In this paper
AbstractIntroductionRepresentation of the mediumStarting pointThe Yukawa dipole as a superposition of GaussiansThe medium as a Poisson processKernel and broadening as a single functionalMaster formula: spectrum differential in k and with a cutCenter-by-center transverse integralsIntegrals over and over Integral over Spectrum integrated over all kArnold’s route: the Wronskian identityGelfand–Yaglom route: as a multilinear determinantOne center: closed-form average over Monte CarloMecke’s formulaDimensionless variables and sampling in the opacity coordinateCase 1: spectrum integrated over Cases 2 and 3: with a cut and differentialLog-uniform sampling in and truncation errorControl variate, all at once, and real partPhysical variables and conventions for expanding mediaCostValidation and comparison with published resultsInternal checksFirst order in opacityComparison with the results of Andrés et al.Summary and OutlookData AvailabilityAcknowledgmentsContinuous route: Jacobi equation and Arnold’s propagatorHTL potential with a magnetic massReferencesWorksPhysics and Astronomy
A Poisson-gas approach to medium-induced gluon radiation
Abstract
The precision era of jet-quenching phenomenology requires a quantitative control of medium-induced gluon radiation. We consider the BDMPS-Z spectrum in the soft-gluon limit for a Yukawa (Gyulassy-Wang) potential. Writing the Yukawa dipole cross section as a Laplace superposition of Gaussians, we express the fully resummed spectrum as an average over Poisson-distributed configurations of Gaussian scattering centers, each of which is solved in closed form. This yields a master formula for the spectrum both differential in the transverse momentum and integrated with a cut on , valid for static and expanding media with arbitrary profiles. Its evaluation by Monte Carlo sampling of the configurations is efficient and trivially parallelizable. The method reproduces the first order in opacity and published results without kinematic cuts, while with a cut we find sizeable relative deviations from previous results. We argue that the same construction extends to the more complex building blocks of an in-medium parton cascade, such as finite energy fraction and emissions off color antennas, and provides a natural starting point for analytical approximations. The code is publicly available.
Introduction
The analysis of hard probes in high-energy heavy-ion collisions is one of the main experimental handles on the properties of the quark–gluon plasma created in those collisions. One of the central theoretical tools to describe hard probes in such an environment is the medium-induced gluon radiation spectrum. With more precise measurements and new observables becoming available, a precise evaluation of this spectrum has become essential.
In this paper we consider the BDMPS-Z formalism [1, 2, 3, 4] for gluon radiation off an eikonal parton in the soft limit. In this description the medium acts as a color background field with which the jet interacts. We neglect recoil effects and assume the background field to be classical and to obey Gaussian statistics. The medium is then characterized by the two-point function
where x and y are transverse positions, is the density of scattering centers at path length , and is the Fourier transform of the elastic scattering potential, whose screening mass may also depend on . In the soft limit the medium enters through an adjoint dipole, whose Gaussian average is with
that is, , with the Fourier transform of , is the elastic collision rate of Ref. [5].
Typically, the potential is taken to be of Yukawa type (Gyulassy–Wang model) [6]. Another option is the HTL potential [7], which shares the tail of the Yukawa potential and, when regularized with a magnetic mass, also admits the representation as a Poisson gas of Gaussian scattering centers used in this paper (Appendix B). In this paper we focus on the Yukawa potential and leave other cases for future studies. This tail is associated with rare but relevant hard kicks. It makes the dipole cross section behave as at small , and it has proven a hurdle that is difficult to overcome analytically.
Traditionally, two analytical limits have been used: GLV [8, 9, 10] and the harmonic oscillator (HO) [1, 11]. The GLV spectrum is the first order of an expansion in the opacity of the medium. The HO approximation keeps only the quadratic term of the dipole cross section, which amounts to a fixed width for every scattering, but it resums multiple scatterings. More recently, the improved opacity expansion (IOE) [12, 13, 14], which expands around the HO, and the resummed opacity expansion (ROE) [15], which extends the description to the Bethe–Heitler regime, have made significant progress.
From the numerical point of view, Andrés et al. [5] proposed to compute the spectrum by solving a PDE, following earlier work by Caron-Huot and Gale [16], and later extended this approach to expanding media with kinematic cuts [17]. Closer to the approach followed in this paper, Feal and Vázquez [18] used a Monte Carlo method. Both approaches resum multiple scatterings to all orders.
In this paper we also tackle the problem in full generality. We write the Gyulassy–Wang dipole cross section as a mixture of Gaussians, which allows us to interpret the medium as a Poisson gas of Gaussian scattering centers, each characterized by its position and its transverse-momentum width (see § 2). Analytical calculations lead us to a master formula that encompasses both the spectrum differential in and the spectrum with a cut on . This master formula is amenable to a standard Monte Carlo integration over configurations, which is efficient, highly parallelizable and able to handle arbitrary profiles of both and .
