic deformation or flat inverse Fourier transforms. In the PROJ setting, SINH is reported as robust and accurate for European and barrier options, and it remains effective when FFT coefficients are degraded by aliasing, underspecified truncation widths, or narrow double-barrier intervals (Boyarchenko et al., 2018, Boyarchenko et al., 2021).

5. Sinh regularization in Lagrangian nonuniform sampling

A distinct, though related, use of the term appears in nonuniform sampling theory. Here the problem is reconstruction of $(-\infty,\infty)$6, where

$(-\infty,\infty)$7

from nonuniform samples associated with a sine-type function $(-\infty,\infty)$8 whose real zeros form a separated set $(-\infty,\infty)$9. The classical Lagrangian sampling expansion is

$\xi=i\omega_1+b\sinh(i\omega+y)$00

The sinh-regularized method multiplies the finite Lagrangian series by a compactly supported window

$\xi=i\omega_1+b\sinh(i\omega+y)$01

with $\xi=i\omega_1+b\sinh(i\omega+y)$02, so that distant samples are suppressed without attenuating the central contribution (Jiang et al., 25 Jan 2026).

For a finite sample set $\xi=i\omega_1+b\sinh(i\omega+y)$03, the auxiliary sine-type function

$\xi=i\omega_1+b\sinh(i\omega+y)$04

yields basis functions

$\xi=i\omega_1+b\sinh(i\omega+y)$05

and the regularized series is

$\xi=i\omega_1+b\sinh(i\omega+y)$06

In the periodic $\xi=i\omega_1+b\sinh(i\omega+y)$07-channel setting, one uses

$\xi=i\omega_1+b\sinh(i\omega+y)$08

The Fourier transform of the window has a Bessel representation, and the tail estimate

$\xi=i\omega_1+b\sinh(i\omega+y)$09

drives the main error bounds (Jiang et al., 25 Jan 2026).

The resulting non-periodic and periodic estimates are

$\xi=i\omega_1+b\sinh(i\omega+y)$10

and

$\xi=i\omega_1+b\sinh(i\omega+y)$11

The comparison paper states that Gaussian regularization yields an exponential term of order

$\xi=i\omega_1+b\sinh(i\omega+y)$12

so the sinh window nearly doubles the decay rate in the exponent at the level of the main asymptotic dependence on $\xi=i\omega_1+b\sinh(i\omega+y)$13. The analysis also relaxes the admissible perturbation condition for the finite-$\xi=i\omega_1+b\sinh(i\omega+y)$14 construction to $\xi=i\omega_1+b\sinh(i\omega+y)$15, whereas the Gaussian comparison cited in that work required $\xi=i\omega_1+b\sinh(i\omega+y)$16 (Jiang et al., 25 Jan 2026).

6. Applications, empirical behavior, and limitations

The conformal-map version of SINH-acceleration was demonstrated on endpoint-singular and complex-singular integrals, Goursat’s infinite integral, adaptive singularity recovery, nonlinear waves, and multidimensional expectations. In the interval example with logarithmic endpoint singularities at $\xi=i\omega_1+b\sinh(i\omega+y)$17 and nearby complex conjugate singularities at $\xi=i\omega_1+b\sinh(i\omega+y)$18 and $\xi=i\omega_1+b\sinh(i\omega+y)$19, the optimized map

$\xi=i\omega_1+b\sinh(i\omega+y)$20

produced markedly faster convergence than the standard DE map, with roughly $\xi=i\omega_1+b\sinh(i\omega+y)$21–$\xi=i\omega_1+b\sinh(i\omega+y)$22 more correct digits for the same $\xi=i\omega_1+b\sinh(i\omega+y)$23. For Goursat’s integral on $\xi=i\omega_1+b\sinh(i\omega+y)$24, prior DE achieved about $\xi=i\omega_1+b\sinh(i\omega+y)$25 relative error with about $\xi=i\omega_1+b\sinh(i\omega+y)$26 function evaluations, whereas the optimized DE map achieved the same error with about $\xi=i\omega_1+b\sinh(i\omega+y)$27 evaluations. In multidimensional expectations over $\xi=i\omega_1+b\sinh(i\omega+y)$28, optimized DE gave significantly lower errors than standard DE for $\xi=i\omega_1+b\sinh(i\omega+y)$29–$\xi=i\omega_1+b\sinh(i\omega+y)$30, although the total cost remained $\xi=i\omega_1+b\sinh(i\omega+y)$31 (Slevinsky et al., 2014).

