Tidal Adaptive Softening in Cosmological Simulations

• Tidal adaptive softening is an adaptive gravitational force-resolution technique that sets softening lengths based on the local tidal field’s anisotropy rather than density alone.
• It employs methods such as tidal-tensor-based prescriptions and anisotropy-gated refinement (e.g., NovA) to prevent spurious fragmentation in anisotropic structures like filaments and sheets.
• Applications in cosmological simulations demonstrate improved force accuracy and conservative dynamics, though its efficacy in fully eliminating artificial halo formation remains limited.

Searching arXiv for recent and foundational papers on tidal adaptive softening and closely related adaptive force-softening methods. Searching arXiv for tidal adaptive softening, NovA, and adaptive gravitational softening. Tidal adaptive softening denotes a class of adaptive gravitational force-resolution schemes for collisionless $N$-body and related cosmological simulations in which the effective softening length, or the refinement level that determines force resolution, is regulated by the local anisotropy or strength of the gravitational tidal field rather than by density alone. In the contemporary literature, the term covers two closely related lines of development. One is an explicitly tidal-tensor-based prescription, in which each particle’s softening is set from the norm of the local tidal tensor, as in the GIZMO implementation of conservative adaptive softening (Hopkins et al., 2022) and its cosmological assessment in warm dark matter simulations (Paun et al., 22 Jul 2025). The other is an anisotropy-gated adaptive mesh refinement strategy, exemplified by NovA in RAMSES, which withholds refinement in sheets and filaments and allows it only when collapse is sufficiently isotropic, thereby making the effective softening conservative in regions of one- or two-dimensional collapse (Hobbs et al., 2015). Across these variants, the shared objective is to prevent the force resolution from becoming spuriously fine in anisotropic structures, where discreteness noise and two-body scattering seed artificial fragmentation.

1. Conceptual definition and scope

Tidal adaptive softening is motivated by a specific failure mode of conventional collisionless simulations: discretization of a collisionless fluid with finite-mass macroparticles generates localized force perturbations that can dominate the physical small-scale signal when the matter power spectrum is suppressed, especially in warm dark matter cosmologies (Paun et al., 22 Jul 2025). In standard formulations with fixed or density-only adaptive softenings, force resolution often tightens as soon as local density increases, even if the physical collapse is highly anisotropic. In sheets and filaments this causes the softening to track the shortest spacing while remaining too small along the long axis, amplifying two-body noise and producing the familiar “beads-on-a-string” fragmentation (Hobbs et al., 2015).

The central idea of tidal adaptive softening is that force resolution should respond to the geometry of collapse set by the tidal field. If matter is compressing along all principal directions, finer resolution is justified; if compression occurs along only one or two directions, refinement should remain conservative (Hobbs et al., 2015). This logic can be implemented directly through the tidal tensor, using the local Hessian of the gravitational potential to set a particle softening scale (Hopkins et al., 2022, Paun et al., 22 Jul 2025), or indirectly through an isotropy test on the particle configuration inside an AMR cell, using the inertia tensor as a proxy for local tidal anisotropy (Hobbs et al., 2015). In both cases, the purpose is not merely adaptive resolution in the generic sense, but adaptive resolution conditioned on anisotropic collapse.

This distinguishes tidal adaptive softening from density-based adaptive softening, including earlier conservative schemes in GADGET that adapt the softening length as $h \propto \rho^{-1/3}$ via a target neighbor number (Iannuzzi et al., 2011). Those methods improve spatial resolution in dense regions and can sharpen inner profiles, but they are not tidal-field adaptations and do not explicitly suppress refinement in filamentary or sheet-like collapse (Iannuzzi et al., 2011).

2. Mathematical formulations

Two mathematical formulations dominate the literature.

The explicitly tidal formulation defines the tidal tensor as

$$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$

or, in the GIZMO-based formulation,

$T = - \nabla \otimes \nabla \hat{\Phi},$

where $\hat{\Phi}$ is a local potential estimator (Hopkins et al., 2022, Paun et al., 22 Jul 2025). A scalar tidal strength is then constructed from the Frobenius norm,

$||T||^2 = \sum_i \lambda_i^2,$

with $\lambda_i$ the eigenvalues of $T$ (Paun et al., 22 Jul 2025). The softening rule is

$$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$

with $\xi \sim 2$ in the GIZMO implementation (Paun et al., 22 Jul 2025). The 2022 conservative formulation further introduces smooth floor or ceiling bounds through

$h \propto \rho^{-1/3}$0

with $h \propto \rho^{-1/3}$1 and $h \propto \rho^{-1/3}$2 typical (Hopkins et al., 2022). This construction is intended to preserve differentiability when imposing minimum or maximum softenings.