Unlike PDE-based approaches, each configuration of the medium is solved in closed form. This provides a clearer picture of the BDMPS-Z spectrum: every configuration is Gaussian and is solved like the harmonic oscillator, with a frequency given by a sum of impulses located at the scattering centers. Feal and Vázquez also use a Monte Carlo method, but they sample momentum kicks in discretized time, whereas we sample Poisson-distributed centers in continuous time. On the other hand, Feal and Vázquez include exact kinematics, large emission angles, a thermal gluon mass and the external legs. We leave the inclusion of this physics in our formalism for future work.
The method reproduces the first order in opacity and the published results without kinematic cuts. With a cut, we find sizeable relative deviations from previous results, both at small gluon energy fraction, where the cut strongly suppresses the spectrum, and at large energy fraction, where they reach 5–20 %. The code is publicly available at https://github.com/manoel-maker/poisson-gas-bdmps_z.
The paper is organized as follows. In Sec. II we set up the calculation: starting from known expressions for the BDMPS-Z spectrum [5, 10] and relying on our representation of the medium, we write the spectrum in a form from which the master formula, on which this paper is based, follows directly. In Sec. III we derive the master formula for the spectrum differential in and with a cut, for arbitrary and , by integrating the transverse positions center by center; the equivalent route through Arnold’s propagator [19] is given in Appendix A. In Sec. IV we obtain the spectrum integrated over all , again by two routes. In Sec. V we present the Monte Carlo method. Although the methodology is standard, we give a self-contained review: we present Mecke’s formula and discuss the importance sampling of the time between the emission and a marked scattering center. Section VI contains the validation and the comparison with published results, and Sec. VII our summary and outlook. Appendix B discusses the HTL potential with a magnetic mass.
Representation of the medium
Starting point
As detailed in the Introduction, we study the medium-induced gluon emission spectrum in the soft-gluon limit. Our starting point is the spectrum differential in , Eqs. (2.2)–(2.4) of Ref. [5], written in position space:
Figure 3.
with , and “(vacuum)” denotes the same expression with . In (2) the medium enters through two factors. The first one is the kernel , which describes the evolution of the dipole between and . The second factor is the broadening factor . It has the form of a kernel without kinetic term: after the kinetic phases of the amplitude and of the conjugate amplitude cancel, the dipole size is frozen at , and the factor describes the broadening of the emitted gluon. That the problem reduces to a single dipole is a consequence of the soft limit, in which the emitter is purely eikonal.
For the Yukawa potential [5], with a local screening mass [17], the dipole cross section gives
The Yukawa dipole as a superposition of Gaussians
Since is a rational function of , its inverse Laplace transform [20] is immediate,
and the integral over in becomes Gaussian. This is, in essence, Schwinger’s proper-time parametrization [21], commonly used in quantum field theory. Performing the Gaussian integral and changing variables to we obtain, for each ,
with
Each term is the dipole factor of a Gaussian scattering center, which transfers a transverse momentum distributed as , with . We refer to as the transverse-momentum width of the center. The screening mass only sets the scale of the widths: with , .
Let us analyze the result. is a probability density, , but it has a heavy tail, for , so that the moments diverge for . This tail is the Coulomb tail of , that is, the hard kicks. It will be relevant in the Monte Carlo of § 5: it is the reason why we integrate exactly over the width of a center whenever possible (§ 5 C), and why we introduce importance sampling at short times (§ 5 E).
The medium as a Poisson process
Beyond the apparent simplification of writing the potential as an average of Gaussians, the representation (6) allows us to interpret the interaction with the medium as a Poisson process. We first discuss an identity valid for Poisson processes in general, and then apply it to our problem.
Consider points in the plane, generated as follows: we divide the plane into cells and, independently for each cell, place a point in it with probability . This is a Poisson process in the plane with intensity . Equivalently, the positions are distributed along with density , and each center draws its width from , independently of everything else (marking theorem [22 Theorem 5.6]). The number of centers is Poisson distributed with mean , the opacity. We denote by the expectation over this process.
Generating functional. For a bounded function ,
Proof. The cells are independent, so the expectation of the product is the product of the expectations. In a cell the factor equals with probability and 1 otherwise; its expectation is . The product over cells tends to the exponential in Eq. (8). □ This is the probability generating functional of a Poisson process [22 Eq. (3.5) and Theorem 3.9]; the proof above is complete and does not require homogeneity.
