WorksPhysics and Astronomy
Bayesian constraints on the transport coefficients and from spin polarization in relativisitic heavy-ion collisions
Abstract
Bayesian analyses in the context of relativistic heavy-ion collisions have so far relied almost exclusively on bulk hadronic observables constructed from momentum degrees of freedom to constrain the transport properties of the quark-gluon plasma. In this work, we perform the Bayesian inference after incorporating the longitudinal spin polarization of hyperons alongside conventional bulk measurements in Pb+Pb collisions at TeV to constrain the shear and bulk viscosity to entropy density ratios, and . We demonstrate that the inclusion of spin polarization, which provides complementary sensitivity to the space-time structure and vorticity of the medium, shifts the posterior distribution of toward larger values, although current uncertainties do not allow a statistically significant separation at the 68% credibility level. Nevertheless, the results establish spin polarization as a valuable probe in quantitative studies of QGP transport properties and indicate that it should be incorporated in comprehensive and systematically constrained Bayesian extractions of the medium's dynamical parameters.
Bayesian constraints on the transport coefficients and from spin polarization in relativisitic heavy-ion collision
Introduction
The remarkable success of relativistic hydrodynamics in describing collective flow observables in heavy-ion collisions has led to the conclusion that the quark–gluon plasma (QGP) behaves as an almost perfect fluid [1–3]. In particular, comparisons of hydrodynamic simulations with elliptic flow measurements indicate that the specific shear viscosity, , is very small and close to the conjectured lower bound [4], making the QGP the most perfect fluid observed in nature. Despite this phenomenological success, the transport coefficients of QCD matter, including both the shear viscosity and the bulk viscosity , have not yet been computed reliably from first principles in the temperature regime relevant for heavy-ion collisions. The strongly coupled and non-perturbative nature of QCD near the confinement scale renders perturbative approaches inadequate, while lattice calculations of transport coefficients remain notoriously challenging [5, 6]. Nevertheless, these coefficients play a central role in governing the space-time evolution of the plasma and are therefore essential for a quantitative characterization of the medium.
To address this difficulty, the community has adopted a top-down strategy in which transport coefficients are inferred by systematically comparing model calculations with experimental data. In recent years, Bayesian inference techniques have emerged as a powerful framework for this purpose [7, 8, 9, 10, 11, 12, 13, 14], enabling a statistically rigorous extraction of these coefficients with quantified uncertainties from a broad set of observables measured at RHIC and the LHC. These analyses have primarily relied on hadronic observables such as multiplicities, mean transverse momenta and anisotropic flow coefficients, which are sensitive to the viscous response of the medium. Such global Bayesian calibrations have significantly improved our understanding of the temperature dependence of and , yet they have so far been restricted to momentum-space observables and have not incorporated spin-related measurements.
In parallel with its near-perfect fluidity, the QGP has also been identified as the most vortical fluid ever observed, as evidenced by the global spin polarization of hyperons measured in heavy-ion collisions [15]. The description of spin observables such as global polarization and spin alignment has stimulated intense theoretical activity from different perspectives (see reviews [16, 17, 18, 19] and references therein).
Recent studies have shown that polarization in heavy-ion collisions is sensitive to the medium’s transport properties [20], with the longitudinal polarization component, , being particularly sensitive to the bulk viscosity and initial flow structure at LHC energies [21]. Moreover, extensions of kinetic theory incorporating spin degrees of freedom indicate that spin polarization itself can modify effective transport coefficients, including bulk and shear viscosity, thereby influencing the thermodynamic evolution of the fluid [22]. These developments motivate the inclusion of spin polarization measurements in a unified Bayesian analysis of QGP transport coefficients, as their demonstrated sensitivity to bulk viscosity can provide additional and independent constraints beyond conventional measurements constructed from momentum degrees of freedom. In this work, we perform a systematic Bayesian study of and by incorporating longitudinal polarization data as an additional constraint alongside traditional hadronic observables. Because the computation of spin polarization requires full -dimensional hydrodynamic simulations and a realistic event-by-event treatment is computationally demanding, we carry out our analysis using event-averaged initial profiles. A detailed event-by-event treatment including pre-equilibrium dynamics, while important, is left for future work. To make the parameter estimation tractable, we construct a Gaussian process emulator trained on a carefully designed set of hydrodynamic calculations, which is subsequently employed within a Markov Chain Monte Carlo (MCMC) framework to extract posterior distributions of the transport coefficients. This study provides the first Bayesian constraints on QGP viscosities that explicitly incorporate spin polarization observables, thereby opening a new avenue for probing the dissipative properties of strongly interacting matter.
Notation: We work in natural units where . We adopt the mostly negative metric convention, . Our code is written in Milne coordinates, , expressed in terms of Cartesian coordinates as defined as and . The inner product of two four-vectors is denoted by . The fluid velocity is denoted by such that . The covariant derivative of a quantity is denoted by . The comoving derivative of a quantity is denoted by . The derivative orthogonal to the 4-velocity is written as . For the Levi-Civitá tensor , we follow the sign convention .
Description of the model
The hydrodynamic evolution is governed by energy-momentum and net baryon number conservation, and which are expressed through the following continuity equations:
Here is the energy-momentum tensor, and is the (net baryon) charge current and the is the covariant derivative of the spacetime. We choose to work in the Landau frame, defined by
with the local energy density and the fluid velocity. With this choice, the energy-momentum tensor and the net baryon current are generically written as
Here is the energy density, is the net-baryon number density, is the thermodynamic equilibrium pressure, is the shear-stress tensor, is the bulk viscous pressure, and is the baryon diffusion current, and is the projector transverse to the fluid velocity . In this study, we consider for simplicity. The equations of motion for and are taken as [23, 24]
where and are corresponding Navier-Stokes limit given by
Here and denote the shear and bulk viscosities, respectively. These are treated as temperature-dependent quantities with no dependence. We parameterize these transport coefficients using the following functional forms [11]
with , and
In particular, we introduce three free parameters each to characterize and . The remaining second-order transport coefficients entering the equations (6)–(7) are not treated as independent parameters. Instead, they are fixed using expressions derived in kinetic theory [24, 25], but evaluated consistently with the chosen equation of state (discussed below). This approach reduces the dimensionality of the parameter space while retaining a physically motivated description of the system. The remaining transport coefficients are expressed below:
We employ the lattice-QCD-based equation of state NEOS-BQS [26, 27, 28, 29, 30, 31] to close the hydrodynamic equations. The square of the speed of sound is computed from the equation of state as follows
As mentioned above, we do not explicitly model the pre-equilibrium stage in this work. Instead, we assume that the system has locally thermalized at a proper time , which is treated as a free parameter, after which a hydrodynamic description becomes applicable. The initial energy-momentum tensor at is constructed using the model described in Refs [32, 33], and is given by
where is a parameter controlling the fraction of longitudinal momentum converted into fluid velocity and is defined as
In the above expressions, and are nucleus thickness functions computed from the positions of participant nucleons using the following expression
where denotes the transverse coordinates of the nucleon and is the width of the nucleon.
