openntn 2.0.0__tar.gz

openntn-2.0.0/CHANGELOG.md

@@ -0,0 +1,1558 @@
+# Changelog
+
+The format follows [Keep a Changelog 1.1.0](https://keepachangelog.com/en/1.1.0/).
+
+## [2.0.0] - 2026-10-26
+
+First release of OpenNTN for Sionna 2.x, which is built on PyTorch: the package
+`openntn`, installed from PyPI with `pip install openntn`, whose channel models give the
+results of 2.0.0a4 and 2.0.0rc1. Against the versions for Sionna 1.x and 0.19, its
+classes and functions take and return PyTorch tensors, its corrections of the models
+change simulation results (the first section lists them), its three example notebooks
+run with Sionna 2.x and PyTorch, and it adds a description of the models
+(`docs/model.md`), a migration guide (`docs/migration.md`) and the citation metadata
+(`CITATION.cff`). Since 2.0.0rc1 the metadata of the distribution has the development
+status "5 - Production/Stable", and the tests of two test files that failed on a GPU are
+fixed; the package is unchanged.
+
+This entry collects the changes since the versions for Sionna 1.x and 0.19, which are
+OpenNTN 1.0.0 and 0.19.0 as to their models: each item ends with the pre-releases whose
+sections below describe it in full, with its measured effect and its conditions, apart
+from the fix of the tests, which is new in this version. A change that concerns only the
+pre-releases, such as a defect that one pre-release brought and a later one fixed, is
+described only in their sections.
+
+### Changes that affect simulation results
+
+- Random numbers: every random number is drawn with PyTorch from the generators of
+  Sionna 2.x instead of with TensorFlow, so the same seed gives other realizations,
+  which also differ between the CPU and a GPU and between eager and compiled calls, and
+  single precision rounds some quantities differently; the statistics of the results do
+  not change. Affected earlier versions: the versions for Sionna 0.19 and 1.x. See
+  2.0.0rc1.
+- Shadow fading applied once: on every link the shadow fading reaches the channel only
+  as the large scale parameter SF of TR 38.901 V16.1.0, clause 7.5, step 4, with σ_SF
+  and the correlations of TR 38.811 V15.4.0 (eq. (6.6-4), Tables 6.7.2-1a to 6.7.2-8b),
+  where a second draw in the basic path loss gave about √2 times σ_SF and remained with
+  `enable_shadow_fading=False`. Affected earlier versions: the versions for Sionna 0.19
+  and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Correlation of the large scale parameters: they are correlated first across the UTs
+  and then across the parameters of each link, in the Cholesky order of TR 38.901
+  V16.1.0, clause 7.5, step 4 (TR 38.811 V15.4.0, clause 6.7.2), where the reversed
+  order left links with UTs of the same link state within the correlation distances
+  without the cross-correlations of the tables. Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- UTs at the same location: UTs at one location in the horizontal plane, in one link
+  state and batch example, share their large scale parameters, as the correlation
+  exp(-d/d_corr) of TR 38.901 V16.1.0, clause 7.5, step 4, is 1 at d = 0; its singular
+  matrix gave the UTs after them too large a variance and NaN gradients (2.0.0a1), NaN
+  parameters (the version for Sionna 1.x) or a result that depends on TensorFlow (the
+  version for Sionna 0.19). Affected earlier versions: 2.0.0a1; the versions for Sionna
+  0.19 and 1.x build the same singular correlation for such UTs. See 2.0.0a2.
+- Elevation angles midway between two reference angles: at exactly 25°, 45°, 65° and 85°
+  every table takes the higher reference angle, where the rounding of ties to the even
+  number took the lower one; TR 38.811 V15.4.0, clauses 6.6.1 and 6.6.2, takes the
+  nearest reference angle and gives no rule for an angle midway between two. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Ray offset angles: on every link ray 16 of every cluster lies at -1.1481 times the
+  intra-cluster angle spread, as in TR 38.901 V16.1.0, Table 7.5-3 (TR 38.811 V15.4.0,
+  clause 6.7.2), instead of -0.1481 times. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Cluster elimination: on every link, clusters with less than -25 dB power compared with
+  the strongest cluster get zero power, as TR 38.901 V16.1.0, clause 7.5, step 6,
+  removes them (TR 38.811 V15.4.0, clause 6.7.2), where they were kept; the shapes of
+  the outputs stay, and the other powers are not renormalized. Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Angular scaling factors: C_phi^NLOS and C_theta^NLOS follow the number of clusters at
+  the elevation angle, as in TR 38.811 V15.4.0, Tables 6.7.2-1aa and 6.7.2-1ab, where
+  each parameter file held one pair for all angles; this concerns urban LOS links at
+  10°, urban NLOS and sub-urban LOS links at 70° to 90° and sub-urban NLOS links at 60°
+  to 90°, and the files no longer have the keys `CPhiNLoS` and `CThetaNLoS`. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Random signs and ray coupling per link: the signs X_n of the cluster angles and the
+  coupling of the rays (TR 38.901 V16.1.0, clause 7.5, steps 7 and 8; TR 38.811 V15.4.0,
+  clause 6.7.2) are drawn for each link, where all UTs of a satellite shared them, which
+  correlated their angles in simulations with more than one UT per satellite. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Ray coupling in the two strongest clusters: on every link their rays are coupled
+  within the sub-clusters of TR 38.901 V16.1.0, Table 7.5-5, as clause 7.5, step 8,
+  requires (TR 38.811 V15.4.0, clause 6.7.2), where they were coupled across the whole
+  cluster; the cluster powers and delays and the mean power of a link do not change.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See
+  2.0.0a2.
+- LOS arrival direction: the zenith angle of arrival of the LOS ray is 180° minus the
+  zenith angle of departure, as in TR 38.901 V16.1.0, clause 7.5, step 1c (TR 38.811
+  V15.4.0, eq. (6.8-1b)), where the zenith angle of departure plus 180° put the arrival
+  direction 180° minus twice the elevation angle off; this concerns every LOS link
+  through the antenna pattern, the array phases and the Doppler shift at the UT.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See
+  2.0.0a2.
+- Sign of the LOS propagation phase: the LOS ray contains exp(-j2π d3D/λ0), as in
+  TR 38.901 V16.1.0, eq. (7.5-29), and TR 38.811 V15.4.0, eq. (6.8-1b), instead of
+  exp(+j2π d3D/λ0); single realizations of LOS links change, the statistics do not.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See
+  2.0.0a2.
+- LOS propagation phase in single precision: the phase 2π d3D/λ0, with d3D of TR 38.811
+  V15.4.0, eq. (6.6-3), is computed in double precision and reduced modulo 2π before the
+  cast, where single precision made it random; this concerns the first-path and the
+  total power of single LOS links in single precision, not their distributions. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1, in single
+  precision, their default precision. See 2.0.0a2.
+- Zenith angle of departure offset: mu_offset,ZOD of TR 38.901 V16.1.0, eq. (7.5-19), is
+  0, as the tables of TR 38.811 V15.4.0, clause 6.7.2, have no offset row, where NLOS
+  links took the expression of TR 38.901 Table 7.5-9 (rural macro cell), in radians
+  added to degrees; the cluster zenith angles of NLOS links at the satellite change
+  little. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+  See 2.0.0a2.
+- Sub-urban NLOS, Ka band, uplink, 80°: the mean of lg ZSD is -3.20, as in TR 38.811
+  V15.4.0, Table 6.7.2-6b, instead of -3.30, which made the median ZSD and the zenith
+  spread of the rays within each cluster at the satellite or aerial platform too small;
+  as the departure spreads of a satellite are zero in both directions (the item on the
+  departure spreads at the satellite in the uplink), the corrected value acts on the
+  uplink of aerial platforms. Affected earlier versions: the versions for Sionna 0.19
+  and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Urban NLOS, Ka band: the cluster ASA at 30°, 40° and 50° is 16.4°, 17.86° and 19.74°,
+  as in TR 38.811 V15.4.0, Table 6.7.2-4b, instead of 18.14°, 16.04° and 17.86° in both
+  directions, which made the spread of the rays within each cluster on the UT side too
+  large at 30° and too small at 40° and 50°. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Departure spreads at the satellite in the uplink: NOTE 8 of TR 38.811 V15.4.0, Tables
+  6.7.2-1a to 6.7.2-8b, by which the departure angle spreads of a satellite are zero,
+  applies in both directions, where only the downlink parameter files applied it; the
+  files hold the tables as printed, so aerial platforms take their spreads in both
+  directions, where the downlink files gave them zero. This concerns the uplink of
+  satellites and both directions of aerial platforms. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Cluster azimuth spread at the satellite: every ray leaves a satellite in the azimuth
+  of its cluster, as NOTE 8 of TR 38.811 V15.4.0, Tables 6.7.2-1a to 6.7.2-8b, makes the
+  departure angle spreads of a satellite zero, where the rays of the urban scenario left
+  at the AOD of their cluster plus c_ASD α_m (TR 38.901 V16.1.0, eq. (7.5-13)); this
+  concerns the urban links of satellites, as the other tables print c_ASD = 0, and not
+  aerial platforms. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1. See 2.0.0a2.
+- Satellite Doppler shift: the Doppler shift due to the motion of the satellite has the
+  factor R/(R+h) of TR 38.811 V15.4.0, clauses 6.7.1 and 6.9.2, without which it was
+  (R+h)/R times too large, and a geostationary satellite, which induced one, induces
+  none, as clause 5.3.2.3 states; this concerns the phase over time of every satellite
+  link with the Doppler shift enabled. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Satellites from 160 km to below 600 km: the Doppler shift due to the motion of the
+  satellite applies to them, where it applied from 600 km on; the Faraday rotation of
+  TR 38.811 V15.4.0, clause 6.8.2, given for a spaceborne BS above the ionosphere, still
+  applies from 600 km, the lowest satellite altitude of TR 38.811. Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Faraday rotation in the uplink: uplink paths to a satellite from 600 km are rotated in
+  the direction of the downlink, as seen from the earth station (ITU-R P.618-13,
+  section 4), between the scatterers and the satellite, where the field of the UT was
+  rotated before the scatterers and in the opposite direction (TR 38.811 V15.4.0,
+  eqs. (6.8-1a) and (6.8-1b), are written for the downlink); this changes the power
+  between single elements and, through scatterers, over two orthogonal elements at the
+  satellite. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1 to 2.0.0a3. See 2.0.0a4.
+- Default atmosphere: the defaults of `temperature` and `atmospheric_pressure` are those
+  of the reference atmosphere of TR 38.811 V15.4.0, clause 6.6.4 (288.15 K and
+  1013.25 hPa), where they were 273 K and 1020 hPa, and the gas loss takes
+  t = T - 273.15 in degrees Celsius, where T - 273 was used; this concerns the gas loss
+  of every link. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1. See 2.0.0a2.
+- Gas attenuation: the gas loss of every link follows Annex 2 of ITU-R P.676-10 as
+  printed (eqs. (22a) to (23d) and (25a) to (28); TR 38.811 V15.4.0, eq. (6.6-8)), with
+  the pressure of clause 6.6.4 as the dry air pressure, where it used curve fits up to
+  54 GHz, with two coefficients that no revision prints, followed no revision above
+  54 GHz and took the pressure as the total pressure. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Atmospheric parameters of `set_topology`: a value of `latitude`,
+  `atmospheric_pressure`, `temperature` or `water_vapor_density` persists until it is
+  passed again and, passed without new topology tensors, recomputes the gas and
+  scintillation losses (TR 38.811 V15.4.0, clauses 6.6.4 and 6.6.6), where every call
+  reset the values that it did not pass and only a new topology recomputed the losses;
+  the `set_topology` method of the models takes them, where it raised `TypeError`.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1. See
+  2.0.0a2.
+- Tropospheric scintillation: in the Ka band the loss is the value of TR 38.811 V15.4.0,
+  Table 6.6.6.2.1-1, at the nearest reference elevation angle, as clause 6.6.6.2.1
+  specifies above 6 GHz, where it was the fade depth of ITU-R P.618-13 exceeded 0.01 %
+  of the time instead of the 1 % of the table, computed from atmospheric and antenna
+  parameters; this concerns every link in the Ka band. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Ionospheric scintillation: in the S band, satellite links get A_IS of TR 38.811
+  V15.4.0, clause 6.6.6.1.4, only for UTs within 20° of the magnetic equator, by
+  `latitude` of `set_topology` (by default 47°, which gives 0 dB), and links to aerial
+  platforms get none, as Table 6.10.1-1 gives the scintillation as negligible for UAS,
+  where A_IS applied at every latitude and to every platform. Affected earlier versions:
+  the versions for Sionna 0.19 and 1.x and 2.0.0a1, and for aerial platforms within 20°
+  of the magnetic equator 2.0.0a2 and 2.0.0a3. See 2.0.0a2 and 2.0.0a4.
+- Antenna element pattern `"dlp"`: it includes the maximum directional gain of 5 dBi of
+  R1-1802551, Table 2, which TR 38.811 V15.4.0, clause 6.4.2, uses, where its maximum
+  was 0 dBi; this concerns every channel with `"dlp"` elements, at either end. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3. See
+  2.0.0a4.
+- Antenna arrays with `polarization="dual"` and `polarization_type="VH"`: the two
+  elements of each pair have the slant angles 0° and 90°, as in TR 38.901 V16.1.0,
+  clause 7.3.2, and R1-1802551, Table 1 (TR 38.811 V15.4.0, clause 6.4.2), where they
+  had the -45° and +45° of `"cross"`; this concerns every such array, at either end, but
+  not the power summed over a pair. Affected earlier versions: the versions for Sionna
+  0.19 and 1.x, and 2.0.0a1 to 2.0.0a3. See 2.0.0a4.
+- UT orientations of `generate_uts_topology` and `gen_single_sector_topology`: every UT
+  faces the satellite or aerial platform, as the generators intend and their docstrings
+  now state, where the UTs at negative x faced away from it (`atan` of y/x instead of
+  the angle of (x, y)); this concerns such UTs with a directional antenna element, about
+  a third of the UTs of `generate_uts_topology` with `drop_area="cell"`. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3. See
+  2.0.0a4.
+
+### Added
+
+- The argument `device`, by default `sionna.phy.config.device`, of `DenseUrban`,
+  `Urban`, `SubUrban`, their scenarios and `SystemLevelScenario`, of the antenna classes
+  and `ChannelCoefficientsGenerator`, and of nine functions of `openntn.utils`; the
+  antenna arrays of a model must be built on its device. See 2.0.0a3 and 2.0.0rc1.