In option pricing, the most emphasized practical gain is robustness under heavy tails, asymmetry, narrow supports, and path dependence. For CGMY/KoBoL densities, FFT truncation on a narrow support produces periodic wrap-around, whereas SINH computes coefficients on deformed contours without periodicity and therefore does not alias. For European options under CGMY with $\xi=i\omega_1+b\sinh(i\omega+y)$32 and $\xi=i\omega_1+b\sinh(i\omega+y)$33, the paper reports that, when the truncation width is underspecified with $\xi=i\omega_1+b\sinh(i\omega+y)$34, SINH attains beyond practical precision of approximately $\xi=i\omega_1+b\sinh(i\omega+y)$35–$\xi=i\omega_1+b\sinh(i\omega+y)$36 with moderate $\xi=i\omega_1+b\sinh(i\omega+y)$37, while FFT stagnates around $\xi=i\omega_1+b\sinh(i\omega+y)$38–$\xi=i\omega_1+b\sinh(i\omega+y)$39. For an up-and-out call with $\xi=i\omega_1+b\sinh(i\omega+y)$40, $\xi=i\omega_1+b\sinh(i\omega+y)$41, $\xi=i\omega_1+b\sinh(i\omega+y)$42, $\xi=i\omega_1+b\sinh(i\omega+y)$43, SINH reaches approximately $\xi=i\omega_1+b\sinh(i\omega+y)$44–$\xi=i\omega_1+b\sinh(i\omega+y)$45 accuracy for $\xi=i\omega_1+b\sinh(i\omega+y)$46, whereas FFT stalls near $\xi=i\omega_1+b\sinh(i\omega+y)$47. In a double barrier call with $\xi=i\omega_1+b\sinh(i\omega+y)$48, naive FFT saturates around $\xi=i\omega_1+b\sinh(i\omega+y)$49–$\xi=i\omega_1+b\sinh(i\omega+y)$50, while SINH converges with empirically fourth-order behavior, and an anti-aliasing FFT requires six times more coefficients during initialization to achieve comparable rates (Boyarchenko et al., 2021).

The nonuniform-sampling formulation shows a different empirical profile. For the test function used in the 2026 study, averaged over $\xi=i\omega_1+b\sinh(i\omega+y)$51 realizations, the non-periodic case with $\xi=i\omega_1+b\sinh(i\omega+y)$52 and $\xi=i\omega_1+b\sinh(i\omega+y)$53 yielded Gaussian error about $\xi=i\omega_1+b\sinh(i\omega+y)$54 and sinh error about $\xi=i\omega_1+b\sinh(i\omega+y)$55; for $\xi=i\omega_1+b\sinh(i\omega+y)$56 and $\xi=i\omega_1+b\sinh(i\omega+y)$57, the errors were about $\xi=i\omega_1+b\sinh(i\omega+y)$58 and $\xi=i\omega_1+b\sinh(i\omega+y)$59; and for $\xi=i\omega_1+b\sinh(i\omega+y)$60 and $\xi=i\omega_1+b\sinh(i\omega+y)$61, about $\xi=i\omega_1+b\sinh(i\omega+y)$62 and $\xi=i\omega_1+b\sinh(i\omega+y)$63. The observed curves were reported to match the predicted main exponential terms $\xi=i\omega_1+b\sinh(i\omega+y)$64 and $\xi=i\omega_1+b\sinh(i\omega+y)$65 (Jiang et al., 25 Jan 2026).

Several limitations recur across the literature. In the conformal-map quadrature setting, the map is approximate, so while it enforces $\xi=i\omega_1+b\sinh(i\omega+y)$66, the constant $\xi=i\omega_1+b\sinh(i\omega+y)$67 need not be fully optimal; however, the cited theorem shows that $\xi=i\omega_1+b\sinh(i\omega+y)$68 enters only logarithmically in the asymptotic bound. Rational Sinc-Padé approximants are less effective at locating branch points and essential singularities than simple poles, which constrains adaptivity. Infinite arrays of singularities can limit the attainable rate to roughly single-exponential under further compositions. In high dimensions, the optimized maps improve constants but do not remove the curse of dimensionality. In the Fourier-pricing formulation, narrow analyticity cones restrict the allowable tilt and reduce the strip width $\xi=i\omega_1+b\sinh(i\omega+y)$69, increasing the number of trapezoid points; branch cuts and pole geometry must still be respected carefully. In the sampling formulation, when $\xi=i\omega_1+b\sinh(i\omega+y)$70 is very close to $\xi=i\omega_1+b\sinh(i\omega+y)$71, the oversampling gap is small and larger $\xi=i\omega_1+b\sinh(i\omega+y)$72 is required because the decay exponent scales with $\xi=i\omega_1+b\sinh(i\omega+y)$73 (Slevinsky et al., 2014, Boyarchenko et al., 2018, Jiang et al., 25 Jan 2026).

A common misconception is to identify SINH-acceleration with a single fixed map such as $\xi=i\omega_1+b\sinh(i\omega+y)$74 or with tanh-sinh quadrature in the narrow sense. The cited literature instead uses the term for a broader analytic strategy: either a polynomially adjusted sinh conformal map that repositions singularity preimages, a sinh contour deformation that exploits strip-and-cone analyticity of characteristic exponents, or a compactly supported sinh window that regularizes truncation in nonuniform sampling. What these constructions share is the use of the sinh function as the mechanism by which analyticity and decay are reshaped into forms that the trapezoidal rule, Sinc approximations, or Lagrangian series can exploit at near-optimal rates (Slevinsky et al., 2014, Boyarchenko et al., 2018, Boyarchenko et al., 2021, Jiang et al., 25 Jan 2026).