For symmetric collisions, the initial energy density is given by
with defined as [32]
Here denotes the Heaviside function. is obtained through the matching condition at the initial time, which results in an expression , with expressed in terms of and . The choice of initial energy-momentum tensor results in zero transverse velocity, and the -component of the velocity is determined numerically by solving the implicit equation
For the initial baryon density, we consider the following profile
where and are given by [34]:
The normalization constant is chosen by demanding that
where denotes the total number of participants. The above condition gives
We fix the parameter . For the net charge density profile, we consider
The hydrodynamic evolution is stopped at the particlization (or switching) hypersurface, defined by a constant energy density . This hypersurface is obtained using the CORNELIUS code [35]. At this hypersurface, we use a hadron sampler [36, 37, 38] to generate particles from fluid elements. The resulting particle set serves as input to the SMASH transport model [39, 40] for subsequent hadron interactions and decays. We compute the spin polarization of -hyperons on the switching hypersurface, including the contributions from thermal vorticity,
and thermal shear,
following the prescription of Refs. [41, 42].1 The spin four-vector is written as
with
In the above equation, is the relativistic Fermi-Dirac distribution with four-temperature and the appropriate combination of chemical potentials for the , hyperons; in the shear contribution, is the time unit vector in the centre-of-mass collision frame. At very high collision energies, the baryon chemical potential is very small and satisfies , such that energy-density becomes primarily a function of temperature. In this limit, the constant energy density switching hypersurface can be approximated by a constant temperature hypersurface after neglecting the gradients of temperature on the hypersurface. This allows the temperature to be factored out of the surface integrals in the density operator. This is known as the isothermal approximation [42]. Under this approximation, defining and , Eqs. (24) reduce to
The spin vector in the rest frame, is obtained from the laboratory frame spin vector as follows:
where is the energy in the laboratory frame. Finally, the spin polarization vector is given by . For simplicity and to reduce computational cost, feed-down contributions from resonance decays are neglected in the present analysis. This is a reasonable approximation, as it has been shown that the inclusion of feed-down corrections does not lead to significant changes in the resulting polarization relative to the case of entirely primary production [21].
Model Parameters
A full -dimensional event-by-event simulation of Pb+Pb collisions at TeV is computationally demanding, particularly for central collisions where the fireball lifetime is comparatively long and particle multiplicities are large. The computational cost is further amplified by the particlization (sampler) stage and the subsequent hadronic transport evolution. To render the Bayesian analysis feasible while maintaining a physically well-motivated framework, we adopt a set of simplifications. We neglect pre-equilibrium dynamics and initialize hydrodynamics at a fixed proper time . The initial condition model contains six parameters and is constructed from thickness functions in which the smoothing of local hotspots is regulated by the nucleon width . For each parameter set, we generate multiple fluctuating initial configurations to obtain the energy-momentum tensor for individual events, and subsequently perform event averaging to produce a smooth profile for hydrodynamic evolution. The transport sector includes six parameters describing the temperature dependence of and (three each), together with the switching energy density, , defining the transition from hydrodynamics to the hadronic afterburner, resulting in a total of fifteen model parameters.
We find that two parameters, and , have negligible impact on the calibrated observables within reasonable ranges; these are therefore fixed to and , reducing the effective dimensionality for the analysis to thirteen parameters. For example, in Figure 1, we show the dependence of the lower bound of the scaled multiplicity range on for various centrality classes. The bound remains approximately constant over a wide range of , demonstrating that the centrality selection is largely insensitive to this parameter. We adopt uniform prior distributions for all model parameters within their respective ranges, reflecting minimal prior knowledge. The prior ranges for the parameters are summarized in Table I.

Figure 1. Scaled multiplicity vs nucleon width
| Parameter | Prior range (min–max) | MAP value with (rounded to two decimal places) | MAP value without (rounded to two decimal places) |
| 1.5–3.5 | 2.54 | 2.50 | |
| 1.0–2.5 | 1.38 | 1.44 | |
| 4.0–6.0 | 5.15 | 4.57 | |
| 1.8–3.0 | 2.14 | 2.74 | |
| 0.0–0.4 | 0.17 | 0.16 | |
| [fm] | 0.6–1.2 | 0.83 | 1.0 |
| 0.05–0.25 | 0.07 | 0.06 | |
| 0.0–3.0 | 0.76 | 1.44 | |
| -1.0–1.0 | 0.07 | 0.23 | |
| 0.0–0.3 | 0.16 | 0.12 | |
| 0.0–0.3 | 0.19 | 0.07 | |
| 0.1–0.3 | 0.16 | 0.15 | |
| [GeV fm] | 0.35–0.65 | 0.54 | 0.48 |
Table I. Prior range and MAP values of the parameters
Experimental data
We consider nine sets of experimental data across seven centrality classes: 0-10%, 10-20%, 20-30%, 30-40%, 40-50%, 50-60%, and 60-70%. The observables included in the Bayesian analysis are listed below: (i) Charged particle pseudorapidity density, from Refs. [45, 46],
(ii) Identified-particle rapidity densities, , , and from Ref. [46],
(iii) Mean transverse momentum of identified particles , , and from Ref. [46],
(iv) Elliptic flow coefficient of charged particles from Ref. [47],
(v) Longitudinal spin polarization of hyperons from Ref. [48].
