[aoymt]

Summary

Four defects in the measurement made density.dat report something other than the observables of the input, or nothing at all. None of them printed an error.

Input	Before	After
No one-site observable	segmentation fault after the measurement	density.dat is written
One-site observable on a part of the sites	nan, and Energy = nan when it is the site Hamiltonian	sum over the sites that have it
Several terms on the same site or bond	only the last term is reported	the terms are summed
Two-site observable with |dx| > 3 or |dy| > 3 (bosons)	left out with a warning, the run finishes normally	input error

One commit per defect.

Changes

Write density.dat when the input has no one-site observable

save_density took the width of the name column from onesite_operator_names[0]. With no one-site observable the vector is empty, and a ground-state calculation died after everything else had been measured and saved. A std.toml with bond Hamiltonians only and no [observable] section is enough.

The width is now taken from the first kind of operator that has any, and is that of Energy when there is none.

quasi1d_ctm and onesite_rdm_output carried a one-site observable they did not use, to avoid this crash. They run without it now.

Leave the sites without an operator out of the density

measure_onesite fills the sites on which an observable is not defined with NaN, and save_density summed over all sites. When the site Hamiltonian itself was given on a part of the sites, as in a unit cell that mixes kinds of sites, Energy was nan although every value in onesite_obs.dat and twosite_obs.dat was right.

• The sites are told apart by site_ops_indices, not by isnan, so a NaN that comes out of a measurement still shows.
• The sum is divided by the number of all sites, not of the sites that have the operator. This is what the two-site observables already are, and it is what makes Energy the energy per site.

Add up the observables of one group on the same site or bond

The measurement stored a value by assignment, keyed by the group and the site or bond. tenes_std writes every [[hamiltonian]] as an observable of group 0, so a Hamiltonian written as a sum of terms on one bond (Ising part and XY part, or Heisenberg and Dzyaloshinskii-Moriya) was evolved as given and its Energy missed all the terms but the last. The same went for two terms on one site.

They are summed now, for one-site, two-site and multi-site observables, including the parity branches of the fermion mode.

A reference cannot be summed. ops of a two-site or multi-site observable, and correlation, name the one-site operator of a group on a site, and site_ops_indices keeps one index for it. An input that refers to a one-site observable given more than once on a site is now rejected.

Reject an observable that is too long to be measured

measure_twosite and measure_multisite contract a window of at most 4 x 4 sites. The simple update has no such limit, because tenes_std carries a longer bond through a chain of nearest-neighbour gates. A Hamiltonian with a bond of four lattice spacings was evolved in full and reported an Energy that missed the bond.

The fermion mode already refuses such an observable, when the input is read and again at the measurement, and tenes_std refuses the Hamiltonian bond. This PR does the same for bosons, in the same places.

• tenes: the new validate_measure_window() (src/iTPS/measure_window.hpp) is called after the fermion checks when the input is read, so that a calculation does not start, and again by measure(), measure_density(), measure_twosite() and measure_multisite(), for a state that was not made from an input file. The warnings are removed.
• tenes_std: rejects the observable after its fermion checks, and says when it comes from a [[hamiltonian]] bond.

The messages of the fermion mode do not change, since its checks run first. The manual already gave this range for observable.twosite and observable.multisite. None of the inputs under sample/ and test/data/ has such an observable.

Behavior changes

• Bosons: a two-site observable with |dx| > 3 or |dy| > 3, or a multi-site observable that does not fit into 4 x 4 sites, is an input error. It used to be a warning.
• tenes_std, bosons: a [[hamiltonian]] bond out of the window is rejected when the automatic bond_hamiltonian would hold it. It is still accepted and evolved when the input gives an energy observable of group 0 of its own, since then nothing out of the window is measured.
• An ops-form observable or a correlation function that refers to a one-site observable given more than once on a site is an input error.
• The value in density.dat of a one-site observable given on a part of the sites is the sum divided by the number of all sites.

Note for reviewers: an existing test is changed

test_bosonic_bond_outside_the_window_is_accepted in test/python/test_fermion_chain.py pinned that tenes_std accepts a bosonic bond out of the window, with the user's own energy observable and with the automatic one. The second case is what this PR changes, so the test is now two:

• test_bosonic_bond_outside_the_window_is_evolved: with the user's own observable, accepted as before.
• test_bosonic_bond_outside_the_window_cannot_be_the_energy: with the automatic one, rejected.