The NovA formulation in RAMSES does not define $h \propto \rho^{-1/3}$3 directly from $h \propto \rho^{-1/3}$4. Instead, it computes the moment of inertia tensor of the particle distribution within each cell about the cell’s centroid,

$h \propto \rho^{-1/3}$5

diagonalizes it to obtain eigenvalues $h \propto \rho^{-1/3}$6, and forms the anisotropy measure

$h \propto \rho^{-1/3}$7

A cell is refined only if

$h \propto \rho^{-1/3}$8

where $h \propto \rho^{-1/3}$9 is a tolerance above unity calibrated from shot-noise realizations (Hobbs et al., 2015). For $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$0, the reported values are $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$1 and $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$2 (Hobbs et al., 2015). This refinement gate embodies the rule “refine only if collapse is occurring along all three axes” (Hobbs et al., 2015).

The physical interpretation given for both approaches is similar. In the Zel’dovich picture, sheets correspond to collapse along one principal direction, filaments to collapse along two, and halos to collapse along three (Hobbs et al., 2015). A tidal-aware method therefore aims to reduce $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$3 only when the local state resembles bona fide three-dimensional collapse.

3. Conservative dynamics and implementation frameworks

A major issue with any adaptive softening law is conservation. If the softening varies in space and time, the equations of motion must be modified to account for the induced dependence of the gravitational Lagrangian on particle positions through the softening field. Density-based conservative softening in GADGET derived this through a symmetrized Lagrangian and explicit $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$4 correction terms (Iannuzzi et al., 2011). The same conservative logic was generalized to arbitrary softening rules, including tidal softening, in the GIZMO formulation (Hopkins et al., 2022).

For adaptive softening in GADGET, the Lagrangian is written as

$$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$5

and the resulting equations of motion contain the standard symmetric softened force plus correction terms involving

$$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$6

and

$$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$7

(Iannuzzi et al., 2011). The paper stresses that omitting the correction term causes loss of energy conservation and biases the low-mass halo population (Iannuzzi et al., 2011).

The 2022 GIZMO treatment extends this beyond density-based rules. Starting from a softened pairwise potential $$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$8, the acceleration is

$$T_{ij} \equiv \frac{\partial^2 \Phi}{\partial x_i \partial x_j},$$9

where $T = - \nabla \otimes \nabla \hat{\Phi},$0 and $T = - \nabla \otimes \nabla \hat{\Phi},$1 are the energy- and momentum-conserving “grad-$T = - \nabla \otimes \nabla \hat{\Phi},$2” terms required for arbitrary adaptive softenings (Hopkins et al., 2022). For the tidal rule, these corrections depend on derivatives of the local tidal estimator and on

$T = - \nabla \otimes \nabla \hat{\Phi},$3

(Hopkins et al., 2022).

The same work also emphasizes symmetrization of pairwise softenings. It argues that force-averaging and simple softening-averaging can produce large errors or over-scattering when $T = - \nabla \otimes \nabla \hat{\Phi},$4, and recommends a “maximum-softening” symmetrization,

$T = - \nabla \otimes \nabla \hat{\Phi},$5

or smooth approximations that converge to this limit (Hopkins et al., 2022). The implementation is public in GIZMO (Hopkins et al., 2022).

4. Relation to local inter-particle spacing and anisotropic collapse

A recurring theoretical theme is the relation between softening and inter-particle spacing. In adaptive schemes tied to density, a common rule is

$T = - \nabla \otimes \nabla \hat{\Phi},$6

with $T = - \nabla \otimes \nabla \hat{\Phi},$7 the local mean inter-particle spacing (Hobbs et al., 2015). For isotropic collapse this is natural, because all principal spacings are comparable. In anisotropic collapse, however, the relevant spacings differ by direction. If the principal axis extents are associated with the eigenvalues of the inertia tensor, then a density-only criterion effectively ties $T = - \nabla \otimes \nabla \hat{\Phi},$8 to the shortest spacing, which is numerically hazardous in filaments and sheets (Hobbs et al., 2015).

NovA’s anisotropy gate is designed to prevent this. By withholding refinement until $T = - \nabla \otimes \nabla \hat{\Phi},$9, it keeps the effective softening of order twice the largest inter-particle spacing in the cell,

$\hat{\Phi}$0

rather than allowing it to undershoot along the long axis (Hobbs et al., 2015). This is why NovA is described as a “tidal-aware” adaptive softening method even though the algorithm is expressed through $\hat{\Phi}$1 instead of $\hat{\Phi}$2 (Hobbs et al., 2015).