In our case, in each slice there is a scattering center with probability , and each center carries a Gaussian potential whose width is drawn from ; the Yukawa potential emerges from the average over the widths. Indeed, with , Eqs. (8) and (6) give, for any path and any time interval ,
with . For a fixed configuration the exponent is quadratic in : a harmonic oscillator whose frequency acts only at the instants .
Note that the number of centers in a configuration is not the order in the opacity expansion. Each power of in that expansion mixes a real and a virtual term, whereas in Eq. (9) the virtual terms are resummed in the Poisson weight of each configuration. This weight is the elastic Sudakov factor that is resummed in the resummed opacity expansion [15], which then expands in the number of real scatterings. Here, instead, each configuration, with any number of centers, is solved in closed form, and the average over configurations sums all of them.
Kernel and broadening as a single functional
Consider the complete path on , equal to between and and equal to (frozen) after . The product of kernel and broadening factor in Eq. (2) then contains
that is, the functional (9) evaluated on a single path. For a single configuration: the centers with enter the kernel, with the varying path; the centers with enter the broadening factor with fixed, , with ; and the centers with do not contribute. Exchanging with the path integral,
Configuration by configuration, everything is Gaussian: an oscillator with an impulsive frequency.
Since centers in disjoint intervals are independent, for fixed the expectation would also factorize, . The advantage of using a single configuration is a different one: configuration by configuration everything is Gaussian in , and when moves a center passes continuously from one factor to the other (Sec. III C), which allows the integral over to be done in closed form.
Master formula: spectrum differential in k and with a cut
By Eqs. (2) and (11), for a fixed configuration we need
Everything in this section refers to a fixed configuration, with arbitrary and , which enter only through the and the ; the expectation is taken at the end.
Center-by-center transverse integrals
For a fixed configuration, the integrand of Eq. (12) is a product of three types of factors: free Fresnel kernels between consecutive instants, ; at each center , the factor , with the transverse position at ; and, at , the broadening factor together with the phase . We integrate over the positions one at a time, in time order.
Accumulated kernel. After integrating over all positions before an instant , the result as a function of the position at is
where the term in is not needed, since only at enters. The family (13) is closed under the two operations that appear, with the rules of Table 1. A free segment is the two-dimensional Gaussian integral
| step | |||
| first segment, length from | |||
| free segment, length | |||
| center |
Table 1. TABLE I. Rules for the accumulated kernel (13).
with , its rule for is simply . A center multiplies by , with
The Fresnel integrals with complex parameters have a precise distributional meaning. Here, moreover, they are absolutely convergent in almost all cases: each intermediate position sits at a center and carries the real damping , and the final position carries with as soon as there is a center after . Only for is the integral over a pure Fresnel integral, to be understood in the distributional sense.
The function . We define
Between centers grows as , so that is linear in with constant slope . At a center, with fixed means
with continuous. Without medium, ; centers with do not enter. Expanding the recursion, is a multilinear polynomial in the ,
where the sum runs over the subsets of the centers in , the term being . This is the Gelfand–Yaglom structure [23] that reappears in Sec. IV B.
Integrals over and over
Broadening as one more kick. Multiplying the kernel at by amounts, in Table 1, to : the accumulated effect of all centers after , applied at once, with no free segments between them. The total coefficient of in the exponent is then
When crosses a center, its kick passes from “after ” to “before ” without changing the sum of kicks applied at that instant: is continuous by construction. This is the transparent form of the statement that kernel and broadening are a single functional (see Sec. III D).
Positivity of . From Eq. (18), . On the first segment is real. At each center increases by . On a free segment with real preserves the sign of , and , so is preserved. Hence , with strict inequality as soon as there is any center after .
Derivative at . From (13) and (15),
Applying , with in two dimensions,
By Eq. (19), the integrand of Eq. (12) is . Since , these are two Gaussian integrals in two dimensions,
Thus, per configuration and for fixed ,
For the integral over we use and . The terms in without cancel, and
With the prefactor of (2) and ,
Integral over
For fixed and configuration, , , , , and are functions of .
(i) Primitive between centers. Between two centers and are constant. Let
Then ; , hence ; and . Adding up,
(ii) Continuity at the centers. and are continuous (Sec. III A), hence so is ; only jumps. At ,
with and : the jump of exactly compensates that of . The value of at itself is ambiguous, but it is a set of measure zero, the piecewise sum only uses the one-sided limits, and the combination that enters is continuous.