+- The argument `aperture_radius` of `AntennaElement`, `PanelArray`, `Antenna` and
+  `AntennaArray`, the radius of the pattern `"aperture"` of TR 38.811 V15.4.0, clause
+  6.4.1, in wavelengths, by default 10, the radius that the pattern used before. See
+  2.0.0a4.
+- The optional argument `strongest_cluster_indices` of `Rays`. See 2.0.0a2.
+- `openntn.__version__`, the version of the installed distribution. See 2.0.0rc1.
+- The optional dependencies `test`, what the test suite of the source distribution
+  needs. See 2.0.0rc1.
+- `docs/model.md`, the description of the channel models. See 2.0.0rc1.
+- `docs/migration.md`, a guide for the users of the versions for Sionna 1.x and 0.19.
+  See 2.0.0rc1.
+- `CITATION.cff`, the citation metadata of the software. See 2.0.0rc1.
+- `example/BER_sims_results.npz` and `example/ML_based_SIMO_receiver_results.npz`, the
+  results of the full runs of `BER_sims.ipynb` and `ML_based_SIMO_receiver.ipynb`. See
+  2.0.0rc1.
+
+### Changed
+
+- The package is built on PyTorch and Sionna 2.x: its classes and functions take and
+  return PyTorch tensors of the same shapes instead of TensorFlow tensors, and its
+  objects are objects of Sionna 2.x, with the attributes `dtype` (instead of `rdtype`),
+  `cdtype` and `device`. Affected earlier versions: the versions for Sionna 0.19 and
+  1.x. See 2.0.0rc1.
+- Import name `openntn`, instead of the module `tr38811` that the installation scripts
+  of the earlier versions linked into Sionna. See 2.0.0a1.
+- Installation: `pip install openntn` installs the package from PyPI with PyTorch, NumPy
+  and Matplotlib, for Sionna 2.x (`sionna>=2.0.1,<3.0`) and Python 3.11 or later,
+  without TensorFlow, and changes no file of Sionna; the link and the line
+  `from . import tr38811` that the earlier installation scripts added to Sionna remain
+  in such an environment, as `openntn` neither needs nor removes them. Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x. See 2.0.0a1 and 2.0.0rc1.
+- The licence is declared as `MIT AND Apache-2.0`: the files derived from Sionna are
+  under the Apache License 2.0, all other files under the MIT License, and both licence
+  texts are included. See 2.0.0a1.
+- The source distribution contains `CHANGELOG.md`, and the metadata of the distribution
+  name the changelog, have keywords and the development status "5 - Production/Stable"
+  ("4 - Beta" in 2.0.0rc1) and list Louis Lagona and Tim Düe as authors. See 2.0.0rc1.
+- For users of the version for Sionna 0.19: the argument `precision`, `"single"` or
+  `"double"`, by default `sionna.phy.config.precision`, replaces the TensorFlow `dtype`,
+  and a call with `dtype` raises `TypeError`; the property `dtype` of the scenarios is
+  the real data type, and the antenna classes, `ChannelCoefficientsGenerator`,
+  `LSPGenerator`, `RaysGenerator` and `SystemLevelScenario` are objects of Sionna 2.x
+  instead of plain classes. Affected earlier version: the version for Sionna 0.19. See
+  2.0.0rc1.
+- Counts such as `num_ant`, `batch_size` and `num_ut` are Python ints instead of
+  TensorFlow integer tensors, and the antenna indices of `PanelArray` are `torch.int64`.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x. See 2.0.0rc1.
+- `Topology`, `LSP` and `Rays` are plain classes, and the abstract methods of
+  `SystemLevelScenario` are no longer enforced. Affected earlier versions: the version
+  for Sionna 1.x (the three classes) and the versions for Sionna 0.19 and 1.x (the
+  abstract methods). See 2.0.0rc1.
+- `openntn.channel_coefficients`, `openntn.lsp` and `openntn.rays` define `__all__`, so
+  a star import binds only their two classes. Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x. See 2.0.0rc1.
+- Inputs and edge cases: `set_topology` converts the topology tensors to the precision
+  and device of the model and accepts NumPy arrays, `AntennaElement.field` takes only
+  PyTorch tensors, and several functions of `openntn.utils` treat edge cases otherwise.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x. See 2.0.0rc1.
+- Graph execution: `torch.compile` takes the place of `tf.function`; `set_topology` runs
+  eagerly, outside the compiled function, so code that traced it calls it before, and
+  the call of the models compiles without graph breaks and with CUDA graphs recorded
+  once per shape of the topology, as `set_topology` writes the tensors that it derives
+  from the topology in place: they are buffers of the model, and the next `set_topology`
+  of the same shapes overwrites a tensor taken from it, so clone it to keep its values.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x. See 2.0.0a3, 2.0.0a4
+  and 2.0.0rc1.
+- A model stays on the device and in the precision of its construction: moved or cast,
+  for example with `.to()`, `.cuda()` or `.double()`, it raises `ValueError` in its next
+  `set_topology` or call. Affected earlier versions: 2.0.0a3, whose call raised it only
+  after the last `set_topology`, and 2.0.0a1 and 2.0.0a2, which had no such check. See
+  2.0.0a3 and 2.0.0a4.
+- Antenna arrays built in another precision than the model raise a `ValueError` that
+  names both precisions. Affected earlier versions: 2.0.0a1 to 2.0.0a3; the versions for
+  Sionna 0.19 and 1.x raised an `InvalidArgumentError` of TensorFlow. See 2.0.0a4.
+- The parameters of `set_topology` after `latitude` are keyword-only, so a call in the
+  earlier order, with `lwc` and `rain_rate` after `latitude`, raises `TypeError`, where
+  2.0.0a2 to 2.0.0a4 gave their values to the unchecked `atmospheric_pressure` and
+  `temperature`. Affected earlier versions: the versions for Sionna 0.19 and 1.x and
+  2.0.0a1, whose scenarios take these parameters by position in the earlier order, and
+  2.0.0a2 to 2.0.0a4, whose scenarios and models take them by position. See 2.0.0rc1.
+- `enable_pathloss` and `enable_shadow_fading` must be `True` or `False`, otherwise
+  `TypeError`, where any value was taken by its truth value. Affected earlier versions:
+  the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a4. See 2.0.0rc1.
+- `always_generate_lsp` and `doppler_enabled` must be `True` or `False`
+  (`doppler_enabled` of `set_topology` also `None`), otherwise `TypeError`, also when a
+  call passes the average street width and building height of the earlier versions by
+  position. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1
+  to 2.0.0a3. See 2.0.0a4.
+- Invalid inputs, such as an elevation angle outside 10° to 90° or a carrier frequency
+  outside 1.9 GHz to 4 GHz and 19 GHz to 40 GHz, raise `ValueError` instead of
+  `AssertionError`, also under `python -O`, which removed the checks; a `ut_array` or
+  `bs_array` of another type raises `TypeError`. Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3. See 2.0.0a4.
+- `AntennaPanel.polarization` replaces the misspelt `AntennaPanel.porlarization`,
+  without an alias. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1 to 2.0.0a3. See 2.0.0a4.
+- `Topology` takes the argument `los_phase`, which the scenarios provide; code that
+  builds a `Topology` itself has to pass it. See 2.0.0a2.
+- `set_topology` raises `ValueError` unless `bs_loc` holds one satellite (160 km to
+  35 786 km) or aerial platform (8 km to 50 km) at the same height in every batch
+  example; before, a second BS made the channel call fail, and with heights that
+  differed across the batch the 3D distance, the path loss and the switches of the
+  satellite took the height of the first batch example, as in the versions for Sionna
+  0.19 and 1.x and 2.0.0a1. See 2.0.0a2.
+- `set_topology` raises `ValueError` for a `latitude` outside -90° to 90°, the latitude
+  of the UTs relative to the magnetic equator, where any value was accepted and had no
+  effect. See 2.0.0a2.
+- `basic_pathloss` of the scenarios is the free space path loss plus the clutter loss of
+  NLOS links, without the shadow fading draw of its own that it had, as in the versions
+  for Sionna 0.19 and 1.x and 2.0.0a1. See 2.0.0a2.
+- `set_topology` raises `ValueError` for indoor UTs, as the models have neither the
+  building entry loss of TR 38.811 V15.4.0, clause 6.6.3, nor clusters for them, where
+  their channel coefficients were NaN. Affected earlier versions: 2.0.0a1, and the
+  versions for Sionna 0.19 and 1.x. See 2.0.0a2.
+- The three notebooks of `example/` run with Sionna 2.x, PyTorch and `openntn`; with
+  `OPENNTN_NOTEBOOK_MODE` unset, `BER_sims.ipynb` and `ML_based_SIMO_receiver.ipynb`
+  load their stored results, `full` computes the full runs and `reduced` runs short
+  ones. See 2.0.0rc1.
+- `BER_sims.ipynb` takes the noise variance of `ebnodb2no` without the factor of the
+  bits per symbol, so that its curves lie 3.01 dB to the left of those of the earlier
+  notebook. See 2.0.0rc1.
+- `ML_based_SIMO_receiver.ipynb` evaluates the receiver from -10 to 0 dB, where the
+  earlier notebook evaluated it from -5 to 0 dB. See 2.0.0rc1.
+- Seven changes of the notebooks: `BER_sims.ipynb` points the antenna of the satellite
+  at the UT, generates the topology at the elevation angle of the model and decodes
+  log-likelihood ratios in cell 7; `ML_based_SIMO_receiver.ipynb` evaluates the receiver
+  beside the baselines with perfect CSI and with LS channel estimation and writes its
+  weights with `numpy.savez`; `channel_response.ipynb` generates the topologies at the
+  elevation angles of the models and places the geostationary satellite at 35 786 km
+  instead of 36 000 km. See 2.0.0rc1.
+- Other changes of the notebooks: the device, single precision and the seed 42 in their
+  first code cell, in `BER_sims.ipynb` the batch size and the limits of the notebook for
+  Sionna 1.x, in `channel_response.ipynb` no Doppler sweep of the condition number and
+  other figures, and texts that describe what the code computes. See 2.0.0rc1.
+
+### Removed
+
+- `TDL` and `CDL`, with their modules; the link-level models NTN-TDL and NTN-CDL of
+  TR 38.811 V15.4.0, clause 6.9, are not implemented, and their parameter files stay in
+  `openntn/models`. See 2.0.0a2.
+- The rain and cloud attenuation of TR 38.811 V15.4.0, clause 6.6.5, which no model
+  computed: the scenario property `additional_pathloss` and
+  `utils.compute_pathloss_additional`. See 2.0.0a2.
+- The parameters `lwc` and `rain_rate` of `set_topology` and the scenario properties of
+  the same names, which had no effect. See 2.0.0a2.
+- The parameters `relative_humidity`, `diameter_earth_antenna` and `antenna_efficiency`
+  of `set_topology` and the scenario properties of the same names, which had no effect;
+  a call that passes one of the five removed parameters raises `TypeError`. See 2.0.0a2.
+- The arguments `average_street_width` and `average_building_height` of the models and
+  their scenarios and the scenario properties of the same names, which no computation
+  used. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to
+  2.0.0a3. See 2.0.0a4.
+- `compute_pathloss_basic`, `compute_pathloss_entry`, `compute_pathloss_gas` and
+  `compute_pathloss_scintilation` of `openntn.utils` as public functions; the properties
+  of the scenarios give the path loss terms. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3. See 2.0.0a4.
+- `utils.compute_stallite_doppler`, which nothing used. See 2.0.0a2.
+- `utils.gen_single_sector_topology_interferers`, which failed in every scenario. See
+  2.0.0a2.
+- The unused model data files. See 2.0.0a1.
+- `example/Example_NTN_sim.pkl`, replaced by `example/BER_sims_results.npz`. See
+  2.0.0rc1.
+
+### Fixed
+
+- The seed of Sionna alone fixes every random number of a simulation on one device, as
+  all are drawn from the generators of Sionna 2.x, where the earlier versions drew most
+  of them from the global generator of TensorFlow, which that seed does not seed.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x. See 2.0.0rc1.
+- `precision="double"` works throughout, where `gen_single_sector_topology`,
+  `set_topology` and other functions of the version for Sionna 1.x raised an error of
+  TensorFlow. Affected earlier version: the version for Sionna 1.x; the version for
+  Sionna 0.19 was not checked. See 2.0.0rc1.
+- The distribution installs only the package `openntn`, where the `setup.py` of the
+  earlier versions also listed their tests as a top-level package `test`. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x (by their `setup.py`; they were
+  not built). See 2.0.0rc1.
+- The gradients of `matrix_ut_distance_2d` and of the large scale parameters with
+  respect to the UT locations are finite, where the square root of the zero distance of
+  a UT to itself made them NaN; no simulation result changes. Affected earlier versions:
+  2.0.0a1, and the versions for Sionna 0.19 and 1.x. See 2.0.0a2.
+- Documentation of the scenario properties: `doppler_enabled` describes the flag that
+  switches the Doppler shift due to the motion of the satellite, where it had the
+  description of the antenna efficiency, and `gas_pathloss`, `scintillation_pathloss`,
+  `entry_pathloss` and `free_space_pathloss` cite TR 38.811 V15.4.0, where they cited
+  TR 38.901, clause 7.4.1; no result changes. Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x, and 2.0.0a1. See 2.0.0a2.
+- Documentation: the docstrings name `openntn`, the current paths of Sionna and the
+  arguments of the signatures; no result changes. Affected earlier versions: 2.0.0a1 to
+  2.0.0a3; the versions for Sionna 0.19 and 1.x have the same texts, apart from the
+  module paths. See 2.0.0a4.
+- Documentation: corrected docstrings, for example that of
+  `SystemLevelScenario.los_zoa`, which gives the zenith angle of arrival of the LOS path
+  as 180° minus its zenith angle of departure, where it said "considered 90 degrees"; no
+  result changes. Affected earlier versions: 2.0.0a1 to 2.0.0a4; the versions for Sionna
+  0.19 and 1.x have the same texts, apart from the arguments `precision` and `device`.
+  See 2.0.0rc1.
+- The messages of the checks of the panel spacings of `PanelArray` spell "Panel" where
+  they said "Pannel". Affected earlier versions: the versions for Sionna 0.19 and 1.x,
+  and 2.0.0a1 to 2.0.0a4. See 2.0.0rc1.