The experimental data listed in items (i)-(iv) are not available in the 0-10% centrality class but are reported separately for the 0-5% and 5-10% centrality bins. We therefore merge these two bins to construct the corresponding 0-10% observables following the procedure described below.
The combined experimental values of and are taken as the arithmetic mean of the corresponding values in the two centrality classes, while the associated uncertainty is taken to be the larger of the two quoted uncertainties. For the mean transverse momentum, the value in the merged centrality class is computed as
where and denote the 0-5% and 5-10% centrality classes, respectively. The corresponding uncertainty is taken as the larger of the two quoted uncertainties.
For the elliptic flow coefficient , an additional step is required. Our hydrodynamic simulations employ event-averaged smooth initial profiles and therefore predict the mean elliptic flow , while experimental measurements include contributions from event-by-event flow fluctuations. To make the experimental data compatible with our model, we follow Ref. [49] and characterize the magnitude of flow fluctuations by a variance . The elliptic flow coefficients measured using the two- and four-particle cumulant methods satisfy
From these relations, the mean elliptic flow is obtained as
Measurements of and are available for Pb+Pb collisions at TeV [47]. We use these measurements to extract in each experimental centrality bin, with uncertainties propagated in quadrature. The values in the 0–5% and 5–10% classes are then merged using the same procedure as for the mean transverse momentum.
Computational details
The initial condition model employed in this work satisfies the scaling at high collision energies. The TRENTo model [50] provides this behavior for the reduced thickness parameter . We therefore use the TRENTo code [51] to construct the initial conditions and to perform centrality selection.
We generate minimum-bias events in the impact-parameter range 0–20 fm using the following TRENTo configuration:
cross-section = 7.0
normalization = 1.0
reduced-thickness = 0.0
fluctuation = 1.0
nucleon-min-dist = 0.0
nucleon-width = 0.9
b-min = 0.0
b-max = 20.0
grid-max = 10.0
grid-step = 0.2
For each event, we store the transverse coordinates of the participating nucleons and compute the integrated reduced thickness . We assume the final-state charged-particle multiplicity to be proportional to this quantity and sort the events in descending order of the integrated reduced thickness. Since centrality classes are defined by percentiles of the multiplicity distribution, the overall normalization (set here to unity) does not affect the centrality determination.
Centrality classes are defined by selecting fixed percentiles of the sorted event ensemble in intervals of 10%. For each centrality class, we select 5000 events from the minimum-bias sample. For each selected event, the nuclear thickness functions are constructed using Eq. (15), and the corresponding energy–momentum tensor is evaluated. The initial condition for the hydrodynamic evolution is then obtained by averaging over all events in a given centrality class, yielding a smooth, event-averaged energy–momentum tensor .
The viscous hydrodynamics is solved numerically in full dimensions using the code developed in Refs. [20, 52, 53]. The code, written in Fortran, employs the Godunov-type relativistic Harten-Lax-van Leer-Einfeldt (HLLE) approximate Riemann solver to compute numerical fluxes at fluid cell interfaces, as described in Refs. [54, 55]. Although inspired by the publicly available vHLLE framework [55, 56], the code includes several extensions, such as multi-threaded parallelization via OpenMP [57] and the evolution of spin degrees of freedom within a spin-hydrodynamic framework [58]. For the present analysis, only the background viscous hydrodynamic evolution is employed.
Our analysis is based on a finite ensemble of model evaluations performed at design points in the model parameter space. The design matrix is constructed using maximin Latin hypercube sampling (LHS), which provides an approximately space-filling coverage of the multidimensional parameter space [7, 11, 59]. The corresponding simulation outputs contain the predicted observables for Pb+Pb collisions at TeV. The observable vector entering the Bayesian analysis has dimension , corresponding to nine observables evaluated in seven centrality classes (refer § 1.4).
Specifically, the Latin hypercube samples are generated in the unit hypercube with using the lhs package in R [60], and are subsequently mapped to the physical model parameters by linearly scaling each dimension to its corresponding parameter range. This procedure corresponds to assuming independent uniform prior distributions for all model parameters within the specified bounds. The design consists of a total of 800 parameter sets. Fixed random seeds are used in the LHS construction to ensure reproducibility. We verify that the chosen prior ranges (listed in Table I) encompass the experimental data, as illustrated in Figure 3-Figure 5.
Figure 3. Top panel shows the model predictions obtained from parameter values sampled uniformly from the prior ranges for Pb+Pb collisions at TeV. Bottom panel compares the experimental data with model predictions evaluated at the median of the posterior parameter distribution in the analysis including spin polarization. The band represents the corresponding posterior predictive uncertainty.
Figure 5. Left panel shows the model predictions obtained from parameter values sampled uniformly from the prior ranges for Pb+Pb collisions at TeV for longitudinal component of spin polarization. The right panel instead shows the experimental data with model predictions evaluated at the median of the posterior parameter distribution in the analysis including spin polarization for the longitudinal polarization.
The design points and their corresponding model outputs are stored in separate data files and are read into the analysis framework as input-output pairs. Prior to training the emulator, design points for which one or more observables are undefined or contain non-numerical values are removed from the dataset. In total, seven such design points are excluded, leaving valid samples.
The filtered dataset is divided into a training set and an independent validation set. A subset of 100 design points is randomly selected and reserved exclusively for validation, while the remaining samples are used for training the Gaussian process model emulator (discussed below). The random selection is performed using a fixed random seed to ensure reproducibility of the training-validation split. The training dataset is used to construct the emulator, while the validation dataset is employed to test the emulator’s accuracy and robustness.