(iii) Piecewise sum. , with three types of contribution. (a) Endpoint . , , , , and : a purely imaginary constant, independent of the configuration. It is exactly the vacuum contribution, which cancels in the subtraction. (b) Endpoint . grows linearly and , so . (c) Centers . With and continuous and the jump (16),
with . For there are no later centers and the contribution vanishes. We arrive at the master formula, valid for arbitrary and :
with given by (25), , the slopes just before and after center , , and the expectation over the Poisson process of § 2 C. Differentiating with respect to ,
a mixture of complex Gaussians in .
Reading. The gluon is emitted at ; the centers between and build through the jumps (16); then comes a “special” center ; the centers after only count through inside , that is, through the broadening of the already emitted gluon. The profile of the medium only enters through where the centers fall () and through their widths ().
Spectrum integrated over all k
Integrating (2) over all gives , and since the broadening factor disappears. With and (19) at ,
where the vacuum has . This is the structure of Eqs. (3.7)–(3.8) of Arnold [19]. It is also the limit of the master formula: with , Eq. (26) gives
The next two subsections bring this spectrum to the form .
Arnold’s route: the Wronskian identity
Integral over [19 Eqs. (3.9)–(3.15)]. and solve the same equation in , with constant Wronskian , evaluated at , where and . Dividing by , , and for the vacuum . At , before the first center, and exactly, so the lower limit does not contribute. For , is a straight line, with , and . Hence
Integral over [19 Eqs. (3.19)–(3.20)]. For there are no centers after and . Then , and with , ,
The real part removes the branch ambiguity of the logarithm.
Link with the master formula. Doing the same integral with the primitive of Sec. III C at , where , the left-hand side of Eq. (30) equals . Therefore
and Eqs. (29) and (31) coincide. Arnold’s paper [19] treats precisely an arbitrary , i.e., expanding media, in the HO approximation; here the same identity is applied configuration by configuration.
Gelfand–Yaglom route: as a multilinear determinant
Lemma. , with and .
Proof. Multilinearity. By the rules of Table 1, or with the transfer matrices (free segment ; center , ), each center acts linearly in its , so that is multilinear, , with the sum over subsets of centers. Coefficients. Let . The coefficient can be computed with only the centers in . After the last one, , and . Since does not depend on , the coefficient of in is times that of , which is the problem with centers. With , . Determinant. Since is diagonal, , with the principal minors. Subtracting from each row of the previous one gives a triangular matrix with diagonal , so . □
This is the analogue, for the final slope, of the multilinear polynomial (17). In practice is computed with the Möbius recursion: with (the of Sec. III A with ), between centers and at each center, so that
where is the value just before center .
One center: closed-form average over
Each factor of Eq. (33) only depends on the width of its own center and on the earlier centers. For the center at , , with
and (Sec. II B). The average over the width is then elementary,
since and . This is the piece used by the Monte Carlo of the integrated spectrum (Sec. V C). With a single center it gives the first order in opacity, , which for a static medium is the GLV result, Eq. (B.4) of Ref. [5] in the limit (no kinematic cut; see Sec. V B).
Monte Carlo
Equations (26), (27) and (31) are averages over configurations of the Poisson process. The Monte Carlo method is the standard one [24]: independent samples of an estimator whose expectation is the quantity sought, and the result . What is not trivial is the choice of . We use four variance-reduction techniques, all of general use: (i) a marked center (Mecke’s formula), which replaces a sum over centers by an integral over the position of one “special” center; (ii) exact integration of one variable, the width of the marked center, whenever possible; (iii) importance sampling of the time between the emission and the marked center; and (iv) a control variate, the estimator without cut, whose mean is the integrated spectrum. Everything is written for arbitrary and ; the static medium is a particular case.
Mecke’s formula
For a function of a point and of the configuration of the Poisson process of Sec. II C,
This is Mecke’s formula [25], [22 Theorem 4.1]. With the cells of Sec. II C it follows in a few lines: the cell at contributes to the sum only if it contains a center, which happens with probability ; by the independence between cells, the rest of the configuration has the same law as if that cell did not exist, and an infinitesimal cell does not change that law. □
In practice one samples the position of the marked center with density , and only those centers that influence its term. Each sample contributes a single term instead of a sum, and centers that do not influence it need not be sampled.
Dimensionless variables and sampling in the opacity coordinate
We measure times in units of and transverse momenta in units of a reference screening mass :
with (Sec. II B). Then with
For a static medium is the opacity, and with , where, as in Ref. [5], and . The kinematic cut then corresponds to , with ; means no cut.
Sampling. With the opacity coordinate , the probability of a center in a slice is : in the variable the process is homogeneous with unit rate on , where is the total opacity. We draw points uniformly in , their number being Poisson distributed with mean equal to the length of the interval in , and map them back with . Stretches without medium () receive no centers. In Mecke’s formula, .