+- The tests in `tests/topology_orientations_test.py` and the test `test_dlp_pattern` in
+  `tests/antenna_pattern_test.py` failed on a GPU, where Sionna's default device is
+  `cuda:0`: they converted tensors on the GPU to NumPy arrays without moving them to
+  the CPU. The package is not affected. Affected earlier versions: 2.0.0a4 and
+  2.0.0rc1, whose source distributions contain these tests.
+
+## [2.0.0rc1] - 2026-10-21
+
+Release candidate of OpenNTN 2.0.0 for Sionna 2.x, whose channel models give the results
+of 2.0.0a4: it adds a description of the models (`docs/model.md`), a migration guide for
+the users of the versions for Sionna 1.x and 0.19 (`docs/migration.md`) and the citation
+metadata (`CITATION.cff`), ports the three example notebooks to Sionna 2.x and PyTorch,
+takes the parameters of `set_topology` after `latitude` by keyword only, and accepts
+only `True` or `False` for `enable_pathloss` and `enable_shadow_fading`.
+
+The entries of this section also complete the sections of 2.0.0a1 to 2.0.0a4: an entry
+that names a pre-release, such as "since 2.0.0a1", describes a change that this
+pre-release brought and its section does not list.
+
+### Changes that affect simulation results
+
+- Random numbers: since 2.0.0a1 the package draws every random number with PyTorch,
+  from the generators of Sionna 2.x, instead of with TensorFlow, so the same seed gives
+  other realizations of every random quantity: the UT drops of the topology generators,
+  the LOS states, the large scale parameters, the cluster delays, powers and angles, the
+  ray coupling, the cross-polarization ratios, the initial phases and the direction of
+  the orbit. The move to PyTorch changed neither the number nor the order nor the
+  distributions of the draws; the corrections listed in the sections of the
+  pre-releases change some of them. The realizations also depend on where they are
+  drawn: the same seed gives other realizations on the CPU than on a GPU, and a call
+  compiled with `torch.compile` draws other random numbers than an eager call, unless
+  `torch._inductor.config.fallback_random` is set. In single precision, some
+  intermediate quantities are rounded differently, because they are now computed in
+  double precision before the cast or by the kernels of PyTorch, for example the cosine
+  of the elevation angle in the satellite speed, the Faraday rotation angle, the
+  constant of the free space path loss, the Bessel function of the pattern `"aperture"`
+  and the antenna positions, so results in single precision can differ within rounding
+  (the size of these differences was not measured). None of these differences changes
+  the statistics of the results. With the versions for Sionna 0.19 and 1.x, the seed of
+  Sionna alone did not fix the random numbers (see "Fixed"). Affected earlier versions:
+  the versions for Sionna 0.19 and 1.x.
+
+### Added
+
+- `example/BER_sims_results.npz` and `example/ML_based_SIMO_receiver_results.npz`, the
+  results of the full runs of `BER_sims.ipynb` and `ML_based_SIMO_receiver.ipynb` with
+  their metadata (the seed, the versions, the date, the device, the precision, the
+  parameters and the run times), as NumPy arrays read with
+  `numpy.load(..., allow_pickle=False)`.
+- The argument `device` of `AntennaElement`, `AntennaPanel`, `PanelArray`, `Antenna`,
+  `AntennaArray` and `ChannelCoefficientsGenerator`, and of the functions
+  `subcarrier_frequencies`, `time_frequency_vector`, `sample_bernoulli`,
+  `drop_uts_in_sector`, `set_3gpp_scenario_parameters`, `generate_uts_topology`,
+  `gen_single_sector_topology`, `exp_corr_mat` and `one_ring_corr_mat` of
+  `openntn.utils`: the device on which they create their tensors, by default
+  `sionna.phy.config.device`. The antenna arrays of a model must be built on the device
+  of the model (see 2.0.0a3). Since 2.0.0a1; the versions for Sionna 0.19 and 1.x have
+  no such argument, as TensorFlow places the tensors itself.
+- `openntn.__version__`, the version of the installed distribution, read from its
+  metadata; in a source tree that is not installed it is `"0+unknown"`. Since 2.0.0a1;
+  the versions for Sionna 0.19 and 1.x have no `__version__`.
+- The optional dependencies `test`, installed with `pip install openntn[test]`: what
+  the test suite of the source distribution needs, pytest since 2.0.0a1 and mpmath
+  since 2.0.0a2.
+- `docs/model.md`, the scope, the conventions and the limits of the channel models: what
+  they implement of TR 38.811 and TR 38.901, where they take a side and what they leave
+  out, with their calibration against TR 38.821 V16.2.0 and their speed on a GPU.
+- `docs/migration.md`, a guide for the users of the versions for Sionna 1.x and 0.19:
+  the package and its import, the installation, PyTorch tensors, precision, device and
+  random numbers, the removed and renamed names, the arguments by keyword and the checks
+  of the inputs, the results that change and why, and the example notebooks.
+- `CITATION.cff`, the citation metadata of the software in the Citation File Format
+  1.2.0, with the paper that presents the framework among its references.
+
+### Changed
+
+- The package is built on PyTorch and Sionna 2.x: since 2.0.0a1 its classes and
+  functions take and return PyTorch tensors (`torch.Tensor`) instead of TensorFlow
+  tensors, with the same shapes, for example the topology tensors of `set_topology` and
+  the path coefficients and delays of a call of `DenseUrban`, `Urban` and `SubUrban`,
+  so the code around them has to use PyTorch and Sionna 2.x as well. Its objects are
+  objects of Sionna 2.x: the real data type is the attribute `dtype` instead of
+  `rdtype`, the complex one is `cdtype` as before, and the attribute `device` is new;
+  an invalid precision raises the error of Sionna 2.x ("Invalid precision") instead of
+  "'precision' must be 'single' or 'double'". Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x, which are built on TensorFlow (for the version for Sionna
+  0.19, see also the entry for its users below).
+- Graph execution: `torch.compile` takes the place of TensorFlow's `tf.function`.
+  `set_topology` runs eagerly, outside the compiled function, and the call of
+  `DenseUrban`, `Urban` and `SubUrban` can be compiled, without graph breaks since
+  2.0.0a3 and with CUDA graphs across topologies of the same shapes since 2.0.0a4 (see
+  "Fixed" in these versions). The versions for Sionna 0.19 and 1.x could run
+  `set_topology` inside a function traced by `tf.function`, as the earlier
+  `BER_sims.ipynb` did; such code has to call `set_topology` before the compiled
+  function. Affected earlier versions: the versions for Sionna 0.19 and 1.x.
+- For users of the version for Sionna 0.19: the classes take the argument `precision`,
+  `"single"` or `"double"`, instead of a TensorFlow `dtype`, as in the version for
+  Sionna 1.x: `AntennaElement`, `AntennaPanel`, `PanelArray`, `Antenna`, `AntennaArray`,
+  `ChannelCoefficientsGenerator`, `DenseUrban`, `Urban`, `SubUrban`, their scenarios and
+  `SystemLevelScenario`, and so do the functions `subcarrier_frequencies`,
+  `time_frequency_vector`, `drop_uts_in_sector`, `set_3gpp_scenario_parameters`,
+  `generate_uts_topology`, `gen_single_sector_topology`, `exp_corr_mat` and
+  `one_ring_corr_mat` of `openntn.utils`; a call with `dtype` raises `TypeError`.
+  Without `precision`, they take `sionna.phy.config.precision`, which sets the
+  precision of every object at once. The property `dtype` of the scenarios is the real
+  data type (`torch.float32` or `torch.float64`) instead of the complex TensorFlow
+  dtype, which is now `cdtype`. `AntennaElement`, `AntennaPanel`, `PanelArray` (and with
+  it `Antenna` and `AntennaArray`), `ChannelCoefficientsGenerator`, `LSPGenerator`,
+  `RaysGenerator` and `SystemLevelScenario` are objects of Sionna 2.x, PyTorch modules
+  with a precision and a device, instead of plain classes (an abstract base class in
+  the case of `SystemLevelScenario`, see the entry on its abstract methods). Since
+  2.0.0a1. Affected earlier version: the version for Sionna 0.19; the version for
+  Sionna 1.x took `precision` already.
+- The parameters of `set_topology` after `latitude` are taken by keyword only:
+  `atmospheric_pressure`, `temperature` and `water_vapor_density` of `DenseUrban`,
+  `Urban`, `SubUrban` and their scenarios, and `doppler_enabled` of the scenarios. A
+  call that passes them by position raises `TypeError`; `latitude` stays the eighth
+  parameter, its position in the scenarios of all earlier versions. In the versions for
+  Sionna 0.19 and 1.x and in 2.0.0a1, the `set_topology` method of the scenarios took
+  `lwc` and `rain_rate` by position after `latitude`. Since their removal in 2.0.0a2, a
+  call written for that order gave their values to `atmospheric_pressure` and
+  `temperature`, which are not checked: with a liquid water content of 0.5 g/m³ and a
+  rain rate of 100 mm/h, for example, the gas loss was computed for 0.5 hPa and 100 K,
+  2.1180 dB instead of 0.5156 dB at 20 GHz and 30° (dense urban, downlink, 600 km), and
+  a rain rate of 10 mm/h made `set_topology` fail with an `OverflowError` in the gas
+  loss (computed with the code of 2.0.0a4). Such a call now raises `TypeError`, and so
+  does a call that passes the atmospheric parameters by position in the order of
+  2.0.0a2 to 2.0.0a4, which has to name them. Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x and 2.0.0a1, whose scenarios take these parameters by position
+  in the earlier order, and 2.0.0a2 to 2.0.0a4, whose scenarios and models take them by
+  position.
+- `enable_pathloss` and `enable_shadow_fading` of `DenseUrban`, `Urban`, `SubUrban` and
+  their scenarios must be `True` or `False`; other values raise `TypeError`. Before, any
+  value was accepted and taken by its truth value. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a4, which accept any value.
+- Counts are Python ints instead of TensorFlow integer tensors: the properties
+  `num_rows` and `num_cols` of `AntennaPanel` and `PanelArray`, `num_rows_per_panel`,
+  `num_cols_per_panel`, `num_panels`, `num_panels_ant` and `num_ant` of `PanelArray`
+  (and of `Antenna` and `AntennaArray`), and `batch_size`, `num_ut`, `num_bs`,
+  `num_clusters_los`, `num_clusters_nlos` and `num_clusters_max` of the scenarios. The
+  antenna indices `ant_ind_pol1` and `ant_ind_pol2` of `PanelArray` are `torch.int64`
+  instead of `tf.int32`. Since 2.0.0a1. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x.
+- `Topology`, `LSP` and `Rays` are plain classes, without the attributes of Sionna's
+  `Object`, from which the version for Sionna 1.x derived them. The abstract methods of
+  `SystemLevelScenario` are no longer enforced, as it derives from Sionna's `Object`, a
+  PyTorch module, instead of an abstract base class: a subclass that lacks one is built
+  without an error and fails when the method is used. Since 2.0.0a1. Affected earlier
+  versions: the version for Sionna 1.x (the three classes) and the versions for Sionna
+  0.19 and 1.x (the abstract methods).
+- The modules `openntn.channel_coefficients`, `openntn.lsp` and `openntn.rays` define
+  `__all__`, so a star import binds only their two classes, for example `LSP` and
+  `LSPGenerator` from `openntn.lsp`; it also bound the names that the module imported,
+  such as `tf`. Since 2.0.0a1. Affected earlier versions: the versions for Sionna 0.19
+  and 1.x.
+- Accepted inputs and edge cases: `set_topology` converts the topology tensors to the
+  precision and device of the model and accepts NumPy arrays; `AntennaElement.field`
+  takes PyTorch tensors of any floating precision and device, but no Python numbers or
+  NumPy arrays; `deg_2_rad`, `rad_2_deg` and `wrap_angle_0_360` accept Python numbers
+  and tensors of any precision; `drop_uts_in_sector` and `gen_single_sector_topology`
+  accept integer heights; `sample_bernoulli` no longer takes a TensorFlow dtype as its
+  precision; `exp_corr_mat` raises `ValueError` for a coefficient of magnitude 1 or
+  more instead of an error of TensorFlow, and no longer rejects a NaN coefficient;
+  `time_lag_discrete_time_channel` returns Python ints and computes its ceiling in
+  double precision (one sample more where the product lies just above an integer);
+  with `normalize=True`, `cir_to_ofdm_channel` and `cir_to_time_channel` leave a link
+  whose energy is NaN unnormalized, where the division made every coefficient of such
+  a link NaN, and the first takes its frequencies only as a PyTorch tensor, cast to the
+  precision and device of the delays; `time_to_ofdm_channel` returns a response with
+  more taps than the FFT size instead of an error. Since 2.0.0a1. Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x.
+- Installation: `pip install openntn` installs the package with its dependencies and
+  changes no file of Sionna (since 2.0.0a1). The installation scripts of the versions
+  for Sionna 0.19 and 1.x linked their package into Sionna, as `sionna.channel.tr38811`
+  or `sionna.phy.channel.tr38811`, and added the line `from . import tr38811` to the
+  `__init__.py` of Sionna's channel package; `openntn` neither needs nor removes this
+  link and this line, which remain in an environment with such an installation (the
+  line until Sionna is reinstalled). The distribution declares PyTorch, NumPy and
+  Matplotlib as dependencies besides Sionna 2.x, with the lower bounds of Sionna 2.0.1;
+  it needs neither TensorFlow nor `importlib_resources`, which the earlier versions
+  imported without declaring it. Affected earlier versions: the versions for Sionna
+  0.19 and 1.x.
+- The source distribution contains `CHANGELOG.md`. The metadata of the distribution
+  names the changelog among the project URLs, has keywords and the development status
+  "4 - Beta", and lists Louis Lagona and Tim Düe as authors, in this order.
+- The three notebooks of `example/`, `channel_response.ipynb`, `BER_sims.ipynb` and
+  `ML_based_SIMO_receiver.ipynb`, run with Sionna 2.x and PyTorch, with the channel
+  models, the antennas and the topology generator of `openntn`, instead of TensorFlow
+  and `sionna.phy.channel.tr38811`. The environment variable `OPENNTN_NOTEBOOK_MODE`,
+  read in their first code cell, sets what they compute: unset, `BER_sims.ipynb` and
+  `ML_based_SIMO_receiver.ipynb` load their stored results and `channel_response.ipynb`
+  runs with its own batch sizes; `full`, the full runs, which for
+  `ML_based_SIMO_receiver.ipynb` include the training of the receiver and need a GPU;
+  `reduced`, short runs with small batches that write no file.