The model outputs are standardized to have zero mean and unit variance for each observable. This preprocessing is performed using a linear rescaling,
where and denote the sample mean and standard deviation of the observable across the training dataset. To reduce the dimensionality of the observable space and remove correlations among observables, we perform a principal component analysis (PCA) on the standardized model outputs using the scikit-learn Python library [61]. The full PCA spectrum is first computed to quantify the fraction of total variance carried by each principal component. We retain the leading principal components, which together account for a 97.06% of the total variance when performing analysis including spin polarization and 98.19% when performing analysis excluding spin polarization. The remaining variance is incorporated into the likelihood function as truncation uncertainty, denoted as , and therefore does not bias the parameter inference. Including additional principal components was found to degrade the emulator calibration, as the subleading modes are dominated by statistical noise and are not reliably constrained by the training data, leading to unstable Gaussian-process hyperparameters and spurious constraints in the Bayesian analysis.
The PCA yields a transformation matrix of dimension which operates on standardized model outputs and projects them onto the retained principal components. In the present analysis , resulting in a three-dimensional principal-component representation. The transformation reads
The experimental observables are treated consistently within the same linear framework. Let denote the vector of experimental mean values and let denotes the covariance matrix of experimental uncertainties in observable space, assumed diagonal due to the absence of bin-to-bin correlations. The experimental observables are first standardized
Projection onto principal-component space yields
and the corresponding covariance matrix transforms as
The log-likelihood function used for Bayesian inference is given by
where the total covariance in principal-component space is
with denoting the emulator predictive covariance in principal-component space. Each retained principal component is emulated independently using a Gaussian process regressor (GPR) from the scikit-learn library [61, 62]. We employ a composite kernel consisting of a constant kernel multiplied by a squared-exponential (RBF) kernel, augmented by a white-noise term to account for residual numerical noise [63]. The characteristic length scales of the RBF kernel are initialized and bounded according to the physical ranges of the model parameters. Emulator performance is assessed using a train-test split of the design set. The coefficient of determination implemented as GPR.score in the scikit-learn library for the three retained principal components is found to be 99.86%, 96.88%, and 95.97%, respectively for the analysis with spin-polarization data. After confirming satisfactory predictive performance, the Gaussian process emulator is retrained using the full design dataset for the subsequent Bayesian analysis. The accuracy of the full emulator is further evaluated by reconstructing the observables at the validation design points (,). Predictions obtained in principal-component space are mapped back to observable space using the in-
verse PCA transformation followed by inverse scaling. Emulator performance is quantified using the root-mean-square error (RMSE) in observable space defined as
where denotes the model output. When normalized by the standard deviation of the validation data in observable space, we obtain an overall normalized RMSE of 3.5%.
To further assess emulator robustness, we perform a five-fold cross-validation of the Gaussian-process emulators in principal-component space. The design set is randomly partitioned into five subsets, and for each split the emulator is trained on four subsets and evaluated on the remaining one. For each principal component, we analyze the normalized residuals (-scores) [64], defined as the difference between the true and emulated values divided by the predicted emulator uncertainty. For a well-calibrated emulator, these residuals are expected to follow a standard normal distribution. Our results for the aggregated -score distribution is shown in the top panel of Figure 2 and is found to be consistent with a standard normal distribution. We additionally evaluate coverage probabilities, confirm
Figure 2. (Top) Distribution of normalized residuals (-scores) for the retained principal components. The dashed curve indicates a standard normal distribution. (Bottom) Comparison between the true model values and emulator predictions for the three principal components in the analysis including spin polarization. The diagonal line denotes perfect agreement.
Figure 4. Top panel shows the model predictions obtained from parameter values sampled uniformly from the prior ranges for Pb+Pb collisions at TeV for and . Bottom panel compares the experimental data with model predictions evaluated at the median of the posterior parameter distribution in the analysis including spin polarization for and . The band represents the corresponding posterior predictive uncertainty.
ing that approximately 69.4% and 95% of the residuals fall within the corresponding one- and two-sigma confidence intervals. Finally, direct comparison between true and emulated principal-component values are performed to verify the absence of systematic bias as shown in the bottom panel of Figure 2.
The trained Gaussian-process emulators are used to predict the principal components for any parameter vector , providing both the mean and diagonal emulator covariance. Predictions are mapped back to observable space via inverse PCA and inverse standardization, with emulator uncertainties propagated accordingly, including a truncation term for discarded variance, as discussed above. Uniform priors are assumed within the predefined parameter bounds, and the likelihood is defined as a multivariate Gaussian in principal-component space, incorporating emulator, experimental, and truncation covariances. Posterior distributions of the model parameters are sampled using the parallel-tempered MCMC algorithm implemented in ptemcee library [65, 66], with multiple temperatures and walkers to ensure efficient exploration. After burn-in, the converged chains are used to infer the posterior distributions of all model parameters.
Results
To illustrate the outcome of the Bayesian parameter inference, we evaluate the emulator at the median of the posterior parameter distribution and compare the resulting model predictions with experimental data across all observables and centrality classes as shown in Figure 3-Figure 5. The emulator provides the mean prediction in observable space together with the associated emulator uncertainty, obtained by propagating Gaussian-process uncertainties from principal-component space and including the PCA truncation variance. The shaded bands in the figures indicate the resulting uncertainty on the model prediction. For comparison, we also show predictions obtained from parameter values sampled uniformly within the prior ranges, highlighting the degree to which the data constrain the model parameters. These comparisons demonstrate that the posterior-median parameter set yields a significantly improved description of the measured observables.
To assess the overall consistency of the inferred model with experimental data, we perform posterior predictive checks. While the posterior-median comparison described above illustrates the predictions corresponding to a representative inferred parameter set, posterior predictive checks probe the full predictive distribution of the model and provide a consistency test of the Bayesian inference. Parameter samples are drawn from the posterior distribution and propagated through the trained emulator to generate replicated datasets in the observable space. The resulting ensemble of replicated observables is used to construct pointwise posterior predictive intervals, defined by the 5th and 95th percentiles, together with the median prediction. These intervals are compared directly to the experimental measurements across all observables and centrality classes, as shown in Figure 6. Good agreement of the data with the posterior predictive bands indicates that the inferred parameter distributions provide a statistically consistent description of the measured observables. We emphasize that this procedure tests the full predictive distribution of the model rather than a single best-fit parameter set.