Case 1: spectrum integrated over
By Eq. (33), , where only depends on the centers before . By Eqs. (36) and (35), with the width of the marked center integrated exactly,
with , and where just before is generated by the centers in . One sample consists of: (a) uniform in (stratified in the code) and ; (b) Poisson centers in , uniform in , each with ; (c) the Möbius recursion between centers and at each center, from up to ; and (d) . Centers after are not sampled. The exact integration over the width of the marked center removes the largest source of variance, the Coulomb tail (). As a further control variate we use the same estimator without earlier centers, i.e., the first order in opacity, whose mean is computed by quadrature.
Cases 2 and 3: with a cut and differential
In Eq. (26) there remain the integral over and the sum over the marked center . Applying Eq. (36) to and passing to units of , , so that
With , . One sample consists of: (a) log-uniform in and uniform in , with and weight ; (b) the centers between the emission and the marked center, Poisson, uniform in , with their ; starting from at , they are propagated with the free stretches and the jumps (16), which gives , and, with the of the marked center, ; (c) the centers after the marked center, Poisson, uniform in , which only enter through ; (d) the centers before the emission, which do not contribute and are not sampled; and (e) the estimators
with and , so that
Here the width of the marked center cannot be integrated in closed form, since it enters and, through it, inside the exponential; it is sampled.
Log-uniform sampling in and truncation error
If is small, is large and, without intermediate centers, over a wide range of . With uniform the estimator would be unbounded; with the weight of the log-uniform sampling it is bounded. The log-uniform sampling needs a lower limit, and the region is omitted, with
This is a bias, not a statistical error, and we quantify it.
Physical argument. The region only contributes when , i.e., for a tiny : a center with a huge width , which gives a transverse kick . The condition states that is the formation time of a gluon with that momentum, . At the boundary , that is, , which is of order for the parameters below. Such a kick is so hard that the gluon comes out with , outside the kinematic cut and outside the soft-gluon approximation in which the whole formula is written.
Size. At first order in opacity (a single center, the marked one) the omitted part can be computed exactly by quadrature. For a static medium with the kinematic cut , the omitted fraction ranges from to for between 0.01 and 10 and –. Without cut (, ) it is , in agreement with the analytic estimate divided by the first order. With several centers the order of magnitude should be the same, since the omitted region is dominated by the hard kicks of a single center.
Control variate, all at once, and real part
Control variate. In the same sample we compute , the estimator without cut, whose mean is the integrated spectrum . We take from an independent run of the estimator of § 5 C, , and return with . The variance is reduced by a factor , and the error of is propagated, . Since the case-1 estimator is much more precise than , the control variate is effective at little cost.
All at once. only enters the last factor of Eq. (41), so every configuration serves a whole grid of , and the covariance between different is obtained as well.
Real part, sample by sample. Re and are linear and commute, so we take the real part of each sample before averaging. This has two advantages: the variance of does not exceed that of , and the imaginary vacuum constant from (Sec. III C) disappears in each sample. What cannot be done is to take the real part inside the computation of a configuration, since Re does not commute with products, quotients or logarithms. The oscillator is genuinely complex, . and are propagated as complex numbers, and their phases are the LPM interference.
Physical variables and conventions for expanding media
Time is measured from the production point, and there may be vacuum before the medium; reversing the profile in time gives a different result. Following [17], we use with , so that with , and the kinematic cut corresponds to with (static medium: ).
Cost
The cost per sample does not depend on the profile, which only changes the Poisson intensity and . On a laptop (two physical cores), samples of the estimator of § 5 take about 3 s and give a relative error of –; samples of the estimators (41), for a grid of six values of , take 3–5 s, with relative errors of about 1% for the differential spectrum. The tasks are independent, with a single final reduction; with two processes the speed-up is 1.6.
Validation and comparison with published results
All numbers in this section are obtained with the publicly available code, with the statistical errors shown. The code and the scripts that produce the figures were written with the assistance of an AI tool (see Acknowledgments) and validated as described in this section. The curves of Refs. [5, 17] are read from the vector graphics of the PDF files, which gives the curves with the precision of the drawing; as a check of this reading and of the normalization, the published first-order (GLV) curves of Ref. [5] are reproduced to 0.9996–1.0002 with and .
Internal checks
The closed-form integral of Sec. III C agrees with a direct quadrature of the integrand to ; the multilinear polynomial (17) and the determinant of Sec. IV B agree with the recursions to machine precision; and the integrated spectrum obtained from the estimator of § 5 C agrees with the limit of the estimator of § 5 D, which is a different estimator of the same quantity. The integral over of the differential estimator reproduces the estimator with a cut sample by sample. The bias from the truncation was discussed in § 5 E.