The explicitly tidal rule reaches a similar endpoint from a different direction. In a smooth Keplerian background with enclosed mass $\hat{\Phi}$3 at scale $\hat{\Phi}$4, the tidal norm is estimated as

$\hat{\Phi}$5

leading to a Hill-radius scaling $\hat{\Phi}$6 (Hopkins et al., 2022). For a cubic spline kernel, the reported relation is $\hat{\Phi}$7 (Hopkins et al., 2022). In isotropic cusps with local density $\hat{\Phi}$8, the tidal norm tends to

$\hat{\Phi}$9

so that

$||T||^2 = \sum_i \lambda_i^2,$0

that is, of order the local mean inter-particle spacing (Hopkins et al., 2022). This suggests that the tidal rule interpolates between a Hill-radius criterion in external tides and a density-like spacing criterion when local self-gravity dominates.

5. Numerical performance in idealized and cosmological tests

The numerical record is mixed and depends strongly on the problem class.

In controlled anisotropic-collapse tests, tidal-aware methods show clear gains. NovA was tested on the asymmetric perturbed plane wave or “Valinia” collapse and produced no numerical fragmentation while still capturing fine caustics and shells around the collapsing regions (Hobbs et al., 2015). In the same class of idealized filamentary-collapse tests, the 2025 cosmological study reports that tidal adaptive softening in GIZMO significantly improves force accuracy over fixed softening (Paun et al., 22 Jul 2025). Using an anti-symmetric perturbed plane-wave collapse with $||T||^2 = \sum_i \lambda_i^2,$1, $||T||^2 = \sum_i \lambda_i^2,$2, $||T||^2 = \sum_i \lambda_i^2,$3, a $||T||^2 = \sum_i \lambda_i^2,$4 PM mesh, and $||T||^2 = \sum_i \lambda_i^2,$5, the authors find that TAS yields approximately $||T||^2 = \sum_i \lambda_i^2,$6 more particles with zero spurious $||T||^2 = \sum_i \lambda_i^2,$7-shift than fixed softening, and suppresses large artificial displacements along the unperturbed axis (Paun et al., 22 Jul 2025).

In warm dark matter cosmological runs, NovA also improves over standard RAMSES. For a thermal relic with $||T||^2 = \sum_i \lambda_i^2,$8 in cosmologies with $||T||^2 = \sum_i \lambda_i^2,$9, $\lambda_i$0, $\lambda_i$1, and $\lambda_i$2, NovA converges more rapidly than standard mass-only refinement and produces little or no spurious halos on small scales, with nearly an order-of-magnitude fewer low-mass halos below $\lambda_i$3 at $\lambda_i$4 compared to standard runs (Hobbs et al., 2015). However, the same work notes that AHF likely mislabels some caustics and criss-crossing filaments as halos and that one or two particularly massive filaments appear to fragment in any NovA version where refinement is allowed (Hobbs et al., 2015).

By contrast, the 2025 study of Mostoghiu Paun et al. reports that explicit tidal adaptive softening does not substantially reduce spurious halos in full cosmological WDM simulations when paired with optimized initial conditions (Paun et al., 22 Jul 2025). Their production runs use monofonIC with 3LPT, anti-aliasing via Orszag’s $\lambda_i$5-rule padding, optional PLT corrections, a late start at $\lambda_i$6, $\lambda_i$7, $\lambda_i$8 dark matter particles, a $\lambda_i$9 PM mesh, and $T$0 of the mean inter-particle spacing at the initial conditions (Paun et al., 22 Jul 2025). In these runs, both fixed and TAS cases show the characteristic low-mass upturn in the cumulative halo mass function near $T$1, indicating spurious halos (Paun et al., 22 Jul 2025). The ratio TAS/fixed peaks around that scale, with TAS yielding up to $T$2 more halos than fixed, corresponding to approximately $T$3 near the 100-particle mass (Paun et al., 22 Jul 2025). Table-based counts obtained with an empirical proto-halo sphericity cut give 10,452 total WDM objects with 5,973 flagged as spurious in the fixed-softening run, versus 13,060 total with 8,004 flagged as spurious in the TAS run (Paun et al., 22 Jul 2025).

The same study also reports a measurable shift in WDM halo formation times under TAS. Using the half-maximum mass scale factor $T$4, the mean values in WDM are $T$5 and $T$6, a difference $T$7; no analogous shift is found in CDM, where the means are both approximately $T$8 (Paun et al., 22 Jul 2025). This indicates that the softening prescription can bias inferred formation histories even when it does not cure fragmentation.

6. Tidal adaptive softening for subhalo tidal evolution

A more recent and distinct use of the term concerns the numerical convergence of tidally stripped subhalos rather than artificial filament fragmentation. In this context, tidal adaptive softening means adapting force resolution to the instantaneous tidal radius of each subhalo (Chiang et al., 30 Oct 2025). The key result is a universal force-resolution criterion: in AMR simulations, the minimum instantaneous cell size must satisfy

$T$9

and in tree codes the Plummer-equivalent softening must satisfy

$$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$0

(Chiang et al., 30 Oct 2025). Here $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$1 is the minimum tidal radius experienced since infall.