Figure 6. Posterior predictive checks for the Bayesian analysis of Pb+Pb collisions at TeV. The shaded bands correspond to the 5th–95th percentile intervals of the posterior predictive distribution obtained by propagating posterior parameter samples through the emulator trained including spin polarization.. The solid line indicates the median prediction. Experimental data are shown for comparison.
Using the posterior samples obtained from the Bayesian analysis, we reconstruct the temperature dependence of the shear and bulk viscosity coefficients in Figure 7-Figure 8. For each coefficient, we evaluate the functions Eqs. (8)-(9) over a dense temperature grid using a large ensemble of posterior samples and construct pointwise credibility intervals. The median and maximum-a-posteriori (MAP) estimates are shown alongside the 60%, 90%, and 100% credible intervals. We find that the extracted shear viscosity is essentially unchanged when longitudinal spin polarization data are included in the Bayesian analysis, indicating that
Figure 7. Posterior distributions of the temperature-dependent shear viscosity obtained from the analysis with and without including in the left and right panels, respectively. The median estimate is shown as a solid line and the maximum-a-posteriori (MAP) estimate as a dashed line. The shaded bands correspond to the 60%, 90%, and 100% credible intervals, with decreasing color intensity indicating increasing credibility level.
Figure 8. Posterior distributions of the temperature-dependent bulk viscosity obtained from the analysis with and without including in the left and right panels, respectively. The median estimate is shown as a solid line and the maximum-a-posteriori (MAP) estimate as a dashed line. The shaded bands correspond to the 60%, 90%, and 100% credible intervals, with decreasing color intensity indicating increasing credibility level.
Figure 10. Marginalized one- and two-dimensional posterior distributions of selected model parameters inferred from MCMC sampling using an emulator trained on bulk and spin polarization observable. The diagonal panels show one-dimensional marginal posteriors, while the off-diagonal panels display the corresponding two-dimensional joint posterior distributions. Dashed lines indicate posterior median values, and the quoted uncertainties represent central 90% Bayesian credible intervals.
bulk hadronic observables already provide strong constraints on shear transport. In contrast, the median value of the bulk viscosity is significantly enhanced-by approximately a factor of two-when spin polarization data are incorporated, although the MAP estimates remain comparable in the two analyses. Notably, both parameter analyses yield a satisfactory
Figure 11. Marginalized one- and two-dimensional posterior distributions of selected model parameters inferred from MCMC sampling using an emulator trained on only bulk observables without spin. The diagonal panels show one-dimensional marginal posteriors, while the off-diagonal panels display the corresponding two-dimensional joint posterior distributions. Dashed lines indicate posterior median values, and the quoted uncertainties represent central 90% Bayesian credible intervals.
description of conventional bulk observables, suggesting that spin polarization provides complementary sensitivity primarily to bulk viscous effects rather than shear transport. This behavior is consistent with the expectation that longitudinal component of spin polarization is more sensitive to the space-time structure of the expansion dynamics of the medium, which are directly influenced by bulk viscosity.
To visualize the inferred parameter correlations and the extent to which individual model parameters are constrained by the data, we construct corner (pairwise correlation) plots from the posterior samples, shown in Figure 9–Figure 12. For clarity, the parameters are separated into two groups: parameters associated with the initial condition model, initial hydrodynamic time , and switching energy density ; and parameters governing the temperature dependence of the shear and bulk viscosities. For each group, a random subset of posterior samples is used to generate kernel-density estimates of the one- and two-dimensional marginal distributions, with median values and central 90% credible intervals indicated.
Figure 9. Marginalized one- and two-dimensional posterior distributions of selected model parameters inferred from MCMC sampling using an emulator trained on bulk and spin polarization observable. The diagonal panels show one-dimensional marginal posteriors, while the off-diagonal panels display the corresponding two-dimensional joint posterior distributions. Dashed lines indicate posterior median values, and the quoted uncertainties represent central 90% Bayesian credible intervals.
Figure 12. Marginalized one- and two-dimensional posterior distributions of selected model parameters inferred from MCMC sampling using an emulator trained on only bulk observables without spin. The diagonal panels show one-dimensional marginal posteriors, while the off-diagonal panels display the corresponding two-dimensional joint posterior distributions. Dashed lines indicate posterior median values, and the quoted uncertainties represent central 90% Bayesian credible intervals.
We observe that parameters primarily related to the initial condition and hydrodynamic initialization exhibit nearly flat marginal distributions with broad, multi-modal structures and weak correlations. This indicates that, within the present analysis, these parameters are only weakly constrained by the considered experimental observables. Such behavior is expected, as the bulk hadronic observables and longitudinal spin polarization included here are predominantly sensitive to the medium’s transport properties rather than to detailed features of the early-time dynamics or particlization hypersurface.
In contrast, the posterior distributions of the parameters related to transport coefficients remain broadly consistent between the two analyses. The inclusion of longitudinal spin polarization does not lead to a substantial narrowing or broadening of the credible intervals, or to dramatic changes in the correlation structure among transport parameters. However, it does shift the inferred bulk-viscosity distribution toward larger values. In particular, the median of increases by approximately a factor of two when data are incorporated, while the MAP estimate exhibits a modest upward shift. The corresponding 68% credible intervals retain significant overlap beyond MeV, indicating that the two analyses remain statistically compatible within uncertainties. These results suggest that longitudinal spin polarization provides complementary sensitivity to bulk viscous effects. In contrast, the extracted shear viscosity remains largely unchanged, reinforcing the conclusion that conventional bulk observables already provide strong constraints on shear transport.
As an additional consistency check of our analysis, we use the median parameter values extracted from the corresponding posterior distributions to compute differential observables, including the transverse momentum spectra of pions, kaons, and protons, as well as the transverse momentum dependence of the charged-particle elliptic flow coefficient , shown in Figure 13 and Figure 14. We find that these differential hadronic observables are equally well described by both sets of extracted parameters, namely those obtained from analyses with and without spin polarization data. We additionally evaluate the transverse momentum and azimuthal angle dependence of the longitudinal spin polarization shown in Figure 15. The results indicate that the experimentally observed sign of is correctly reproduced only by the parameter set obtained from the Bayesian analysis that includes spin polarization observables. This parameter set corresponds to a bulk viscosity approximately a factor of two larger than that obtained in the analysis without spin observable. These findings further emphasize that spin polarization observables provide complementary and nontrivial constraints on the bulk viscosity of the QCD medium.