First order in opacity
Table II compares the first order in opacity of the master formula, computed by quadrature, with the closed expressions (B.3) and (B.4) of Ref. [5] for the differential spectrum and for the spectrum with the kinematic cut. Without cut, the first order of Sec. IV C reproduces the GLV result to twelve digits.
| master formula | Ref. [5] | |||
| cut, | 0.05 | 0.0625 | 0.0532891 | 0.0532891 |
| 1 | 25 | 0.2943896 | 0.2943896 | |
| 5 | 625 | 0.1201906 | 0.1201826 | |
| cut, | 0.05 | 1.25 | 0.7862587 | 0.7862587 |
| 1 | 500 | 0.5025957 | 0.5025957 | |
| 5 | 12500 | 0.1361501 | 0.1361425 | |
| differential | 0.03 | 0.03 | 1.2358926 | 1.2358926 |
| 0.3 | 1 | 0.1096906 | 0.1096906 | |
| 1 | 10 | 0.0178158 | 0.0178158 |
Table II. First order in opacity (static medium, ): master formula at (quadrature) versus Eqs. (B.4) (cut, ) and (B.3) (differential) of Ref. [5], per unit of .
Comparison with the results of Andrés et al.
The master formula (26) has the same structure as Eq. (3.3) of Ref. [5]: the integral over is done analytically, one scattering is singled out, the kernel runs from the emission to that scattering and the broadening from it onwards, and the integrand vanishes outside the medium. Through Mecke’s formula (§ 5 A), the sum over becomes the integral , the analogue of the insertion in that equation.
Static medium. Figure 1 compares the spectrum for a static medium with , without cut and with the kinematic cut for several values of , with Fig. 3 of Ref. [5]. Without cut, their curve lies 0.1–0.4 % below our result.
Figure 1. Static medium, : spectrum without cut and with the kinematic cut for several . Lines: Fig. 3 of Ref. [5]; points: this work. Lower panel: relative difference, Ref. [5]/this work , in %. Published points below the frame of the published panel () are not compared. The figure was produced with scripts written with the assistance of an AI tool (see Sec. VI and Acknowledgments).
our result. With the cut, their curves lie 0.7–1.2 % below our results for , while at small they lie above, by an amount that grows as and decrease: up to 8.6 % for and 16 % for in the range shown. The statistical errors are below 0.35 %.
Differential spectrum. Figure 2 compares the spectrum differential in with the curves of Figs. 1 and 2 of Ref. [5], for and , at 56 points each. Their values lie below ours by 1.05% on average (rms 1.1%) for and by 1.4% (rms 1.6%, maximum 4.8%) for , with a mean statistical error of 0.5%.
!Spectrum differential in : relative difference between the curves of Figs. 1 () and 2 () of Ref. [5] and this work bbox=0.15,0.08,0.85,0.30
Figure 2. Spectrum differential in : relative difference between the curves of Figs. 1 () and 2 () of Ref. [5] and this work, Ref. [5]/this work − 1 in %, at the published points; the color gives . The figure was produced with scripts written with the assistance of an AI tool (see Sec. VI and Acknowledgments).
Expanding media. Figure 3 compares our results with Fig. 3 of Ref. [17] for : a static medium, a power-law profile, and the hydrodynamic profile read from the inset of that figure, with time measured from ; with this choice the equivalent static medium has , close to the value quoted there. Following that reference, the power-law profile is , , with and , normalized to , and the hydrodynamic one is , with and along the temperature profile of the inset, which gives . Without cut, their curves lie 0–1.6% below our results. With the kinematic cut, at small their curves for the expanding media lie well above ours: by 46% (power law) and 41% (hydrodynamic profile) at , decreasing to 16 % and 13 % at and below 2 % for ; for the static medium the difference at is 3 %. At large their curves with cut fall below ours, by 4–9 % at –10 and by about 20 % at for the expanding media (6 % and 18 % at and 30 for the static one). The statistical errors are below 0.25 %.
Summary and Outlook
We have computed the BDMPS-Z spectrum in the soft-gluon limit for a Yukawa potential, to all orders in opacity, by writing the Yukawa dipole cross section as a Laplace superposition of Gaussians. This turns the medium into a Poisson gas of Gaussian scattering centers, each configuration of which is a harmonic oscillator with an impulsive frequency, solved in closed form. The result is a master formula for the spectrum differential in and with a cut on , valid for arbitrary profiles and , whose limit is the average of the logarithm of a Gelfand–Yaglom determinant. Its Monte Carlo evaluation is efficient, gives the whole dependence from a single run, and is trivially parallelizable. It reproduces the first order in opacity and the published results without cut, while with a kinematic cut we find sizeable relative deviations from the published results, both at small gluon energy fraction, where the cut strongly suppresses the spectrum, and at large energy fraction, where they reach 5–20 %.