+- `BER_sims.ipynb` computes the noise variance with `ebnodb2no` for the Eb/N0 per
+  information bit, without the factor of the bits per symbol by which the earlier
+  notebook multiplied it: the noise variance at an Eb/N0 of x dB is that of the earlier
+  notebook at x + 3.01 dB, so that by this change its curves lie 3.01 dB to the left of
+  those of the earlier notebook (its BER at x is the earlier BER at x + 3.01 dB).
+- `ML_based_SIMO_receiver.ipynb` trains the receiver on Eb/N0 drawn from [-5, 0.5) dB,
+  as before, and evaluates it from -10 to 0 dB in steps of 0.5 dB; the earlier notebook
+  evaluated it from -5 to 0 dB.
+- Seven changes of the notebooks. `BER_sims.ipynb` points the antenna of the satellite
+  at the UT instead of the nadir: with the topology at 50°, the nadir would put the UT
+  40° off the boresight of the pattern `"aperture"`, at -44.64 dB (in the earlier
+  notebook, whose UTs lay at 80°, 10° off, at -30.54 dB); the channel is normalized, so
+  this changes the BER only through the polarization. It generates the topology at the
+  elevation angle of the model, 50°, where the earlier notebook took the default of the
+  generator, 80°, and it decodes log-likelihood ratios in the single transmission of
+  its cell 7, where the earlier notebook fed the decoder hard decisions and printed a
+  coded BER of 0.478 (0.0776 with the version for Sionna 0.19); its BER simulations stay
+  uncoded, with hard decisions. `ML_based_SIMO_receiver.ipynb` evaluates the receiver
+  beside the two baselines with perfect CSI and with LS channel estimation, and writes
+  the weights once, after the training, as float32 arrays with `numpy.savez`, where the
+  earlier notebooks pickled them. `channel_response.ipynb` generates the topologies at
+  the elevation angles of the models, where the earlier notebook took 90° or the
+  default of the generator, 80°, beside models at other angles, and places the
+  geostationary satellite at 35 786 km, the altitude of TR 38.811 V15.4.0, Table 4.5-1,
+  instead of 36 000 km, which the models reject since 2.0.0a2.
+- Other changes of the notebooks: their first code cell sets the device, single
+  precision and the seed 42, so that runs on one device repeat their results, apart from
+  a training of the receiver on a GPU (the earlier notebooks set no seed);
+  `BER_sims.ipynb` simulates each point with a batch size of 16, at most 1000
+  iterations and at most 1000 block errors, as the notebook for Sionna 1.x did, where
+  the notebook for Sionna 0.19 took 128, 1 000 000 and 10 000; `channel_response.ipynb`
+  no longer sweeps the condition number over the Doppler shift (it took one time
+  sample, at t = 0, where the Doppler shift has no effect, so its figure showed random
+  variation only), leaves the paths without power out of its figures, shows the
+  condition numbers as empirical distribution functions with their medians and plots
+  every time sample; and the texts of the three notebooks describe what the code
+  computes and what its results show.
+
+### Removed
+
+- `example/Example_NTN_sim.pkl`, the pickled results of the earlier `BER_sims.ipynb`,
+  replaced by `example/BER_sims_results.npz`.
+
+### Fixed
+
+- The seed of Sionna alone now fixes every random number of a simulation on one device:
+  since 2.0.0a1 the package draws all of them from the generators of Sionna 2.x, which
+  `sionna.phy.config.seed` seeds (see "Changes that affect simulation results"). The
+  seed of Sionna 1.x (`sionna.phy.config.seed`) and that of Sionna 0.19
+  (`sionna.config.seed`) seed Sionna's own generators, but not TensorFlow's global
+  generator. The version for Sionna 1.x drew only the large scale parameters and its
+  Bernoulli draws, such as the LOS states, from Sionna's generator, the version for
+  Sionna 0.19 only the LOS states, and both drew every other random number from
+  TensorFlow's global generator: the UT drops, the cluster delays, powers, signs and
+  angles, the ray coupling, the cross-polarization ratios, the initial phases, the
+  direction of the orbit and the shadow fading of the basic path loss. Their results
+  therefore changed from run to run under the same seed of Sionna; a run of these
+  versions repeats its random numbers only when `tf.random.set_seed` is set as well
+  (read from the code of these versions and of Sionna 0.19.2, 1.0.1 and 1.2.1).
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x.
+- Double precision: `precision="double"` works throughout (since 2.0.0a1), as the
+  reference outputs of the test suite, computed in double precision, show. In the
+  version for Sionna 1.x several functions combined float32 tensors with float64 ones,
+  among them `gen_single_sector_topology` and `set_topology`, which therefore raised an
+  error of TensorFlow in double precision (computed with Sionna 1.2.1 and TensorFlow
+  2.19.0), so that simulations ran in single precision, its default, only. Affected
+  earlier version: the version for Sionna 1.x; the version for Sionna 0.19, which takes
+  a TensorFlow dtype, was not checked.
+- The distribution installs only the package `openntn` (since 2.0.0a1); the tests are
+  part of the source distribution only. The `setup.py` of the versions for Sionna 0.19
+  and 1.x took its packages from `find_packages()`, which also lists the folder of their
+  tests as a top-level package `test`, a name that Python's standard library uses for
+  its own tests. Affected earlier versions: the versions for Sionna 0.19 and 1.x (by
+  their `setup.py`; they were not built).
+- Documentation: the docstring of `SystemLevelChannel` describes its argument
+  `precision`, which the model does not use, as in Sionna, since it takes the precision
+  and the device of its scenario; the docstrings of the argument `precision` of
+  `DenseUrban`, `Urban`, `SubUrban`, their scenarios and `SystemLevelScenario` say that
+  the antenna arrays must be built in that precision; the docstring of
+  `SystemLevelScenario.los_zoa` gives the zenith angle of arrival of the LOS path, 180°
+  minus its zenith angle of departure, where it said "considered 90 degrees"; those of
+  `los_probability` of the scenarios give its shape [batch size, number of BSs, number
+  of UTs]; that of `drop_uts_in_sector` describes its drop, within ± `isd`/2 in x and in
+  y around one point at the elevation angle of the BS, where it described the sector
+  drop of Sionna and `isd` as the distance between two adjacent BSs, and its arguments
+  `elevation_angle` and `device`; the module `openntn` no longer calls itself a
+  sub-package of Sionna; the docstrings of `PanelArray`, `time_to_ofdm_channel`,
+  `drop_uts_in_sector`, `relocate_uts`, `generate_uts_topology` and
+  `gen_single_sector_topology` no longer refer to figures of Sionna's documentation,
+  which the package does not have; that of `relocate_uts` describes the indexing of the
+  sectors in words; the entries `panel_vertical_spacing` and `panel_horizontal_spacing`
+  of `PanelArray` compare the spacings with the height and the width of a panel, as its
+  checks do, where they had height and width swapped; and the summary of
+  `gen_single_sector_topology` places its BS above the origin and its UTs around one
+  point, where it said that the UTs are dropped uniformly in a cell sector. No result
+  changes. Affected earlier versions: 2.0.0a1 to 2.0.0a4; the versions for Sionna 0.19
+  and 1.x have the same texts, apart from the argument `precision`, which the version
+  for Sionna 0.19 does not take (it takes a TensorFlow `dtype`), and the argument
+  `device`, which neither takes.
+- The messages of the checks of the panel spacings of `PanelArray` spell "Panel" where
+  they said "Pannel". Affected earlier versions: the versions for Sionna 0.19 and 1.x,
+  and 2.0.0a1 to 2.0.0a4.
+
+## [2.0.0a4] - 2026-10-19
+
+Fourth pre-release of OpenNTN for Sionna 2.x: its corrections change simulation results
+(the first section names them), and the interface changes, with a renamed property,
+private path loss functions, removed arguments and checks of the inputs that raise
+`ValueError` or `TypeError`.
+
+### Changes that affect simulation results
+
+- Ionospheric scintillation of aerial platforms: links to an aerial platform (8 km to
+  50 km) below 6 GHz no longer get the ionospheric scintillation loss of TR 38.811
+  V15.4.0, clause 6.6.6.1.4; its Table 6.10.1-1 gives the scintillation as negligible
+  for the deployment of UAS including HAPS below 6 GHz, whose links stay below the
+  ionosphere. Within 20° of the magnetic equator such links got 2.2 dB at 2 GHz
+  (1.93 dB at 2.185 GHz); at higher latitudes, the default of 47° included, the loss
+  was already 0. Satellites keep the loss. Affected earlier versions: 2.0.0a2 and
+  2.0.0a3; the versions for Sionna 0.19 and 1.x and 2.0.0a1 gave the loss at every
+  latitude, also to platforms at the heights of aerial platforms, which they accepted
+  without a check.
+- Antenna element pattern `"dlp"`: it now includes the maximum directional gain of
+  5 dBi that R1-1802551, Table 2, gives for the element of the co-phased dual linearly
+  polarized UE array, which TR 38.811 V15.4.0, clause 6.4.2, uses. Its maximum was
+  0 dBi, so every path of a channel with `"dlp"` elements was 5 dB too weak for each end
+  that uses them, 10 dB with `"dlp"` at both ends (computed with this version); the
+  element `"38.901"` already included its 8 dBi. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3.
+- The UTs of `generate_uts_topology` and `gen_single_sector_topology` now face the
+  satellite or aerial platform, as the generators intend: the bearing of a UT at
+  negative x was off by 180° (`atan` of y/x instead of the angle of (x, y)), so its
+  boresight pointed away from the platform, 180° minus twice the elevation angle off
+  the line of sight. This concerns UTs with a directional antenna element: with the
+  element `"38.901"` their gain toward the platform was -22 dBi instead of +8 dBi at
+  elevation angles up to 30° and -3.5 dBi at 60°. `generate_uts_topology` with
+  `drop_area="cell"` places about a third of the UTs at negative x;
+  `gen_single_sector_topology` does so on average for at most about 0.6 % of the UTs at
+  600 km and elevation angles up to 80°, but for up to a fifth at 20 km and a third at
+  8 km at 80°, where the loss was 1 to 2 dB on average (computed with this version). At
+  90° nothing changes. Their docstrings now state the orientations that they set: the
+  BS facing the nadir and each UT facing the BS; they said that the UT orientations
+  were random and that the BS faced the centre of the sector. Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3.
+- Dual-polarized antenna arrays with `polarization_type="VH"`: the two elements of each
+  pair now have the slant angles 0° and 90°, a vertically and a horizontally polarized
+  element, as in TR 38.901 V16.1.0, clause 7.3.2 (polarization model 2) and the "X-pol
+  (0/+90 degree)" configuration of its clause 7.8.2, and as R1-1802551, Table 1, gives
+  for the UE arrays of TR 38.811 V15.4.0, clause 6.4.2. They had the slant angles -45°
+  and +45°, those of `polarization_type="cross"`, because the constructor tested the
+  wrong argument. This concerns every `PanelArray`, `Antenna` and `AntennaArray` built
+  with `polarization="dual"` and `polarization_type="VH"`, at the satellite, the aerial
+  platform or the UT, in all scenarios, bands, directions and link states. The power of
+  each element of such a pair changes; the power summed over the two elements of a pair
+  does not. In the characterization outputs of the test suite, a satellite array of
+  four such pairs and a vertically polarized UT element, the power per antenna pair
+  changed by up to 32 dB, the power summed over the satellite's elements by less than
+  1e-14 relative (computed with this version). Affected earlier versions: the versions
+  for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3.
+- Faraday rotation in the uplink: the polarization of uplink paths to a satellite from
+  600 km is now rotated in the same direction as on the downlink, as seen from the earth
+  station, as ITU-R P.618-13, section 4, states, and between the scatterers near the UT
+  and the satellite, as on the downlink. Before, the uplink rotated the UT's field
+  before the scatterers and in the opposite direction (TR 38.811 V15.4.0, eqs. (6.8-1a)
+  and (6.8-1b), are written for the downlink). The rotation is 108°/f² (f in GHz;
+  6.75° to 29.9° in the S band, 0.07° to 0.30° in the Ka band). This changes the power
+  between single elements (by tens of dB, without a bound near a polarization null on a
+  LOS path) and, for the paths through scatterers, the power that the satellite
+  collects over two orthogonal elements, which no longer depends on the rotation (it
+  did by up to 13 dB for single paths); the power over two orthogonal elements at
+  either end of a LOS path does not change (computed with this version). Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3.
+
+### Added
+
+- The argument `aperture_radius` of `AntennaElement`, `PanelArray`, `Antenna` and
+  `AntennaArray`: the radius a of the circular aperture of the element pattern
+  `"aperture"` of TR 38.811 V15.4.0, clause 6.4.1, in multiples of the wavelength, as
+  the element spacings of `PanelArray`. For an aperture of diameter D at the carrier
+  frequency f, a = D f/(2c), with c the speed of light; the 2 m of TR 38.821 V16.2.0,
+  Table 6.1.1.1-1, at 2 GHz give a = 6.67. It defaults to 10, the radius of Figure
+  6.4.1-1 of TR 38.811, which the pattern always used before, so results do not change;
+  the other patterns do not use it. A value that is a bool or not a real number raises
+  `TypeError`, one that is not positive and finite `ValueError`.
+
+### Changed
+
+- A model moved to another device or cast to another precision, for example with
+  `.to()`, `.cuda()` or `.double()`, now raises `ValueError` in every order:
+  `set_topology` checks the device and the precision of its tensors first, as the call
+  already did after the last `set_topology`. Before, a model moved or cast before
+  `set_topology` or between two topologies failed with an error from PyTorch in
+  `set_topology` or in the call, such as "Tensor on device meta is not on the expected
+  device cpu!" or "expected scalar type ComplexFloat but found Double". Affected
+  earlier version: 2.0.0a3; 2.0.0a1 and 2.0.0a2 had no such check.
+- `AntennaPanel.polarization` replaces the misspelt `AntennaPanel.porlarization`,
+  which is removed without an alias. The property returns the polarization of the
+  panel, `"single"` or `"dual"`, under the name of the same property of Sionna's
+  `AntennaPanel` and of `PanelArray`. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3, which have only `porlarization`.