Figure 13. Transverse momentum spectra of (left) charged pions and (right) charged kaons in Pb-Pb collisions at TeV for different centrality classes. Shown are model results obtained using median values of parameters extracted from Bayesian analyses with (solid lines) and without (dashed lines) inclusion of spin polarization observable. Experimental data is taken from Ref. [46].

Figure 14. Transverse momentum spectra of (left) protons and anti-protons and (right) the elliptic flow coefficient, , of charged hadrons in Pb-Pb collisions at TeV for different centrality classes. The model results are obtained using the median parameter values extracted from Bayesian analyses with (solid lines) and without (dashed lines) the inclusion of spin polarization observable. The experimental data for the -spectra is taken from Ref. [46]. The data for and from Ref. [47] are used to construct mean , as discussed in § 1.4, and is shown in the right panel.
Figure 15. (left) Transverse momentum and (right) azimuthal angle dependence of longitudinal polarization for Pb-Pb collisions at TeV in the 30-50% centrality. Shown are model results obtained using median bulk viscosity values extracted from Bayesian analyses with and without inclusion of spin polarization observable.
Finally, we compare the extracted transport coefficients with results from other analyses in Figure 16. The shear viscosity shown in the left panel is broadly consistent with existing results in the literature. In contrast, significantly larger variations are observed among the reported results for the bulk viscosity. The bulk viscosity obtained in our analysis without spin observable is in close agreement with the JETSCAPE result. However, the bulk viscosity extracted from the analysis including spin polarization observable is comparatively larger than existing results shown in the figure. Such a larger bulk viscosity appears to be necessary in order to reproduce the experimentally observed sign of the longitudinal polarization.
Figure 16. Comparison of the extracted transport coefficients obtained from the present Bayesian analyses with and without spin observables (solid and dashed lines, respectively) with results from J. Bernhard et al. [8] (dash-dotted line), the TRAJECTUM framework [11] (dotted line), and the JETSCAPE collaboration [9] (dash-double-dotted line). The present results, the Bernhard et al. analysis, and the JETSCAPE results employ median posterior parameter values, while the TRAJECTUM result corresponds to mean parameter values.
Summary
In this work, we have performed a Bayesian analysis of Pb+Pb collisions at TeV using a Gaussian-process emulator trained on a finite ensemble of model evaluations. The predictive accuracy of the emulator was assessed by partitioning the design ensemble into training and validation subsets. The Gaussian-process emulator was trained on the training set and evaluated on the validation set, demonstrating nice agreement between predicted and true principal-component values. Posterior inference was carried out both with and without the inclusion of longitudinal spin polarization of hyperons in order to quantify its impact on the extraction of QGP transport properties.
We find that the extracted shear viscosity is largely unchanged when spin polarization data are included, indicating that conventional bulk hadronic observables already provide strong constraints on shear transport. In contrast, the inferred bulk viscosity exhibits an upward shift when is incorporated. In particular, the posterior median increases by approximately a factor of two, while the MAP estimate shows a modest increase. Although the credible intervals of the two analyses retain significant overlap, this shift indicates that longitudinal spin polarization provides additional sensitivity to bulk viscous effects.
Overall, our results demonstrate that spin polarization measurements provide complementary information to conventional bulk observables in Bayesian extractions of QGP transport coefficients, particularly for bulk viscosity. These findings motivate more comprehensive future simulations and global analyses in which polarization observables are systematically incorporated alongside traditional hadronic measurements to further refine constraints on the dissipative properties of the quark–gluon plasma.
Acknowledgements
This work and S.K.S. were supported by ICSC - Centro Nazionale di Ricerca in High Performance Computing, Big Data and Quantum Computing, funded by European Union - NextGenerationEU and by the project PRIN2022 Advanced Probes of the Quark Gluon Plasma funded by ”Ministero dell’Università e della Ricerca”. We further acknowledge the ICSC for awarding this project access to the EuroHPC supercomputer LEONARDO, hosted by CINECA (Italy). S.K.S also thank Jhilam Sadhukhan for helpful discussions and assistance with debugging the analysis script.
References
References
- [4]P. K. Kovtun, D. T. Son, and A. O. Starinets, Viscosity in strongly interacting quantum field theories from black hole physics, Phys. Rev. Lett. 94, 111601 (2005).pubmed.ncbi.nlm.nih.gov/15903845
- [5]N. Astrakhantsev, V. Braguta, and A. Kotov, Temperature dependence of shear viscosity of SU(3)–gluodynamics within lattice simulation, JHEP 04, 101, arXiv:1701.02266 [hep-lat].DOI
- [6]L. Altenkort, A. M. Eller, A. Francis, O. Kaczmarek, L. Mazur, G. D. Moore, and H.-T. Shu, Viscosity of pure-glue qcd from the lattice, Phys. Rev. D 108, 014503 (2023).DOI
- [7]J. E. Bernhard, J. S. Moreland, S. A. Bass, J. Liu, and U. Heinz, Applying bayesian parameter estimation to relativistic heavy-ion collisions: Simultaneous characterization of the initial state and quark–gluon plasma medium, Phys. Rev. C 94, 024907 (2016).DOI
- [8]J. E. Bernhard, J. S. Moreland, and S. A. Bass, Bayesian estimation of the specific shear and bulk viscosity of quark–gluon plasma, Nature Phys. 15, 1113 (2019).DOI
- [9]D. Everett et al. (JETSCAPE Collaboration), Multisystem bayesian constraints on the transport coefficients of qcd matter, Phys. Rev. C 103, 054904 (2021).DOI
- [10]J. E. Parkkila, A. Onnerstad, and D. J. Kim, Bayesian estimation of the specific shear and bulk viscosity of the quark-gluon plasma with additional flow harmonic observables, Phys. Rev. C 104, 054904 (2021).DOI
- [11]G. Nijs, W. van der Schee, U. Gürsoy, and R. Snellings, Bayesian analysis of heavy ion collisions with the heavy ion computational framework trajectum, Phys. Rev. C 103, 054909 (2021).DOI
- [12]L. Vermunt, Y. Seemann, A. Dubla, S. Floerchinger, E. Grossi, A. Kirchner, S. Masciocchi, and I. Selyuzhenkov, Mapping properties of the quark gluon plasma in Pb-Pb and Xe-Xe collisions at energies available at the CERN Large Hadron Collider, Phys. Rev. C 108, 064908 (2023), arXiv:2308.16722 [hep-ph].