We expect the construction not to rely on the soft limit in an essential way: what it requires is that, configuration by configuration, the problem is Gaussian. We expect it to extend to the more complex building blocks of an in-medium parton cascade, such as finite energy fraction and emissions off color antennas (for the latter, see Ref. [26] for a recent numerical treatment). The master formula also provides a natural starting point for analytical approximations, which we will address in future work. Finally, the independence of the samples makes the method well suited to massively parallel hardware.
Data Availability
The code used for all numerical results of this article is openly available at https://github.com/ manoel-maker/poisson-gas-bdmps_z. The scripts that produce the figures and the curves read from Refs. [5, 17] are available from the author upon reasonable request.
Acknowledgments
The author used Claude (Anthropic; several model versions, through claude.ai and Claude Code, between June and November 2026) as a research assistant. The representation of the Yukawa dipole cross section as a Laplace superposition of Gaussians, the central derivations, the choice of what to compute and to include, and the conclusions are the author’s; the Monte Carlo strategy was developed in the work with Claude. Claude was used to check derivations, including independent numerical checks; to write and validate the Monte Carlo code and the scripts that produce the figures (see Sec. VI); to search for and verify references; and to draft the English text of parts of the manuscript from the author’s notes. The abstract and § 1 and § 2 follow the author’s own drafts. The author checked the derivations by hand, reviewed the code and the text, and takes full responsibility for the content.

Figure 3 (PDF p. 11). Static, power-law and hydrodynamic media with , without cut (left) and with the kinematic cut and (right). Lines: Fig. 3 of Ref. [17]; points: this work. Lower panels: relative difference, Ref. [17]/this work , in %. The figure was produced with scripts written with the assistance of an AI tool (see Sec. VI and Acknowledgments).
Continuous route: Jacobi equation and Arnold’s propagator
The exponent of in Eq. (11) is with . The classical equation is , with as in Eq. (14); in Arnold’s notation [19], with . The Jacobi solution satisfies
with and . Between two centers is linear in , and at each center it is continuous while its slope jumps as in Eq. (16). The products of with the solution are well defined without any prescription: is continuous at , the discontinuity is in , which never multiplies a , and integrating Eq. (43) over gives Eq. (16).
Since the exponent is quadratic, the path integral is the exponential of the classical action times the Van Vleck prefactor.
In two transverse dimensions,
with and . For the classical path is and the classical action is . Comparing with Eq. (13) at gives , and , which is Eq. (15): both routes define the same . Then
which is Eq. (3.3) of Arnold [19] with , , and Eq. (19) follows. In this route the continuity of is checked directly: when crosses , decreases by and increases by , so that changes by and does not change.
HTL potential with a magnetic mass
The HTL potential [7] is, up to a constant prefactor ,
The term , from the unscreened magnetostatic gluons, makes the total elastic cross section logarithmically divergent in the infrared, so that the Poisson intensity of § 2 C would be infinite. A magnetic mass regularizes it,
which has the form of the effective cross section of Ref. [27] (their Eq. (9), with , ). Writing ,
a superposition of Gaussians whose weight is positive for every if and only if . Performing the Gaussian integral over and changing variables to ,
The total cross section is finite, and with and
This is a probability density, positive if and only if , with the same Coulomb tail as in the Yukawa case, for . Everything in § 2 C–V then applies with and : only the sampling of the widths changes. The condition imposed in Ref. [27] is precisely the positivity of this Poisson gas. For the intensity diverges logarithmically: the unregularized HTL potential is still a superposition of Gaussians with positive weights, but not a Poisson gas with finite intensity. The magnetic mass is of order and nonperturbative; Ref. [27] takes from nonperturbative estimates.