+- `always_generate_lsp` of `DenseUrban`, `Urban` and `SubUrban`, and `doppler_enabled`
+  of these models, of their scenarios, of the `set_topology` method of the scenarios
+  and of `Topology`, must be `True` or `False` (`None` in `set_topology`, which keeps
+  the flag); other values raise `TypeError`. Before, any value was accepted and taken
+  by its truth value. The arguments `average_street_width` and
+  `average_building_height`, which came before them, are removed (see "Removed"), so a
+  call that passes the arguments of the earlier versions by position would otherwise
+  take the street width as one of the flags; it now raises `TypeError`. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3,
+  which accept any value.
+- Invalid inputs raise `ValueError` instead of `AssertionError`, also under
+  `python -O`, which removed these checks, so that the invalid inputs went on
+  unchecked: an elevation angle outside 10° to 90°, a carrier frequency outside 1.9 GHz
+  to 4 GHz and 19 GHz to 40 GHz, a `direction` other than `"uplink"` and `"downlink"`,
+  and arrays whose carrier frequency differs from that of the model (`DenseUrban`,
+  `Urban`, `SubUrban` and their scenarios); a first `set_topology` that leaves out one
+  of the topology tensors; the element pattern of `AntennaElement`; the polarization of
+  `AntennaPanel`, `PanelArray`, `Antenna` and `AntennaArray`, the polarization type of
+  the last three, and panel spacings of `PanelArray` that do not exceed the panel size;
+  the properties `ant_pol2`, `ant_ind_pol2` and `ant_pos_pol2` of a single-polarized
+  array; the scenario of `set_3gpp_scenario_parameters` and
+  `gen_single_sector_topology`; and the drop area of `generate_uts_topology`. A
+  `ut_array` or `bs_array` that is not a `PanelArray`, `Antenna` or `AntennaArray`
+  raises `TypeError` instead of `AssertionError`, as an argument of the wrong type. The
+  conditions and the messages are those of the earlier checks. Under `python -O`, a
+  model at an elevation angle of 5° or at 6 GHz, for example, was built without an
+  error (the code before this change). Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3, which check these inputs with `assert`.
+- Antenna arrays built in another precision than the model raise a `ValueError` that
+  names both precisions. Before, the check of the carrier frequencies failed with its
+  message on the frequencies or, at a frequency whose wavelength both precisions hold
+  exactly, such as 2 398 339 664 Hz (0.125 m), the model was built and its call failed
+  with an error from PyTorch. Affected earlier versions: 2.0.0a1 to 2.0.0a3; the
+  versions for Sionna 0.19 and 1.x raised an `InvalidArgumentError` of TensorFlow on
+  the data types (computed with TensorFlow 2.18.1 and 2.19.0).
+
+### Removed
+
+- The path loss functions `compute_pathloss_basic`, `compute_pathloss_entry`,
+  `compute_pathloss_gas` and `compute_pathloss_scintilation` of `openntn.utils` are no
+  longer public. They take a scenario of the package, such as a `DenseUrbanScenario`,
+  and store their results in its tensors, so only the scenarios use them, in
+  `set_topology`. The path loss terms stay available as the properties
+  `basic_pathloss`, `gas_pathloss`, `scintillation_pathloss`, `entry_pathloss` and
+  `free_space_pathloss` of the scenarios. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1 to 2.0.0a3, whose `utils` module has them.
+- The arguments `average_street_width` and `average_building_height` of `DenseUrban`,
+  `Urban`, `SubUrban` and their scenarios, and the scenario properties of the same
+  names: no computation used them. TR 38.811 has no such parameters; the models took
+  them over from the rural macro model (RMa) of TR 38.901 in Sionna, whose path loss
+  uses them. A call that passes one of them by name raises `TypeError`; for a call by
+  position, see "Changed". Affected earlier versions: the versions for Sionna 0.19 and
+  1.x, and 2.0.0a1 to 2.0.0a3, which accept them.
+
+### Fixed
+
+- With `torch.compile(mode="reduce-overhead")`, which captures the call of
+  `DenseUrban`, `Urban` and `SubUrban` as CUDA graphs, a new topology of the same
+  shapes no longer makes the next call record its graphs again: `set_topology` now
+  writes the tensors that it derives from the topology (the means and standard
+  deviations of the large scale parameters, the ZOD offset, the path loss terms, the
+  cluster mask and the large scale parameters drawn with the topology) into the
+  existing tensors when their shapes stay the same and no gradients are involved, as it
+  already did for the topology itself. Before, it replaced up to 13 of the inputs of
+  the captured graphs, and PyTorch recorded and kept one more set of graphs after every
+  new topology. Measured on an RTX 4090 with torch 2.14.1 (DenseUrban at 2 GHz, 8 and
+  64 topologies of the same shapes): one recording in all instead of one per topology,
+  and the first call after `set_topology` as fast as the next calls, 0.44 to 1.93 ms
+  (computed with this version). A topology with another batch size still records new
+  graphs, which PyTorch keeps. The results are unchanged. The tensors derived from the
+  topology are now buffers of the model and appear in its `state_dict()`. A tensor
+  taken from the model after `set_topology`, such as the path loss terms or the means
+  of the large scale parameters, is now overwritten by the next `set_topology` of the
+  same shapes, as the topology tensors already were; clone it to keep its values.
+  Affected earlier versions: 2.0.0a1 to 2.0.0a3; the versions for Sionna 0.19 and 1.x
+  are built on TensorFlow and are not affected.
+- Documentation: the docstrings referred to the classes of this package in the module
+  `sionna.phy.channel.tr38811`, which does not exist with the package `openntn`, and to
+  `sionna.ofdm.ResourceGrid` instead of `sionna.phy.ofdm.ResourceGrid`. The examples of
+  `DenseUrban`, `Urban`, `SubUrban` and `gen_single_sector_topology` built the model
+  without the required `elevation_angle`, the last one with an argument `o2i_model` that
+  no model takes, and the signatures at the top of the docstrings of the models left out
+  `elevation_angle` and `doppler_enabled`. The docstrings of the models and their
+  scenarios named the antenna arrays `rx_array` and `tx_array` instead of `ut_array` and
+  `bs_array`, those of the scenarios described `always_generate_lsp`, which only the
+  models take, and those of the models named the first input of the call
+  `num_time_steps` instead of `num_time_samples`. The argument `doppler_enabled` of the
+  models, their scenarios, `SystemLevelScenario` and its `set_topology` had no
+  description. `set_3gpp_scenario_parameters` and `gen_single_sector_topology`
+  described a parameter `min_bs_ut_dist` that they do not take, `exp_corr_mat` a
+  parameter `dtype` instead of `precision` and `device`, and `time_to_ofdm_channel` its
+  parameter `rg` as `resource_grid`; `AntennaPanel` described its property
+  `horizontal_spacing` as the vertical spacing. They now name `openntn`, the current
+  paths of Sionna and the arguments of the signatures, and describe `doppler_enabled`.
+  No result changes. Affected earlier versions: 2.0.0a1 to 2.0.0a3; the versions for
+  Sionna 0.19 and 1.x have the same texts, apart from the module paths, which named the
+  module that their installation linked into Sionna.
+
+## [2.0.0a3] - 2026-10-13
+
+Third pre-release of OpenNTN for Sionna 2.x: the models take the argument `device`, and
+`torch.compile` captures their channel generation without graph breaks; the simulation
+results are unchanged.
+
+### Added
+
+- The argument `device` of `DenseUrban`, `Urban`, `SubUrban`, their scenarios and
+  `SystemLevelScenario`: the device on which the model creates its tensors and
+  computes, for example `"cuda:0"`. It defaults to `sionna.phy.config.device` at the
+  construction, the device that the models took before. The antenna arrays `ut_array`
+  and `bs_array` must be built on the same device (the argument `device` of the
+  antenna classes); otherwise the construction raises `ValueError`. Before, Sionna
+  2.0.1 accepted such an array, and Sionna 2.2.0 raised a `ValueError` of its own with
+  the same text. `set_topology` now draws the random LOS states on the device of the
+  model; before, it drew them on the device in `sionna.phy.config.device` at the time
+  of the call, which failed when that was not the device of the model (2.0.0a1 and
+  2.0.0a2).
+
+### Changed
+
+- A model that was moved to another device or cast to another precision after its last
+  `set_topology`, for example with `.to()`, `.cuda()` or `.double()`, raises
+  `ValueError` when it is called: the message names the device and the precision of the
+  construction and those of the tensors, and advises to build the model with the device
+  and precision it is to use. The models stay on the device and in the precision of
+  their construction, as Sionna's models do: `.to()` moves or casts the registered
+  tensors, but not the device and the precision of the model and its parts. Before, the
+  call failed inside the channel generation with a `RuntimeError` such as "Tensor on
+  device cpu is not on the expected device meta!" after a move, or "expected scalar
+  type Double but found ComplexFloat" after a cast to double precision. A model moved
+  or cast at another point, before `set_topology` or before a new topology, still fails
+  with an error from PyTorch, in `set_topology` or in the call.
+
+### Fixed
+
+- `torch.compile` captures the channel generation of `DenseUrban`, `Urban` and
+  `SubUrban` without graph breaks, also with `fullgraph=True`, and the results are
+  unchanged. The Faraday rotation of TR 38.811 V15.4.0, clause 6.8.2, converted its
+  angle ψ = 108°/f², with f in GHz (eq. (6.8-2)), from a tensor to a Python number in
+  every call, and `DenseUrban` held its number of rays per cluster, 20, as a tensor,
+  which the generation of the rays converted to a Python number to use it as a slice
+  bound. With the code of 2.0.0a2, `torch._dynamo.explain` found 11 graph breaks in a
+  call of `Urban` or `SubUrban` and 19 in a call of `DenseUrban` (20 with
+  `always_generate_lsp=True`), in both link directions and both precisions, and
+  `fullgraph=True` failed. The angle is now computed once, when the model is built, and
+  the number of rays per cluster is the integer 20 in the three models. Affected earlier
+  versions: 2.0.0a1 and 2.0.0a2; the versions for Sionna 0.19 and 1.x are built on
+  TensorFlow and are not affected.
+- With 2.0.0a2, on a machine with a CUDA GPU, 7 cases of
+  `tests/differentiable_topology_test.py` and `tests/lsp_spatial_correlation_test.py`
+  failed, because the tests built tensors on the CPU: defects of the tests, not of the
+  package. The tests now build these tensors on the device of the topology.
+  `tests/lsp_spatial_correlation_test.py` now reads the parameter files through the
+  installed package, so that it also runs against an installed package. No result of the
+  package changes. Affected earlier version: 2.0.0a2.
+
+## [2.0.0a2] - 2026-10-12
+
+Second pre-release of OpenNTN for Sionna 2.x. Its corrections change simulation results;
+they are listed first, each with the earlier versions it affects.
+
+### Changes that affect simulation results
+
+Elevation angles in these entries are the reference angles of TR 38.811 (10° to 90° in
+steps of 10°); the models use the parameters of the reference angle nearest to the
+elevation angle.
+
+- Ray offset angles: ray 16 of every cluster is now placed at -1.1481 times the
+  intra-cluster angle spread, as in TR 38.901 V16.1.0, Table 7.5-3, which TR 38.811
+  V15.4.0, clause 6.7.2, uses; it was placed at -0.1481 times that spread. This
+  concerns the azimuth and zenith angles of arrival and departure of every cluster, in
+  all scenarios, link states, bands and directions. With the wrong value the 20 offsets
+  had a mean of 0.050 instead of 0 and a standard deviation of 0.966 instead of 1.000,
+  so the rays of each cluster were shifted by 0.05 times the intra-cluster spread and
+  their spread was 3.4 % too small (computed from the table). Affected earlier
+  versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Sub-urban NLOS, Ka band, uplink, 80°: the mean of lg ZSD is now -3.20, as in TR
+  38.811 V15.4.0, Table 6.7.2-6b; the parameter file held -3.30. The median ZSD (the
+  zenith spread on the satellite side) and the zenith spread of the rays within each
+  cluster on that side (TR 38.901 V16.1.0, eq. (7.5-20)) were 21 % too small:
+  10^-3.30 instead of 10^-3.20 degrees, that is 0.00050° instead of 0.00063°. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Urban NLOS, Ka band: the cluster ASA (c_ASA) at 30°, 40° and 50° is now 16.4°, 17.86°
+  and 19.74°, as in TR 38.811 V15.4.0, Table 6.7.2-4b; the parameter files of both
+  directions held 18.14°, 16.04° and 17.86°. The azimuth spread of the rays within each
+  cluster on the UT side was 11 % too large at 30° and 10 % too small at 40° and 50°.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Angular scaling factors: the factors C_phi^NLOS and C_theta^NLOS now follow the
+  number of clusters of the link state at the elevation angle, as in TR 38.811 V15.4.0,
+  Tables 6.7.2-1aa and 6.7.2-1ab, with the numbers of clusters of Tables 6.7.2-3a to
+  6.7.2-6b (TR 38.901 V16.1.0, clause 7.5, step 7: "a scaling factor related to the
+  total number of clusters"). Each parameter file held one pair of factors for all
+  elevation angles, so the factors of another number of clusters were used for urban
+  LOS links at 10° (4 clusters, the factors of 3 used), urban NLOS and sub-urban LOS
+  links at 70° to 90° (2 clusters, the factors of 3 used) and sub-urban NLOS links at
+  60° to 90° (3 clusters, the factors of 4 used), in both bands and both directions;
+  dense urban is not affected. The cluster angles of TR 38.901 eqs. (7.5-9) and
+  (7.5-14) are divided by these factors: they were 0.74 times (azimuth) and 0.72 times
+  (zenith) their specified size where 2 clusters apply, 0.87 and 0.85 times for
+  sub-urban NLOS links at 60° to 90°, and 1.15 and 1.17 times for urban LOS links at
+  10°. The factors are now tables in the code; the keys `CPhiNLoS` and `CThetaNLoS`
+  were removed from the parameter files. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1.
+- Atmospheric parameters of `set_topology` (`latitude`, `atmospheric_pressure`,
+  `temperature`, `water_vapor_density`): a value now persists until it is passed again,
+  as the `set_topology` documentation states, and a value passed without new topology
+  tensors recomputes the gas and scintillation losses (TR 38.811 V15.4.0, clauses 6.6.4
+  and 6.6.6) without drawing anything else again. Before, every call reset the
+  parameters it did not pass to their defaults, and the losses were recomputed only
+  together with a new topology. The `set_topology` method of `DenseUrban`, `Urban` and
+  `SubUrban` now accepts these parameters and forwards them to the scenario; it raised a
+  `TypeError` before. The default temperature and pressure are now those of the
+  reference atmosphere of clause 6.6.4 (see the entry on the default atmosphere).