The tidal radius is defined through

$$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$2

with $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$3, $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$4 the instantaneous enclosed bound mass profile of the remnant, and $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$5 the host-halo potential (Chiang et al., 30 Oct 2025). The minimum tidal radius is

$$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$6

Across AMR simulations spanning subhalo concentrations $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$7, constant anisotropies $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$8, infall mass ratios $$\epsilon_a = \xi \left(\frac{G m_a}{||T||_a}\right)^{1/3},$$9, circular and eccentric orbits with $\xi \sim 2$0, and profiles including NFW, Plummer, and Hernquist, the $\xi \sim 2$1-element rule is reported to be universal (Chiang et al., 30 Oct 2025).

When the rule is violated, the bound mass fraction is systematically underestimated. The reported correction fit is

$\xi \sim 2$2

where $\xi \sim 2$3 is measured from a very-high-resolution reference run (Chiang et al., 30 Oct 2025). The same work also gives a universal scatter formula for discreteness-noise-driven variance in the bound mass fraction,

$\xi \sim 2$4

and notes that as many as $\xi \sim 2$5 of first-order subhalos in the Bolshoi AMR simulation are likely to be force and/or mass unresolved (Chiang et al., 30 Oct 2025).

This subhalo-focused literature broadens the meaning of tidal adaptive softening. Instead of using the tidal tensor to choose a particle-scale smoothing directly, it uses a physically relevant tidal scale, the instantaneous tidal radius, as the target quantity that must be resolved. A plausible implication is that the term now encompasses both a local field-based prescription and an object-centered adaptive-resolution criterion.

7. Debates, limitations, and current interpretation

Three issues dominate current discussion.

The first is whether tidal adaptive softening actually solves artificial fragmentation in cosmological WDM runs. The answer depends on the variant and setup. NovA substantially suppresses spurious refinement in anisotropic structures and strongly reduces low-mass spurious halos relative to standard AMR (Hobbs et al., 2015). However, explicit tidal-tensor softening in GIZMO, when paired with late-start 3LPT initial conditions, improves force accuracy in idealized filament collapse but does not remove the low-mass WDM mass-function upturn in full cosmological boxes and can slightly increase low-mass counts near the 100-particle scale (Paun et al., 22 Jul 2025). The 2025 study interprets this to mean that, in such runs, the dominant source of spurious fragmentation lies deeper in the $\xi \sim 2$6-body discretization than what TAS alone addresses (Paun et al., 22 Jul 2025).

The second issue is the dependence of spurious-halo identification on post-processing choices. The empirical sphericity-based discriminator used in (Paun et al., 22 Jul 2025) is calibrated from proto-halo inertia tensors,

$\xi \sim 2$7

with sphericity $\xi \sim 2$8, and an empirical mean relation

$\xi \sim 2$9

with $h \propto \rho^{-1/3}$00 and $h \propto \rho^{-1/3}$01 below $h \propto \rho^{-1/3}$02 (Paun et al., 22 Jul 2025). But that same paper shows that moving from $h \propto \rho^{-1/3}$03 to $h \propto \rho^{-1/3}$04 lowers mean proto-halo sphericities by approximately $h \propto \rho^{-1/3}$05, or about $h \propto \rho^{-1/3}$06, in both CDM and WDM, so the same empirical cut no longer cleanly separates the two (Paun et al., 22 Jul 2025). This suggests that shape-based cleaning is sensitive to initial-condition choices.

The third issue concerns what should count as the most physically justified adaptive variable. Density-based softening in GADGET is conservative and improves small-scale clustering and inner-profile convergence, but it is not tidal-aware (Iannuzzi et al., 2011). Tidal-tensor softening in GIZMO is explicitly translation- and Galilean-invariant and equivalence-principle respecting, avoids the gauge dependence of potential-based rules and the pathological behavior of acceleration-based rules, and imposes negligible timestep penalties compared to neighbor-based schemes (Hopkins et al., 2022). NovA, although not expressed directly through $h \propto \rho^{-1/3}$07, encodes the same collapse geometry through an isotropy criterion on the particle distribution and enforces a conservative softening in anisotropic regions (Hobbs et al., 2015). These methods therefore occupy different points in the design space rather than constituting a single algorithm.

Taken together, the literature supports a restrained interpretation. Tidal adaptive softening is best understood as a family of tidal-aware force-resolution strategies that align numerical resolution with the geometry or strength of gravitational collapse. Its most robust successes are in controlling anisotropy-driven force errors, preserving conservative dynamics, and improving convergence for tidally stripped subhalos when the instantaneous tidal scale is explicitly resolved (Hopkins et al., 2022, Chiang et al., 30 Oct 2025). Its efficacy as a standalone cure for cosmological warm-dark-matter fragmentation is more limited and remains contingent on the broader discretization, initialization, and halo-identification pipeline (Hobbs et al., 2015, Paun et al., 22 Jul 2025).