References
- [1]R. Baier, Y. L. Dokshitzer, A. H. Mueller, S. Peigné, and D. Schiff, Radiative energy loss of high-energy quarks and gluons in a finite volume quark-gluon plasma, Nucl. Phys. B 483, 291 (1997), hep-ph/9607355.DOI
- [2]R. Baier, Y. L. Dokshitzer, A. H. Mueller, S. Peigné, and D. Schiff, Radiative energy loss and $p_t$ broadening of high-energy partons in nuclei, Nucl. Phys. B 484, 265 (1997), hep-ph/9608322.DOI
- [3]B. G. Zakharov, Fully quantum treatment of the Landau-Pomeranchuk-Migdal effect in QED and QCD, JETP Lett. 63, 952 (1996), hep-ph/9607404.DOI
- [4]B. G. Zakharov, Radiative energy loss of high-energy quarks in finite size nuclear matter and quark gluon plasma, JETP Lett. 65, 615 (1997), hep-ph/9704255.DOI
- [5]C. Andres, L. Apolinário, and F. Dominguez, Medium-induced gluon radiation with full resummation of multiple scatterings for realistic parton-medium interactions, JHEP 07 (2020), 114, 2002.01517.DOI
- [6]M. Gyulassy and X.-N. Wang, Multiple collisions and induced gluon bremsstrahlung in QCD, Nucl. Phys. B 420, 583 (1994), nucl-th/9306003.DOI
- [7]P. Aurenche, F. Gelis, and H. Zaraket, A simple sum rule for the thermal gluon spectral function and applications, JHEP 05 (2002), 043, hep-ph/0204146.
- [8]M. Gyulassy, P. Levai, and I. Vitev, Reaction operator approach to non-Abelian energy loss, Nucl. Phys. B 594, 371 (2001), nucl-th/0006010.DOI
- [9]M. Gyulassy, P. Levai, and I. Vitev, Non-Abelian energy loss at finite opacity, Phys. Rev. Lett. 85, 5535 (2000), nucl-th/0005032.DOI
- [10]U. A. Wiedemann, Gluon radiation of hard quarks in a nuclear environment: opacity expansion, Nucl. Phys. B 588, 303 (2000), hep-ph/0005129.DOI
- [11]C. A. Salgado and U. A. Wiedemann, Calculating quenching weights, Phys. Rev. D 68, 014008 (2003), hep-ph/0302184.DOI
- [12]Y. Mehtar-Tani, Gluon bremsstrahlung in finite media beyond multiple soft scattering approximation, JHEP 07 (2019), 057, 1903.00506.
- [13]Y. Mehtar-Tani and K. Tywoniuk, Improved opacity expansion for medium-induced parton splitting, JHEP 06 (2020), 187, 1910.02032.
- [14]J. Barata and Y. Mehtar-Tani, Improved opacity expansion at NLO for medium induced gluon radiation, JHEP 10 (2020), 176, 2004.02323.DOI
- [15]J. H. Isaksen, A. Takacs, and K. Tywoniuk, A unified picture of medium-induced radiation, JHEP 02 (2023), 156, 2206.02811.DOI
- [16]S. Caron-Huot and C. Gale, Finite-size effects on the radiative energy loss of a fast parton in hot and dense strongly interacting matter, Phys. Rev. C 82, 064902 (2010), 1006.2379.
- [17]C. Andres, L. Apolinário, F. Dominguez, and M. Gonzalez Martinez, In-medium gluon radiation spectrum with all-order resummation of multiple scatterings in longitudinally evolving media, JHEP 11 (2024), 025, 2307.06226.DOI
- [18]X. Feal and R. Vazquez, Intensity of gluon bremsstrahlung in a finite plasma, Phys. Rev. D 98, 074029 (2018), 1811.01591.DOI
- [19]P. B. Arnold, Simple formula for high-energy gluon bremsstrahlung in a finite, expanding medium, Phys. Rev. D 79, 065025 (2009), 0808.2767.DOI
- [20]D. V. Widder, The Laplace Transform, Princeton Mathematical Series, Vol. 6 (Princeton University Press, Princeton, 1941).
- [21]J. Schwinger, On gauge invariance and vacuum polarization, Phys. Rev. 82, 664 (1951).
- [22]G. Last and M. Penrose, Lectures on the Poisson Process, IMS Textbooks (Cambridge University Press, Cambridge, 2017).DOI
- [23]I. M. Gel’fand and A. M. Yaglom, Integration in functional spaces and its applications in quantum physics, J. Math. Phys. 1, 48 (1960).DOI
- [24]F. James, Monte Carlo theory and practice, Rep. Prog. Phys. 43, 1145 (1980).
- [25]J. Mecke, Stationäre zufällige Maße auf lokal-kompakten Abelschen Gruppen, Z. Wahrsch. verw. Geb. 9, 36 (1967).
- [26]C. Andres, L. Apolinário, N. Armesto, A. Cordeiro, F. Domínguez, P. Guerrero-Rodríguez, and J. G. Mil-hano, Gluon radiation from a QCD antenna with realistic parton-medium interactions, arXiv:2607.00629 [hep-ph] (2026).
- [27]M. Djordjevic and M. Djordjevic, Generalization of radiative jet energy loss to non-zero magnetic mass, Phys. Lett. B 709, 229 (2012), 1105.4359.DOI