If the acceptance in the second case was intended, the last commit is the one to drop or to turn into a warning from tenes_std.

Documentation

• output_format.rst: what the value in density.dat is for an observable on a part of the sites.
• observable_section.rst: observables of the same group are summed; an observable out of range is an input error.
• standard_format.rst: terms on the same site or bond are summed; the range of a Hamiltonian bond.
• NEWS.md.

English and Japanese.

Tests

• New: observable_density, an end-to-end test with five cases, one or two per commit. Every case is a ferromagnetic Heisenberg chain whose exact ground state is the fully polarized product state, so every expected value is known in closed form. All five cases fail with a build of develop.
• New: six cases in test/python/test_tenes_std.py for the range check of tenes_std.
• ctest: 78/78 (Apple clang, no MPI, Release).

Not tested:

• an MPI build,
• the check at the measurement on a state built without an input file (it calls the same function as the check at read time).

🤖 Generated with Claude Code

[yomichi]

Thanks for this. I checked the four fixes, and they behave as described. Below are a few requests, mainly about the tests and the documentation.

1. terms_on_the_same_site_and_bond cannot tell a sum from "the first term only" on the bond.
In the fully polarized state = 0, so with the bond split into ISING + XY, both a correct sum and an implementation that keeps only the first term give -0.25. I replaced the += in twosite_obs.cpp with emplace (first term wins) and all five cases still passed. The test also passes on develop if the two terms are given in the other order. Could you split the bond into two terms whose values are both nonzero, as is done for the field with field(0.25) + field(0.75)?

2. (Optional) Multisite and fermion sums are not covered.
The += in multisite_obs.cpp and in the parity branches of the fermion mode (add_local_obs) are not reached by any test. A multisite case in observable_density would cover the first one cheaply.

3. standard_format.rst: "a bond out of this range is an error" is too strong.
As the PR description says, and as test_bosonic_bond_outside_the_window_is_evolved pins, tenes_std accepts such a bond when the input gives its own group-0 two-site observable. Could the sentence say that the bond is an error when it is measured as the automatic bond_hamiltonian?

4. output_format.rst: "divided by the number of all the sites in the unit cell".
save_density counts only the sites with physical_dims > 1, and simple_format.rst says "per non-vacancy site" for the kagome lattice. Could this say "the number of sites in the unit cell, not counting vacancies"? Same for the Japanese version.

5. (Optional) Nothing pins that the check runs at read time.
If I remove validate_measure_window() from main.cpp, observable_out_of_range still passes, because measure() raises the same error, only after the whole calculation. The check of multisite observables in C++ is not covered either; only the one in tenes_std is. A case that checks that no output/ files are written (or that the run fails before the simple update) would pin the first.

Not tested: an MPI build.

[aoymt]

Thank you for the careful review. All five points are addressed in 5843479 (tests) and 67b87cf (documentation).

1. The bond is now a quarter and three quarters of the Heisenberg term, as the field already was. I had only checked the order Ising, XY against develop.
2. multisite_terms_on_the_same_sites covers the sum in multisite_obs.cpp. The parity branches of the fermion mode are still not covered; that needs a fermionic input of its own, and I left it out of this PR.
3. standard_format.rst now says the bond is an error when it is measured as the automatic bond_hamiltonian, and that it is accepted with a group-0 observable of the user's own.
4. output_format.rst and NEWS.md now say "not counting vacancies". English and Japanese.
5. tenes makes the output directory after reading the input and before anything else, so the two out-of-range cases now require that it is not there. multisite_observable_out_of_range covers the multi-site check in C++.

Each new check fails when the code is broken accordingly: first term wins and last term wins in twosite_obs.cpp, last term wins in multisite_obs.cpp, and no check in main.cpp.

ctest: 78/78 (Apple clang, no MPI, Release). An MPI build is still not tested.

[yomichi]

Thanks, all five points are addressed. I checked 67b87cf with mutations: first term wins and last term wins in twosite_obs.cpp, last term wins in multisite_obs.cpp, no check in main.cpp, and no multi-site check in validate_measure_window() each make observable_density fail, and the unmodified code passes. Leaving the fermion parity branches to a later PR is fine.