+  `latitude` decides the ionospheric scintillation loss (see the entry on the
+  ionospheric scintillation). The parameters `lwc`, `rain_rate`, `relative_humidity`,
+  `diameter_earth_antenna` and `antenna_efficiency`, which had no effect and which the
+  scenario reset in the same way, are removed (see "Removed"). This concerns every
+  simulation that sets atmospheric parameters. For example, at 30° in the urban
+  scenario, a temperature of 300 K instead of the default changes the gas loss from
+  0.0728 dB to 0.0654 dB at 2.2 GHz and from 0.5156 dB to 0.4939 dB at 20 GHz (computed
+  with this version); where earlier versions had reset a value or not applied it, they
+  used the defaults instead. Affected earlier versions: the versions for Sionna 0.19 and
+  1.x, and 2.0.0a1.
+- Cluster elimination: clusters with less than -25 dB power compared with the
+  strongest cluster of their link are now removed, as in TR 38.901 V16.1.0, clause
+  7.5, step 6, which TR 38.811 V15.4.0, clause 6.7.2, uses; before, all clusters were
+  kept with their own delays and angles. The threshold applies to the cluster powers of
+  TR 38.901 eq. (7.5-6) on LOS and NLOS links, before the LOS specular component is
+  added. A removed cluster gets zero power, so the shapes of the outputs do not
+  change, and the remaining powers are not renormalized. This concerns all scenarios,
+  link states, bands and directions. In 10,000 links per scenario, link state and
+  elevation angle (S band, downlink, computed with 2.0.0a1), 0.6 % to 21 % of the links
+  had at least one such cluster. A removed cluster has less than 10^-2.5 (0.32 %) of
+  the power of the strongest cluster, so the cluster power of a link is now lower by
+  less than 0.32 %, 0.63 % or 0.94 % (0.014 dB, 0.028 dB or 0.042 dB) with 2, 3 or 4
+  clusters; on LOS links the specular component is not affected, so the power of the
+  link falls by less. Affected earlier versions: the versions for Sionna 0.19 and 1.x,
+  and 2.0.0a1.
+- LOS arrival direction: the zenith angle of arrival of the LOS ray is now 180° minus
+  the zenith angle of departure, so that the LOS arrival direction is the reversed
+  departure direction, as in TR 38.901 V16.1.0, clause 7.5, step 1c, and clause 7.1,
+  which TR 38.811 V15.4.0, clause 6.7.2 and eq. (6.8-1b), uses. It was the zenith angle
+  of departure plus 180°, which together with the azimuth angle of arrival points at
+  the elevation of the satellite but in the opposite azimuth: the arrival direction was
+  off by 180° minus twice the elevation angle, that is by 160°, 120° and 60° at 10°,
+  30° and 60°, and correct only at 90°. This concerns the LOS ray of every LOS link,
+  in all scenarios, bands and directions: the antenna pattern at the UT, the phases
+  across a UT antenna array, and the Doppler shift of the LOS ray due to the motion of
+  the UT. With one 38.901 element at the UT and a 1 x 4 dual-polarized 38.901 panel at
+  the satellite, the mean power of the first path (the LOS ray and the first cluster)
+  was 10.0 dB, 12.7 dB and 9.0 dB too low at 10°, 30° and 60°, in both directions
+  (dense urban, S band, 1000 forced LOS links, path loss and shadow fading off,
+  computed); with one element with an isotropic pattern at the UT the power does not
+  change. The zenith angles of arrival of the clusters keep their distribution,
+  because TR 38.901 eq. (7.5-18) reflects them into [0°, 180°]; their realizations
+  change. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Sign of the LOS propagation phase: the LOS ray now contains exp(-j2π d3D/λ0), as in
+  TR 38.901 V16.1.0, eq. (7.5-29), and TR 38.811 V15.4.0, eq. (6.8-1b); it contained
+  exp(+j2π d3D/λ0). The minus sign is the one consistent with the Doppler term of the
+  same equation: when a UT moves towards the satellite, d3D shrinks and -2π d3D/λ0
+  grows at the rate of the Doppler phase. This concerns every
+  LOS link, in all scenarios, bands and directions. The statistics do not change,
+  because the phases of the rays of the first cluster, to which the LOS ray is added,
+  are uniformly distributed; single realizations of the channel coefficients of LOS
+  links change. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1.
+- LOS propagation phase in single precision: the phase 2π d3D/λ0 of the LOS ray is now
+  computed in double precision and reduced modulo 2π before the cast to the model
+  precision, with d3D of TR 38.811 V15.4.0, eq. (6.6-3), evaluated in double precision
+  in a form without cancellation, and with the carrier frequency as given (in single
+  precision, 30 GHz, for example, is not representable). Before, d3D and the phase were
+  computed in the model precision. In single precision one unit in the last place of
+  d3D at 600 km (6.25 cm) is 0.4 to 6 wavelengths between 2 and 30 GHz, and d3D was off
+  by up to 0.77 m at 600 km (0.5 m at 90°), so the phase was off by up to 3.1 rad, that
+  is random. The LOS ray is added coherently to the first cluster, so the first-path
+  power and the total power of single LOS links were affected; their distributions
+  were not. The phase is now within 1e-6 rad of the specified value in single and
+  double precision; for elevation angles from 10° to 90° in steps of 0.5°, 600 km and
+  carrier frequencies of 2.0, 2.2, 2.5, 3.0, 3.5, 4.0, 20, 22.5, 25, 27.5 and 30 GHz,
+  the largest errors are 4.3e-7 rad in single and 2.3e-7 rad in double precision
+  (computed). `distance_3d` is within half a unit in the last place in single
+  precision and within 2 units in double precision (600 km, steps of 0.25°,
+  computed). In double precision the results change only within rounding.
+  This concerns every LOS link in single precision, in all scenarios, bands and
+  directions. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1, in single precision, their default precision.
+- Zenith angle of departure offset: the offset mu_offset,ZOD of TR 38.901 V16.1.0,
+  eq. (7.5-19), is now 0 for every link. TR 38.811 V15.4.0, clause 6.7.2, replaces the
+  parameter tables of TR 38.901 with its own tables, which have no offset row. Before,
+  NLOS links used the expression of TR 38.901 Table 7.5-9 (rural macro cell,
+  arctan((35 - 3.5)/d_2D) - arctan((35 - 1.5)/d_2D)) in all three scenarios, computed
+  in radians and added to angles in degrees. This concerns the zenith angles of the
+  clusters of NLOS links at the satellite, in all scenarios, bands and directions; LOS
+  links had no offset. Because the zenith spread of the clusters at the satellite is
+  zero, the offset was the whole deviation of these clusters from the LOS zenith
+  direction: less than 6e-6° at elevation angles of 10° to 60° (horizontal UT
+  distances of 346 km and more) and at most 0.031° at 90°, for a UT 32 m from the point
+  below the satellite (computed from the expression). The clusters of NLOS links are
+  now centred on the LOS zenith direction at the satellite. The effect on the channel
+  is small: of the reference outputs of the test suite (600 km), only the per-antenna
+  powers of urban NLOS links at 90° changed by more than a relative 1e-5. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Random signs and ray coupling per link: the random sign X_n of the cluster angles
+  (TR 38.901 V16.1.0, clause 7.5, step 7, eqs. (7.5-11), (7.5-16) and (7.5-19)) and the
+  random coupling of the rays within a cluster (step 8) are now drawn for each link
+  between a satellite and a UT, as TR 38.901 defines them; TR 38.811 V15.4.0, clause
+  6.7.2, uses this procedure unchanged. Before, both were drawn once per satellite and
+  shared by all its UTs, so the clusters of all UTs of a satellite had the same signs of
+  their angle offsets and the same pairing of ray angles. This concerns simulations
+  with more than one UT per satellite, in all scenarios, link states, bands and
+  directions: the angles of different UTs were correlated, which matters for
+  multi-user and interference studies. For example, for two UTs in the dense urban
+  scenario (S band, 30°, 2000 drops), the sign of the AOA offset of a cluster (other
+  than the strongest) agreed in 69 % (NLOS) to 76 % (LOS) of the clusters instead of
+  50 %, and on average 14 to 16 of the 20 rays of a cluster had the same position in
+  the AOA order instead of 1 (computed with the earlier version). The statistics of a
+  single link do not change. Because more random numbers are drawn, the realizations
+  of the cluster angles, of the ray coupling and of the random variables drawn after
+  them (cross-polarization ratios, initial phases) change for every link. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Ray coupling in the two strongest clusters: the rays of the two strongest clusters
+  of a link are now coupled within the sub-clusters of TR 38.901 V16.1.0, Table 7.5-5,
+  as clause 7.5, step 8, requires ("within a cluster n, or within a sub-cluster in the
+  case of two strongest clusters"), and step 11 splits the same two clusters into
+  sub-clusters (eq. (7.5-26)); TR 38.811 V15.4.0, clause 6.7.2, uses these steps
+  unchanged. The two strongest clusters are those with the largest powers of eq.
+  (7.5-6) after the cluster elimination. Before, the rays of every cluster were coupled
+  across the whole cluster, so a sub-cluster, at its own delay, carried angle offsets
+  drawn from all 20 rays of its cluster instead of those that Table 7.5-5 assigns to
+  it, and step 11 sorted the clusters again, which could select a different second
+  cluster when fewer than two clusters of a link have non-zero power (34 of 2100
+  sampled links in the sub-urban scenario, LOS, S band, 30°). This concerns all
+  scenarios, link states, bands and directions. The angular spread of each sub-cluster
+  now follows Table 7.5-5: the rms ray offsets of the sub-clusters R1, R2 and R3 are
+  0.987, 1.006 and 1.025 times the intra-cluster spread (computed from Tables 7.5-3
+  and 7.5-5), where before they were random around 1; the total angular spread of a
+  cluster, the cluster powers and delays and the mean power of a link do not
+  change. Realizations of the channel coefficients change for every link; no
+  additional random numbers are drawn. Affected earlier versions: the versions for
+  Sionna 0.19 and 1.x, and 2.0.0a1.
+- Correlation of the large scale parameters: the large scale parameters (delay
+  spread, angular spreads, Ricean K-factor, shadow fading) are now generated as TR
+  38.901 V16.1.0, clause 7.5, step 4, prescribes, which TR 38.811 V15.4.0, clause
+  6.7.2, uses with its own tables: with the procedure of clause 3.3.1 of WINNER II
+  D1.1.2 (TR 38.901 reference [14]), which first correlates each parameter spatially
+  across the UTs, with the correlation exp(-d/d_corr) of its correlation distance,
+  and then applies the square root of the cross-correlation matrix of the parameters
+  to each link, with the Cholesky decomposition in the order [SF, K, DS, ASD, ASA,
+  ZSD, ZSA] that step 4 specifies. Before, the cross-correlation was applied first and
+  the spatial correlation second, so the order of the Cholesky decomposition had no
+  effect, and the parameters of one link did not have the cross-correlations of the
+  tables when other UTs of the same link state were within the correlation distances
+  (12 m to 50 m in the TR 38.811 tables of the three scenarios). This concerns all
+  scenarios, link states, bands and directions. For example, for two UTs 20 m apart
+  (dense urban, LOS, S band, downlink, 30°, 20,000 drops, computed), the correlation
+  between the logarithms of the zenith spreads of arrival of the two UTs is now 0.507
+  instead of 0.265, of the azimuth spreads of arrival 0.440 instead of 0.268 and of the
+  delay spreads 0.473 instead of 0.512, against 0.505, 0.435 and 0.473 computed from the
+  tables; the shadow fading and the K-factor keep their correlations exp(-d/d_corr),
+  0.582 and 0.189 (measured 0.587 and 0.191). In the same setting, for the second of two
+  UTs 5 m apart, the cross-correlation between its zenith spread of arrival and its
+  shadow fading was -0.776 instead of -0.800 of the table (the first had -0.805). For
+  UTs that are far apart compared with the correlation distances the statistics do not
+  change. The realizations of all large scale parameters change for every link, and with
+  them the cluster delays and powers; the path losses and the link states do not change.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Satellites from 160 km to below 600 km: the Doppler shift due to the motion of the
+  satellite now applies to every satellite, that is to every platform from 160 km to
+  35 786 km, as the zero departure spreads at the satellite (NOTE 8 of TR 38.811
+  V15.4.0, Tables 6.7.2-1a to 6.7.2-8b) did before; it applied from 600 km on. The
+  Faraday rotation of clause 6.8.2 still applies from 600 km on, as in the earlier
+  versions: the clause gives it "In case of propagation for spaceborne BS above the
+  ionosphere", and 600 km is the lowest satellite altitude of TR 38.811 (Table 4.5-1,
+  LEO; Table 5.1-1, non-GEO down to 600 km), while a lower satellite flies within the F
+  region of the ionosphere, below much of its electron content. This concerns satellites
+  from 160 km to below 600 km, in all scenarios, link states, bands and directions:
+  their channel coefficients now contain the phase drift of the satellite Doppler shift.
+  TR 38.811, clause 3.1, places LEO satellites from 500 km on; the models accept
+  satellites from 160 km, very low orbits included. Results at 600 km and above and of
+  aerial platforms (8 km to 50 km) do not change. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Satellite Doppler shift: the Doppler shift due to the motion of the satellite is now
+  f_c (v_sat/c) (R/(R+h)) cos(α) cos(2πU), with the speed v_sat of a circular orbit at
+  the height h, the radius R of the Earth, the elevation angle α and U uniform per batch
+  example (the direction of the orbit relative to the UTs, as before); its largest value
+  is the shift of TR 38.811 V15.4.0, clauses 6.7.1 and 6.9.2, f_d = (v_sat/c) (R/(R+h))
+  cos(α) f_c. The factor R/(R+h) was missing, so the shift was (R+h)/R times too large:
+  by 9.4 % at 600 km, 18.8 % at 1200 km and 157 % at 10 000 km. At 10° the largest shift
+  was 24.84 ppm of the carrier frequency at 600 km and 23.84 ppm at 1200 km and is now
+  22.70 ppm and 20.06 ppm (TR 38.821 V16.2.0, Table 6.1.1.1-8, gives 24 ppm and 20 ppm
+  as the maximum Doppler shifts for elevation angles of at least 10°). A geostationary
+  satellite (35 786 km) now induces no Doppler shift, as TR 38.811, clause 5.3.2.3,
+  states ("the Geostationary satellite is fixed and therefore no Doppler shift is
+  induced except that due to possible UE motion"); it had up to 10.1 ppm, 202 kHz at
+  20 GHz. As in clause 5.3.4.3, the rotation of the Earth is neglected. This concerns
+  the phase of the channel coefficients over time on every satellite link with the
+  Doppler shift enabled, in all scenarios, link states, bands and directions; powers,
+  delays and angles do not change, and of the reference outputs of the test suite only
+  the time correlations changed. Affected earlier versions: the versions for Sionna 0.19
+  and 1.x, and 2.0.0a1.