- [13]S. A. Jahan, H. Roch, and C. Shen, Bayesian analysis of (3 + 1)D relativistic nuclear dynamics with the rhic beam energy scan data, Phys. Rev. C 110, 054905 (2024).arxiv.org/abs/2408.00537
- [14]N. Götz, I. Karpenko, and H. Elfner, Bayesian analysis of a (3 + 1)D hybrid approach with initial conditions from hadronic transport, Phys. Rev. C 112, 014910 (2025).arxiv.org/abs/2503.10181
- [15]L. Adamczyk et al. (STAR), Global Λ hyperon polarization in nuclear collisions: evidence for the most vortical fluid, Nature 548, 62 (2017), arXiv:1701.06657 [nucl-ex].
- [16]F. Becattini and M. A. Lisa, Polarization and Vorticity in the Quark–Gluon Plasma, Ann. Rev. Nucl. Part. Sci. 70, 395 (2020), arXiv:2003.03640 [nucl-ex].DOI
- [17]F. Becattini, Spin and polarization: a new direction in relativistic heavy ion physics, Rept. Prog. Phys. 85, 122301 (2022), arXiv:2204.01144 [nucl-th].DOI
- [18]W.-B. Dong, X.-L. Sheng, Y.-L. Yin, and Q. Wang, Global polarization and spin alignment in heavy-ion collisions: past, present and future, PoS SPIN2023, 229 (2024), arXiv:2401.15576 [nucl-th].arxiv.org/abs/2401.15576
- [19]F. Becattini, M. Buzzegoli, T. Niida, S. Pu, A.-H. Tang, and Q. Wang, Spin polarization in relativistic heavy-ion collisions, Int. J. Mod. Phys. E 33, 2430006 (2024), arXiv:2402.04540 [nucl-th].arxiv.org/abs/2402.04540
- [20]S. K. Singh and J. Alam, Suppression of spin polarization as an indicator of QCD critical point, Eur. Phys. J. C 83, 585 (2023).DOI
- [21]A. Palermo, E. Grossi, I. Karpenko, and F. Becattini, A polarization in very high energy heavy ion collisions as a probe of the quark–gluon plasma formation and properties, Eur. Phys. J. C 84, 920 (2024), arXiv:2404.14295 [nucl-th].arxiv.org/abs/2404.14295
- [22]D.-X. Wei, Impact of spin polarization on transport and thermodynamic coefficients, Phys. Rev. C 112, 064921 (2025).DOI
- [23]G. S. Denicol, H. Niemi, E. Molnar, and D. H. Rischke, Derivation of transient relativistic fluid dynamics from the Boltzmann equation, Phys. Rev. D 85, 114047 (2012), [Erratum: Phys.Rev.D 91, 039902 (2015)], arXiv:1202.4551 [nucl-th].DOI
- [24]G. S. Denicol, S. Jeon, and C. Gale, Transport coefficients of bulk viscous pressure in the 14-moment approximation, Phys. Rev. C 90, 024912 (2014).DOI
- [25]E. Molnár, H. Niemi, G. S. Denicol, and D. H. Rischke, Relative importance of second-order terms in relativistic dissipative fluid dynamics, Phys. Rev. D 89, 074010 (2014).DOI
- [26]A. Monnai, B. Schenke, and C. Shen, Equation of state at finite densities for qcd matter in nuclear collisions, Phys. Rev. C 100, 024907 (2019).DOI
- [27]A. Bazazov et al. (HotQCD), Equation of state in ( 2+1 )-flavor QCD, Phys. Rev. D 90, 094503 (2014), arXiv:1407.6387 [hep-lat].
- [28]A. Bazazov et al. (HotQCD), Fluctuations and Correlations of net baryon number, electric charge, and strangeness: A comparison of lattice QCD results with the hadron resonance gas model, Phys. Rev. D 86, 034509 (2012), arXiv:1203.0784 [hep-lat].
- [29]H. T. Ding, S. Mukherjee, H. Ohno, P. Petreczky, and H. P. Schadler, Diagonal and off-diagonal quark number susceptibilities at high temperatures, Phys. Rev. D 92, 074043 (2015), arXiv:1507.06637 [hep-lat].
- [30]A. Bazavov et al., The QCD Equation of State to O(μ_B^6) from Lattice QCD, Phys. Rev. D 95, 054504 (2017), arXiv:1701.04325 [hep-lat].DOI
- [31]https://sites.google.com/view/qcdneos/.
- [32]C. Shen and S. Alzhrani, Collision-geometry-based 3D initial condition for relativistic heavy-ion collisions, Phys. Rev. C 102, 014909 (2020), arXiv:2003.05852 [nucl-th].DOI
- [33]S. Ryu, V. Jupic, and C. Shen, Probing early-time longitudinal dynamics with the Λ hyperon’s spin polarization in relativistic heavy-ion collisions, Phys. Rev. C 104, 054908 (2021).DOI
- [34]G. S. Denicol, C. Gale, S. Jeon, A. Monnai, B. Schenke, and C. Shen, Net baryon diffusion in fluid dynamic simulations of relativistic heavy-ion collisions, Phys. Rev. C 98, 034916 (2018), arXiv:1804.10557 [nucl-th].DOI
- [35]P. Huovinen and H. Petersen, Particlization in hybrid models, The European Physical Journal A 48, 171 (2012).DOI
- [36]I. A. Karpenko, P. Huovinen, H. Petersen, and M. Bleicher, Estimation of the shear viscosity at finite net-baryon density from A + A collision data at √sNN = 7.7 − 200 GeV, Phys. Rev. C 91, 064901 (2015), arXiv:1502.01978 [nucl-th].DOI
- [37]A. Schäfer, I. Karpenko, X.-Y. Wu, J. Hammelmann, and H. Elfner, Particle production in a hybrid approach for a beam energy scan of au+au/pb+pb collisions between √sNN = 4.3 gev and √sNN = 200.0 gev, The European Physical Journal A 58, 230 (2022).
