I noticed a potential issue with the spin handling in the SIMPLE_EE Jastrow implementation.
In jastrow_ee, nspins is used to separate r_ee into parallel-spin and antiparallel-spin pairs:
r_ees = [
jnp.split(r, nspins[0:1], axis=1)
for r in jnp.split(r_ee, nspins[0:1], axis=0)
]
The code then applies different cusp coefficients in 2D:
cusp_parallel = interaction_strength / 3
cusp_anti = interaction_strength
This suggests that the Jastrow is intended to distinguish between same-spin and opposite-spin electron pairs.
However, when the Jastrow is applied in networks.py, the code uses
nspins_merged = (n_total, 0)
instead of the actual spin counts.
As a result, jastrow_ee interprets all electrons as belonging to the same spin sector. Therefore, all electron-electron pairs are treated as parallel-spin pairs, and the antiparallel-spin contribution is never included.
For a system containing both spin species, this appears to give
[
U =
\sum{i<j}
-\frac{(1/3)\alpha^2}{\alpha+r{ij}},
]
rather than
[
U =
\sum_{\sigma_i=\sigma_j}
-\frac{(1/3)\alpha^2}{\alpha+r_{ij}}
+
\sum_{\sigma_i\neq\sigma_j}
-\frac{\alpha^2}{\alpha+r_{ij}}.
]
Could you clarify whether using
nspins_merged = (n_total, 0)
is intentional? If the purpose of the 2D Jastrow is to satisfy the spin-dependent electron-electron Coulomb cusp conditions, it seems that the actual (n_up, n_down) should instead be passed to jastrow_apply.
I noticed a potential issue with the spin handling in the SIMPLE_EE Jastrow implementation.
In jastrow_ee, nspins is used to separate r_ee into parallel-spin and antiparallel-spin pairs:
r_ees = [
jnp.split(r, nspins[0:1], axis=1)
for r in jnp.split(r_ee, nspins[0:1], axis=0)
]
The code then applies different cusp coefficients in 2D:
cusp_parallel = interaction_strength / 3
cusp_anti = interaction_strength
This suggests that the Jastrow is intended to distinguish between same-spin and opposite-spin electron pairs.
However, when the Jastrow is applied in networks.py, the code uses
nspins_merged = (n_total, 0)
instead of the actual spin counts.
As a result, jastrow_ee interprets all electrons as belonging to the same spin sector. Therefore, all electron-electron pairs are treated as parallel-spin pairs, and the antiparallel-spin contribution is never included.
For a system containing both spin species, this appears to give
[
U =
\sum{i<j}
-\frac{(1/3)\alpha^2}{\alpha+r{ij}},
]
rather than
[
U =
\sum_{\sigma_i=\sigma_j}
-\frac{(1/3)\alpha^2}{\alpha+r_{ij}}
+
\sum_{\sigma_i\neq\sigma_j}
-\frac{\alpha^2}{\alpha+r_{ij}}.
]
Could you clarify whether using
nspins_merged = (n_total, 0)
is intentional? If the purpose of the 2D Jastrow is to satisfy the spin-dependent electron-electron Coulomb cusp conditions, it seems that the actual (n_up, n_down) should instead be passed to jastrow_apply.