+- Departure spreads at the satellite in the uplink: NOTE 8 of TR 38.811 V15.4.0, Tables
+  6.7.2-1a to 6.7.2-8b ("For satellite (e.g.GEO/LEO), the departure angle spreads are
+  zeros, i.e. µlgASD and µlgZSD are –∞, and corresponding standard deviations are
+  zeros"), now applies to satellites in both directions, with zero cross-correlations of
+  ASD and ZSD with the other large scale parameters, as the downlink parameter files had
+  it. The satellite is the departure end of the rays as the generator draws them in both
+  directions, so the ASD and ZSD of the uplink are spreads at the satellite as well.
+  Before, only the downlink parameter files applied the note: in the uplink the large
+  scale parameters ASD and ZSD took the values of the tables (medians of up to 0.78° and
+  0.09°), the rays of each cluster kept the intra-cluster zenith spread (3/8)·10^µlgZSD
+  of TR 38.901 V16.1.0, eq. (7.5-20), at the satellite (rays up to 0.07° from their
+  cluster), and the cross-correlations with ASD and ZSD entered the generation of the
+  other large scale parameters, whose realizations therefore change (their distribution
+  per link does not). The parameter files now hold the tables as printed, and the models
+  apply the note for satellites (160 km to 35 786 km). Aerial platforms (8 km to 50 km)
+  now take the spreads of the tables in both directions; in the downlink the parameter
+  files had given them zero spreads. This concerns the uplink of satellites and both
+  directions of aerial platforms, in all scenarios, link states and bands. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Cluster azimuth spread at the satellite: every ray now leaves a satellite in the
+  azimuth of its cluster, that is in the LOS direction. TR 38.811 V15.4.0, NOTE 8 of
+  Tables 6.7.2-1a to 6.7.2-8b, makes the departure angle spreads of a satellite zero,
+  and clause 6.9.1 states for the link-level models derived from this model that "the
+  azimuth and elevation angular spreads of departure, ASD and ZSD, can be considered
+  zero, and all cluster AOD and ZOD angles the same". Before, the rays of the urban
+  scenario left the satellite at the AOD of their cluster plus c_ASD α_m (TR 38.901
+  V16.1.0, eq. (7.5-13), applied to the departure), with the cluster ASD c_ASD of the
+  urban Tables 6.7.2-3a to 6.7.2-4b (0.08° to 6.63°) and the offsets α_m of Table 7.5-3:
+  up to 14.3° in azimuth at 90°, and up to 0.74° from the LOS direction (urban NLOS, S
+  band, 70°; computed). Seen from a satellite at 600 km, a scatterer 1 km from the UT is
+  at most 0.03° (10°) to 0.1° (90°) off the direction of the UT. The dense urban and
+  sub-urban tables print c_ASD = 0, so only urban links of satellites (160 km to
+  35 786 km) are affected, in both directions, link states and bands; aerial platforms
+  keep the c_ASD of the tables. In the reference outputs of the test suite (600 km), the
+  power of single urban links changed by up to 0.79 dB and the power per antenna pair by
+  up to 0.07 times the largest of its configuration (NLOS, 90°); the total power of a
+  configuration of six links changed by less than 0.1 dB. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Default atmosphere: the defaults of `temperature` and `atmospheric_pressure` are now
+  the mean annual global reference atmosphere of TR 38.811 V15.4.0, clause 6.6.4 ("For
+  all UEs, [...] T = 288.15 K, p = 1013.25 hPa, ρ = 7.5 g/m³"). The defaults were 273 K
+  and 1020 hPa, and the formulas of ITU-R P.676 and P.618-13 that took the temperature
+  in degrees Celsius used T - 273 instead of T - 273.15. The gas loss of this version,
+  by ITU-R P.676-10, takes t = T - 273.15 (see "Gas attenuation"), and the scintillation
+  loss no longer uses the temperature (see "Tropospheric scintillation"). With the gas
+  formulas of the earlier versions, the defaults give a lower gas loss, for example
+  0.2056 dB instead of 0.2405 dB at 2.2 GHz and 10°, 1.4779 dB instead of 1.5711 dB at
+  20 GHz and 10°, 1.4019 dB instead of 1.5941 dB at 30 GHz and 10°, and 0.2434 dB
+  instead of 0.2768 dB at 30 GHz and 90° (urban, 600 km, computed). Where the
+  temperature was set, the conversion alone changes the gas loss of these formulas at
+  288.15 K by at most 0.0031 dB in the bands that the models accept (at 40 GHz and 10°,
+  1013.25 hPa, 7.5 g/m³). This concerns the gas loss in all scenarios, link states,
+  bands and directions; the scintillation loss no longer depends on the temperature.
+  Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Gas attenuation: the gas loss now follows Annex 2 of ITU-R P.676-10 (09/2013) as
+  printed: the specific attenuation of dry air of eqs. (22a) to (22u), every branch from
+  1 GHz to 350 GHz, that of water vapour of eqs. (23a) to (23d), the equivalent heights
+  of eqs. (25a) to (25e), (26a) and (26b), and the slant path of eqs. (27) and (28)
+  with TR 38.811 V15.4.0, eq. (6.6-8). The atmospheric pressure is the dry air pressure
+  p, as clause 6.6.4 calls its 1013.25 hPa; Annex 2 takes the total pressure p + e, with
+  the water vapour partial pressure e = ρT/216.7 (Annex 1, eq. (4)), and the temperature
+  t = T - 273.15 in degrees Celsius. Clause 6.6.4 takes "The method of Annex 2 in ITU-R
+  P.676" without a version, and clause 2 makes the latest version apply to such a
+  reference ("For a non-specific reference, the latest version applies."). P.676-10 is
+  the last revision whose Annex 2 gives the specific attenuations and the equivalent
+  heights in closed form, the form of the earlier versions; with the reference
+  atmosphere of clause 6.6.4 the revision in force, P.676-13 (08/2022), gives at 10° in
+  the bands that the models accept between 0.2738 dB more (22.41 GHz; 0.2663 dB more at
+  22.74 GHz, outside the 0.5 GHz around the water vapour line at 22.235 GHz that clause
+  6.6.4 excludes from the method) and 0.0676 dB less (40 GHz) (computed). Before, the
+  loss used curve fits in the form of Annex 2 of P.676-6 to P.676-10 up to 54 GHz, with
+  two coefficients of the water vapour equivalent height that no revision prints (-8.1
+  and 1.5, where P.676-6 to P.676-11 print -8.6 and 1.58), and followed no revision
+  above 54 GHz: from 54 GHz to 66 GHz it used expressions of P.676-5, one of them with
+  the factor of another equation, and from 66 GHz to 120 GHz the dry-air term of eq.
+  (22e) lacked the factor f² rp² 10^-3 and had 1.15 ξ5 in the exponent of (f - 66)
+  instead of added to its power. The second made the zenith loss of dry air jump from
+  10.8 dB at 65.9 GHz to 985 dB at 66.1 GHz, and together they made the attenuation
+  1.6 % to 7.4 % of the printed value from 69 GHz to 120 GHz (dry air, zenith, 288.15 K,
+  1013.25 hPa, computed with the earlier formulas). Above 120 GHz dry air did not
+  attenuate, the limit of eq. (25e) on the equivalent height of dry air was missing, and
+  the atmospheric pressure was taken as the total pressure. With the reference
+  atmosphere of clause 6.6.4 and at 10°, the loss changes against the earlier formulas
+  with t = T - 273.15 by +0.0040 dB at 2 GHz, +0.0041 dB at 2.2 GHz, +0.0068 dB at
+  20 GHz (1.4846 dB instead of 1.4779 dB) and +0.0186 dB at 30 GHz (1.4205 dB instead of
+  1.4019 dB); in the bands that the models accept (1.9 GHz to 4 GHz and 19 GHz to
+  40 GHz) the change lies between -0.0112 dB (22.29 GHz; -0.0089 dB at 22.74 GHz,
+  outside the excluded range) and +0.0384 dB (40 GHz), relative between -0.37 %
+  (22.25 GHz) and +2.04 % (3.75 GHz), mostly from the dry reading of the pressure
+  (computed). In dB the change scales with 1/sin of the elevation angle. This concerns
+  all scenarios, link states, bands and directions. `utils.compute_pathloss_gas` raises
+  `ValueError` for carrier frequencies outside 1 GHz to 350 GHz, the range of Annex 2;
+  the models accept only the S and Ka bands. As clause 6.6.4 states, the method does not
+  apply to UE altitudes above 10 km and to frequencies within 0.5 GHz of the centres of
+  resonance lines; the models do not check either. Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Tropospheric scintillation: in the Ka band the scintillation loss is now the value of
+  TR 38.811 V15.4.0, Table 6.6.6.2.1-1, at the reference elevation angle nearest to the
+  elevation angle: 1.08 dB at 10°, 0.48 dB at 20°, 0.30 dB at 30°, 0.22 dB at 40°,
+  0.17 dB at 50°, 0.13 dB at 60° and 0.12 dB from 70° to 90°, as clause 6.6.6.2.1
+  specifies ("As baseline for system-level simulations above 6 GHz, the fading due to
+  scintillation PL_S is equal to the tropospheric attenuation level at 99% of the time
+  derived from Figure 6.6.6.2.1-1 and summarized in Table 6.6.6.2.1-1"). The table is
+  given for 20 GHz in Toulouse; it is used at every carrier frequency from 6 GHz on, as
+  TR 38.821 V16.2.0 uses it at 20 GHz and 30 GHz (Tables 6.1.3.2-1 and 6.1.3.3-1). An
+  elevation angle midway between two reference angles takes the higher one. Before, the
+  loss was the fade depth of ITU-R P.618-13, section 2.4.1, exceeded 0.01 % of the time,
+  where the table gives the depth exceeded 1 % of the time; by that method the depth
+  exceeded 0.01 % of the time is 2.40 times the depth exceeded 1 %. It was computed at
+  the elevation angle itself, from the temperature, the relative humidity and the
+  diameter and efficiency of the Earth-stationed antenna. That computation also
+  evaluated the effective path length of eq. (41) of P.618-13 as
+  2h_L/sqrt(sin²θ + 2.35·10^-4 + sin θ) instead of
+  2h_L/(sqrt(sin²θ + 2.35·10^-4) + sin θ), which changed the loss by -1.85 % at 10° and
+  by +6.68 % at 90° (20 GHz, the default antenna, computed). With the defaults of the
+  earlier versions the loss was 1.5507 dB, 0.4090 dB and 0.1660 dB at 20 GHz and 10°,
+  30° and 90°, and 1.8943 dB at 30 GHz and 10° (urban, 600 km, computed); it is now
+  1.08 dB, 0.30 dB, 0.12 dB and 1.08 dB. This concerns the scintillation loss of every
+  link in the Ka band, in all scenarios, link states and directions; it no longer
+  depends on the atmospheric parameters (see "Removed"). Affected earlier versions: the
+  versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Ionospheric scintillation: in the S band the scintillation loss is now the ionospheric
+  scintillation A_IS of TR 38.811 V15.4.0, clause 6.6.6.1.4, only for UTs within 20° of
+  the magnetic equator, and 0 dB at higher latitudes, as the clause specifies ("For
+  system-level simulations below 6 GHz, PL_S is equal to A_IS from Equation (6.6-13) for
+  latitudes of maximum ±20°. For latitudes between ±20° and ±60° of latitude, PL_S = 0.
+  [...] the choice of PL_S = 0 is also applied for latitude above ±60°."). The latitude
+  is the `latitude` of `set_topology`, the latitude of the UTs relative to the magnetic
+  equator, around which the clause places the region of the model; its default of 47°
+  gives 0 dB, as the calibration of TR 38.821 V16.2.0 assumes (Table 6.1.1.1-5, NOTE 2:
+  "the ionospheric scintillation loss shall be considered equal to zero (i.e., the UEs
+  are located between 20 and 60 degrees of latitude)"). Before,
+  A_IS = 1.1 (f/4 GHz)^-1.5/√2 dB applied at every latitude, and `latitude` had no
+  effect. This concerns the scintillation loss of every link in the S band, in all
+  scenarios, link states and directions: with the default latitude it is now 0 dB
+  instead of 2.2 dB at 2 GHz and 1.9069 dB at 2.2 GHz; within 20° of the magnetic
+  equator it is unchanged. Affected earlier versions: the versions for Sionna 0.19 and
+  1.x, and 2.0.0a1.