- [38]https://github.com/smash-transport/smash-hadron-sampler.
- [39]J. Weil et al., Particle production and equilibrium properties within a new hadron transport approach for heavy-ion collisions, Phys. Rev. C 94, 054905 (2016).DOI
- [40]A. Wergieluk et al., smash-transport/smash: Smash-3.1 (2024).
- [41]F. Becattini, G. Inghirami, V. Rolando, A. Beraudo, L. Del Zanna, A. De Pace, M. Nardi, G. Pagliara, and V. Chandra, A study of vorticity formation in high energy nuclear collisions, Eur. Phys. J. C 75, 406 (2015), [Erratum: Eur.Phys.J.C 78, 354 (2018)], arXiv:1501.04468 [nucl-th].DOI
- [42]F. Becattini, M. Buzzegoli, and A. Palermo, Spin-thermal shear coupling in a relativistic fluid, Phys. Lett. B 820, 136519 (2021), arXiv:2103.10917 [nucl-th].DOI
- [43]S. Y. F. Liu and Y. Yin, Spin polarization induced by the hydrodynamic gradients, JHEP 07, 188, arXiv:2103.09200 [hep-ph].DOI
- [44]X.-L. Sheng, F. Becattini, and D. Roselli, An improved formula for Wigner function and spin polarization in a decoupling relativistic fluid at local thermodynamic equilibrium (2025), arXiv:2509.14301 [nucl-th].DOI
- [45]J. Adam et al. (ALICE Collaboration), Centrality dependence of the charged-particle multiplicity density at midrapidity in pb-pb collisions at √sNN = 5.02 TeV, Phys. Rev. Lett. 116, 222302 (2016).DOI
- [46]S. Acharya et al. (ALICE Collaboration), Production of charged pions, kaons, and (anti-)protons in pb-pb and inelastic pp collisions at √sNN = 5.02 tev, Phys. Rev. C 101, 044907 (2020).DOI
- [47]J. Adam et al. (The ALICE Collaboration), Anisotropic flow of charged particles in pb-pb collisions at √sNN = 5.02 TeV, Phys. Rev. Lett. 116, 132302 (2016).DOI
- [48]S. Acharya et al. (ALICE Collaboration), Polarization of Λ and Λ̄ hyperons along the beam direction in pb-pb collisions at √sNN = 5.02 TeV, Phys. Rev. Lett. 128, 172005 (2022).
- [49]J.-Y. Ollitrault, A. M. Poskanzer, and S. A. Voloshin, Effect of flow fluctuations and nonflow on elliptic flow methods, Phys. Rev. C 80, 014904 (2009).
- [50]J. S. Moreland, J. E. Bernhard, and S. A. Bass, Alternative ansatz to wounded nucleon and binary collision scaling in high-energy nuclear collisions, Phys. Rev. C 92, 011901(R) (2015).DOI
- [51]https://github.com/Duke-QCD/trento.
- [52]S. K. Singh, R. Ryblewski, and W. Florkowski, Spin dynamics with realistic hydrodynamic background for relativistic heavy-ion collisions, Phys. Rev. C 111, 024907 (2025), arXiv:2411.08223 [hep-ph].DOI
- [53]S. K. Singh and J.-e. Alam, Effects of the qcd critical point on the spectra and flow coefficients of hadrons, Phys. Rev. D 107, 074042 (2023).DOI
- [54]D. H. Rischke, S. Bernard, and J. A. Maruhn, Relativistic hydrodynamics for heavy-ion collisions. i. general aspects and expansion into vacuum, Nuclear Physics A 595, 346 (1995).DOI
- [55]I. Karpenko, P. Huovinen, and M. Bleicher, A 3+1 dimensional viscous hydrodynamic code for relativistic heavy ion collisions, Comput. Phys. Commun. 185, 3016 (2014), arXiv:1312.4160 [nucl-th].DOI
- [56]https://github.com/yukarpenko/vhlle/tree/stable_ebe.
- [57]https://www.openmp.org/.
- [58]Sapna, S. K. Singh, and D. Wagner, Spin polarization of Λ hyperons from dissipative spin hydrodynamics, Phys. Rev. C 112, 054902 (2025).DOI
- [59]M. Stein, Large sample properties of simulations using latin hypercube sampling, Technometrics 29, 143 (1987), https://www.tandfonline.com/doi/pdf/10.1080/00401706.1987.10488205.DOI
- [60]R. Carnell, lhs, https://github.com/bertcarnell/lhs.
- [61]F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and Édouard Duchesnay, Scikit-learn: Machine learning in python, Journal of Machine Learning Research 12, 2825 (2011).arxiv.org/abs/1201.0490
- [62]GPy, GPy: A gaussian process framework in python, http://github.com/SheffieldML/GPy (2014).
- [63]J. E. Bernhard, Bayesian parameter estimation for relativistic heavy-ion collisions, Ph.D. thesis, Duke U. (2018), arXiv:1804.06469 [nucl-th].arxiv.org/abs/1804.06469
- [64]T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. (Springer, New York, 2009).
- [65]D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman, emcee: The mcmc hammer, Publications of the Astronomical Society of the Pacific 125, 306 (2013).
- [66]W. D. Vousden, W. M. Farr, and I. Mandel, Dynamic temperature selection for parallel tempering in markov chain monte carlo simulations, Monthly Notices of the Royal Astronomical Society 455, 1919 (2015), https://academic.oup.com/mnras/article-pdf/455/2/1919/18514064/stv2422.pdf.DOI