+- Shadow fading applied once: the shadow fading now reaches the channel once, as the
+  large scale parameter SF of TR 38.901 V16.1.0, clause 7.5, step 4, which step 12
+  applies to the channel coefficients. TR 38.811 V15.4.0, clause 6.6.2, eq. (6.6-4),
+  writes one shadow fading term SF ~ N(0, σ_SF²) into the basic path loss, with σ_SF of
+  Tables 6.6.2-1 to 6.6.2-3, and Tables 6.7.2-1a to 6.7.2-8b give the cross-correlations
+  of this shadow fading with the other large scale parameters and its correlation
+  distance, which only the large scale parameter carries. Before, the basic path loss
+  contained a second shadow fading draw per link, independent of the first and without
+  these correlations, so the shadow fading applied to the channel had about √2 times
+  σ_SF and weaker correlations, and with `enable_shadow_fading=False` the draw in the
+  basic path loss remained. In the urban scenario (S band, downlink, 50°, 600 km, 4000
+  links with one UT each, LOS and NLOS forced, computed), the standard deviation of the
+  applied shadow fading is now 4.05 dB (LOS) and 6.08 dB (NLOS) instead of 5.60 dB and
+  8.57 dB (Table 6.6.2-2: 4 dB and 6 dB); its correlation with lg DS is now -0.411 and
+  -0.214 instead of -0.273 and -0.126 (Tables 6.7.2-3a and 6.7.2-4a: -0.4 and -0.21),
+  and with lg ZSA -0.776 and -0.356 instead of -0.528 and -0.271 (the tables give -0.8
+  and -0.38; with the limit of ZSA to 52° of step 4 the expected values are -0.771 and
+  -0.348); with `enable_shadow_fading=False` it is now 0 dB instead of a standard
+  deviation of 4.07 dB and 6.09 dB. Because two draws per link fewer are made, the
+  realizations of the large scale parameters, and with them of the cluster delays and
+  powers, change for every link; their statistics do not change, and neither do the link
+  states. This concerns all scenarios, link states, bands and directions. Affected
+  earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- Elevation angles midway between two reference angles: an elevation angle of 25°, 45°,
+  65° or 85° now takes the values of the higher reference angle, 30°, 50°, 70° or 90°,
+  in every table: the LOS probability, the shadow fading and the clutter loss, the large
+  scale parameters, the number of clusters and the cluster parameters. TR 38.811
+  V15.4.0, clauses 6.6.1 and 6.6.2, takes the reference angle nearest to the elevation
+  angle and gives no rule for an angle midway between two; the models now take the
+  higher one in every lookup. Before, Python's rounding of ties to the even number took
+  the lower angle, 20°, 40°, 60° or 80°, at these four angles (the tropospheric
+  scintillation table of this version already took the higher one); 15°, 35°, 55° and
+  75° took the higher angle and still do. For example, at 25° the LOS probability of the
+  urban scenario is now 0.493 instead of 0.386, and the clutter loss of the urban NLOS
+  links in the S band 29.0 dB instead of 30.9 dB (the values of Tables 6.6.1-1 and
+  6.6.2-2 at 30° and 20°). This concerns simulations at exactly 25°, 45°, 65° or 85°, in
+  all scenarios, link states, bands and directions; other elevation angles do not
+  change. Affected earlier versions: the versions for Sionna 0.19 and 1.x, and 2.0.0a1.
+- UTs at the same location: two or more UTs at the same location in the horizontal plane
+  and in the same link state, in one batch example, made the spatial correlation of the
+  large scale parameters (delay spread, angular spreads, Ricean K-factor, shadow fading)
+  of that batch example singular, because the correlation exp(-d/d_corr) of TR 38.901
+  V16.1.0, clause 7.5, step 4, which TR 38.811 V15.4.0, clause 6.7.2, uses, is 1 at the
+  distance d = 0. Its Cholesky factorization then stopped without an error at the second
+  of these UTs and left unfactored entries in the rows of the UTs after them, so the UTs
+  listed after them in that batch example got large scale parameters with a log-domain
+  variance above the specified one: 1 + 2c^2 times it in the case measured, a UT after
+  two co-located UTs with the correlation c to them; for the shadow fading of a UT 7.6 m
+  away (dense urban, LOS, S band, downlink, 30°, double precision, computed) 2.33 times,
+  a standard deviation 1.53 times the specified one. Where rounding left a pivot just
+  above 0 instead, mostly in single precision, the factorization continued past these
+  UTs or stopped at a later one. The gradients of the large scale parameters of that
+  batch example with respect to the UT locations were NaN. Now UTs at one location in
+  one state share their large scale parameters, as the correlation exp(-d/d_corr) = 1 at
+  d = 0 requires and as in the procedure of clause 3.3.1 of WINNER II D1.1.2, which step
+  4 uses, where such UTs read the same grid node; every UT has the specified
+  distribution, and the gradients are finite. This concerns topologies with such UTs,
+  whatever their heights (the distance is measured in the horizontal plane), in all
+  scenarios, link states, bands and directions; random drops practically never place
+  two UTs at one location. Topologies without them are unchanged bit for bit. Affected
+  earlier versions: 2.0.0a1. The versions for Sionna 0.19 and 1.x build the same
+  singular correlation for such UTs. The version for Sionna 1.x factors it with the
+  Cholesky decomposition of TensorFlow, which returns NaN for it without an error
+  (computed with TensorFlow 2.18.1 and 2.19.0 on the CPU), so that every UT of such a
+  batch example got NaN large scale parameters, also with only two UTs at one location;
+  the version for Sionna 0.19 takes a symmetric matrix square root, whose result for
+  such UTs depends on TensorFlow.
+
+### Added
+
+- `Rays` takes the optional argument `strongest_cluster_indices`, the indices of the two
+  strongest clusters of each link, whose rays are coupled within the sub-clusters and
+  which the channel coefficients split into sub-clusters; `RaysGenerator` sets it. Code
+  that constructs `Rays` directly can omit it; the two clusters with the largest powers
+  are then split, as before.
+
+### Changed
+
+- `Topology` takes the new argument `los_phase`, the LOS propagation phase per link,
+  which `ChannelCoefficientsGenerator` uses for the LOS ray; the scenarios provide it
+  as `los_phase`, and the channel models pass it. Code that constructs a `Topology`
+  directly has to pass it.
+- `set_topology` now checks the platform and raises `ValueError` unless `bs_loc` holds
+  one BS at the same height in every batch example, either a satellite from 160 km to
+  35 786 km or an aerial platform from 8 km to 50 km. TR 38.811 V15.4.0, clause 3.1,
+  places aerial platforms at 8 km to 50 km and LEO satellites from 500 km on; the models
+  accept satellites from 160 km, very low orbits included. Table 4.5-1 gives typical
+  altitudes (LEO 600 km to 1500 km, MEO 7000 km to 20 000 km, GEO 35 786 km),
+  Table 5.1-1 the reference deployments (GEO, non-GEO down to 600 km, UAS 8 km to
+  50 km). The models simulate one platform with one nominal elevation angle per call.
+  Before, `set_topology` accepted every input: with a second BS the channel call failed,
+  with heights that differed across the batch the 3D distance, the path loss and the
+  switches of the satellite took the height of the first batch example for every
+  example, and any height was used as given. The versions for Sionna 0.19 and 1.x and
+  2.0.0a1 use the height of the first batch example in the same way.
+- `utils.compute_pathloss_gas` raises `ValueError` for carrier frequencies outside 1 GHz
+  to 350 GHz, the range of Annex 2 of ITU-R P.676-10 (see "Gas attenuation").
+- `set_topology` raises `ValueError` for a `latitude` outside -90° to 90°, and a
+  `latitude` passed without new topology tensors now recomputes the scintillation loss.
+  `latitude` is documented as the latitude of the UTs relative to the magnetic equator
+  (see "Ionospheric scintillation"); before, any value was accepted, and it had no
+  effect.
+- `basic_pathloss` of the scenarios is now the free space path loss plus, for NLOS
+  links, the clutter loss, without a shadow fading term; the shadow fading is the large
+  scale parameter SF, which step 12 applies (see "Shadow fading applied once"). Before,
+  `basic_pathloss` contained a shadow fading draw of its own, as in the versions for
+  Sionna 0.19 and 1.x and 2.0.0a1.
+- `set_topology` raises `ValueError` if a UT is indoor (`in_state` `True`), before
+  anything is stored; the models accept outdoor UTs only, for satellites and aerial
+  platforms. TR 38.811 V15.4.0 considers only outdoor UTs for satellites (clause 6.1.2:
+  "Only outdoor conditions are considered for satellite operations"; Table 5.1-1: "100%
+  Outdoors" for the satellite deployments) and indoor UTs as well for aerial platforms
+  (clause 6.1.2: "additional indoor conditions are also considered for HAPS";
+  Table 5.1-1: "Indoor and Outdoor" for UAS including HAPS), but the models have neither
+  the building entry loss of clause 6.6.3 nor clusters for indoor UTs. Before, indoor
+  UTs were accepted, but all their clusters were masked, so their channel coefficients
+  were NaN (dense urban, 2 GHz, downlink, 30°, 600 km, LOS and NLOS forced, computed
+  with the code before this change: every coefficient of an indoor UT), and no building
+  entry loss applied. The indoor probability of the topology generators was already
+  forced to 0; its documentation now says that the argument is ignored. Affected earlier
+  versions: 2.0.0a1, and the versions for Sionna 0.19 and 1.x, which have the same code
+  for indoor UTs.
+
+### Removed
+
+- `utils.compute_stallite_doppler`, which nothing used: it returned the carrier
+  frequency plus the Doppler shift of TR 38.811 V15.4.0, clauses 6.7.1 and 6.9.2, which
+  the channel now applies (see "Changes that affect simulation results").
+- `TDL` and `CDL`, with their modules `openntn.tdl` and `openntn.cdl`: placeholders that
+  raised `NotImplementedError` when constructed (in the versions for Sionna 0.19 and 1.x
+  they were copies of the TR 38.901 models marked as not working). The link-level models
+  NTN-TDL and NTN-CDL of TR 38.811 V15.4.0, clause 6.9, are not implemented in this
+  version; their parameter files stay in `openntn/models`, and Sionna's `TDL` and `CDL`
+  of TR 38.901 are unchanged.
+- The rain and cloud attenuation of TR 38.811 V15.4.0, clause 6.6.5: the scenario
+  property `additional_pathloss` (it raised `AttributeError`, because no model computed
+  the attenuation) and `utils.compute_pathloss_additional`, which failed when called.
+  The models follow the baseline of clause 6.6.5, "For system-level simulations, the
+  baseline is to consider clear sky conditions only", as before.
+- The parameters `lwc` and `rain_rate` of `set_topology` and the scenario properties of
+  the same names: only the rain and cloud attenuation read them, which no model applied,
+  so they had no effect.
+- The parameters `relative_humidity`, `diameter_earth_antenna` and `antenna_efficiency`
+  of `set_topology` and the scenario properties of the same names: only the computation
+  of the tropospheric scintillation by ITU-R P.618-13 read them, and the scintillation
+  loss now follows TR 38.811 V15.4.0, Table 6.6.6.2.1-1 (see "Tropospheric
+  scintillation"), so they had no effect. A call that passes one of the five removed
+  parameters raises `TypeError`; the scenario of 2.0.0a1 and of the versions for Sionna
+  0.19 and 1.x accepted them.
+- `utils.gen_single_sector_topology_interferers`, which failed in every scenario. The
+  models simulate one satellite or aerial platform per call (see "Changed");
+  `utils.gen_single_sector_topology` generates the topology of one sector.
+
+### Fixed
+
+- `set_topology` no longer modifies the tensors passed to it. The scenario kept the
+  topology tensors (UT and BS locations, orientations, UT velocities, indoor state) as
+  its own storage when they already had the model's data type and device, and a later
+  call wrote its values into them, so a tensor passed in an earlier call changed. The
+  scenario now keeps copies, through which gradients with respect to the tensors
+  passed still flow. The outputs for given inputs are unchanged; results change only
+  for scripts that used such a tensor again after a later `set_topology` call.
+  Affected earlier version: 2.0.0a1; the versions for Sionna 0.19 and 1.x are not
+  affected.
+- Repeated `set_topology` calls with topology tensors that require gradients: every call
+  now works, and the quantities of the scenario, for example the 2D distances and the
+  LOS angles, stay differentiable with respect to the tensors of that call. Before, the
+  scenario wrote the values of a later call into the tensors it had stored in the first
+  call, which carried the graph of that call, so the backward pass of the second
+  differentiable call failed with "Trying to backward through the graph a second time",
+  and after a differentiable call the stored tensors kept requiring gradients in calls
+  without gradients. Calls without gradients still write in place, so that the stored
+  tensors keep their addresses. The path coefficients and delays of the channel models
+  are detached, as before and as in Sionna; their documentation now says so. Affected
+  earlier version: 2.0.0a1; the versions for Sionna 0.19 and 1.x are not affected.
+- Gradients with respect to the UT locations of the scenario property
+  `matrix_ut_distance_2d`, the distances between the UTs in the horizontal plane, and of
+  the large scale parameters, whose spatial correlation is computed from these
+  distances, were NaN in the x and y coordinate of every UT: the distance of a UT to
+  itself, which is always 0, was taken as a square root, whose derivative at 0 is
+  infinite. They are now finite: the distance of a UT to itself has the derivative 0,
+  and the distances are unchanged bit for bit. For example, with the dense urban
+  scenario at 2.2 GHz, downlink, 30°, 600 km and 2 batch examples with 3 UTs each, the
+  gradient of the sum of the distances was NaN in 12 of the 18 entries of the UT
+  locations, and so were the gradients of the delay spread, the angular spreads, the
+  Ricean K-factor and the shadow fading. This concerns users who differentiate with
+  respect to the UT locations through these quantities; the path coefficients and delays
+  are detached and not affected, and no simulation result changes. Affected earlier
+  versions: 2.0.0a1, and the versions for Sionna 0.19 and 1.x, which compute the
+  distances between the UTs in the same way.
+- The test suite now checks the data files of the NTN-CDL and NTN-TDL models in
+  `openntn/models` against TR 38.811 V15.4.0, Tables 6.9.1-1 to 6.9.1-4 and 6.9.2-1 to
+  6.9.2-4, value by value; it called them TR 38.901 profiles without TR 38.811 values.
+  The files equal the tables; no model of the package reads them.
+- Documentation of the scenario properties: `doppler_enabled` had the description of the
+  antenna efficiency; it now states that the flag switches the Doppler shift due to the
+  motion of the satellite, while the Doppler shift due to the motion of the UTs is
+  always applied. `gas_pathloss`, `scintillation_pathloss` and `entry_pathloss` called
+  themselves the basic path loss of TR 38.901, clause 7.4.1; they now name the gas
+  attenuation (TR 38.811 V15.4.0, clause 6.6.4, eq. (6.6-8)), the scintillation loss
+  (clause 6.6.6) and the building entry loss (clause 6.6.3). `free_space_pathloss` now
+  cites eq. (6.6-2) of TR 38.811 instead of TR 38.901, clause 7.4.1. No result changes.
+  Affected earlier versions (the same texts): the versions for Sionna 0.19 and 1.x, and
+  2.0.0a1.
+
+## [2.0.0a1] - 2026-10-07
+
+First pre-release of OpenNTN for Sionna 2.x, which is built on PyTorch.
+
+- Import name `openntn` (`import openntn`), instead of the module `tr38811` that the
+  installation scripts of the earlier versions linked into Sionna.
+- Installation with pip from PyPI (`pip install openntn`) instead of `install.sh`.
+- Sionna 2.x is required (`sionna>=2.0.1,<3.0`), with Python 3.11 or later.
+- The licence is declared as `MIT AND Apache-2.0`: files derived from Sionna are under
+  the Apache License 2.0, all other files under the MIT License. Both licence texts are
+  included.
+- The unused model data files were removed.