pyhelayers HeConfigRequirement.bootstrap_config does not work as documented

[tsunami935]

I am trying to use bootstrapping with OpenFheCKKS, but am unable to do so.

I have the following lines of code:

requirement = pyhelayers.HeConfigRequirement(
    num_slots = 2 ** 16, # Number of slots per ciphertext
    multiplication_depth = 13, # Allow 13 levels of multiplications
    fractional_part_precision = 24, # Set the precision to 1/2^24.
    integer_part_precision = 5, # Set the largest number to 2^5.
    security_level = 128
)
requirement.bootstrappable = True
bootstrap_config = pyhelayers.BootstrapConfig()
bootstrap_config.target_chain_index = 12

The documentation says that bootstrap_config is a property of HeConfigRequirement. I get the following error when I try to set the bootstrap configuration:

    requirement.bootstrap_config = bootstrap_config
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: (): incompatible function arguments. The following argument types are supported:
    1. (self: pyhelayers._pyhelayers_cppwrappers.HeConfigRequirement, arg0: std::optional<helayers::BootstrapConfig>) -> None

I then tried to call HeConfigRequirement.boostrap_config with the function signature given and get a type error:

TypeError: Unregistered type : std::optional<helayers::BootstrapConfig>

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "~/helayers/openfhe_boostrap.py", line 30, in <module>
    requirement.bootstrap_config(bootstrap_config)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Unable to convert function return value to a Python type! The signature was
        (self: pyhelayers._pyhelayers_cppwrappers.HeConfigRequirement) -> std::optional<helayers::BootstrapConfig>

Did you forget to `#include <pybind11/stl.h>`? Or <pybind11/complex.h>,
<pybind11/functional.h>, <pybind11/chrono.h>, etc. Some automatic
conversions are optional and require extra headers to be included
when compiling your pybind11 module.

Is bootstrapping currently supported in pyhelayers? The only demo I noticed with bootstrapping was the Game of Life C++ demo.

[aehud]

It seems like a bug. Bootstrap should work in pyhelayers.
Can you check the version number you are using (print(pyhelayers.VERSION))
And the python version?

[tsunami935]

I am using pyhelayers==1.5.5.3 and Python 3.11.15.

[aehud]

Ok, thanks.
We managed to reconstruct a bug, but it looks a bit different than what you reported.
This line works ok:

requirement.bootstrap_config = bootstrap_config

But an error similar to what you reported occurs later when we actually call bootstrap()

In a few days we'll publish a fix with some python wrappers improvements. We'll update here so you can check if it solves your problem as well.

[tsunami935]

Can you please let me know the timeline for the fixes? This issue with bootstrap is more urgent than the others that I've opened. Thank you.

[aehud]

Ok. It should be ready in a week or two. We'll try to push the bootstrap issue sooner, I'll update.

[aehud]

Update: we now plan to issue a fix to release 1.5.5 with just this issue in two weeks.
The other two issues will be solved in the next release 1.5.6 in about a month.

[aehud]

We released version 1.5.5.4 that is supposed to solve this issue. Can you check if it works for you now?

Note though that your configuration values are not feasible.
The following init worked for me:

requirement = pyhelayers.HeConfigRequirement(
    num_slots = 2 ** 16, # Number of slots per ciphertext
    multiplication_depth = 30, 
    fractional_part_precision = 52, 
    integer_part_precision = 5, 
    security_level = 128
)

requirement.bootstrappable = True
#bootstrap_config = pyhelayers.BootstrapConfig()
#bootstrap_config.target_chain_index = 5
he=pyhelayers.OpenFheCkksContext()
he.init(requirement)

The two commented out lines also run ok if uncommented, but the user provided target_chain_index is ignored, and is set up by the context automatically. You can obtain the bootstrap's start and end chain-indices using:

print(he)

The following code for bootstrapping now works for me:

enc=pyhelayers.Encoder(he)
v=np.array([1.0,2.0,3.0])
c=enc.encode_encrypt(v)
c.set_chain_index(2)
c.bootstrap()
print('after bootstrap CI=',c.get_chain_index())
print(c)

[tsunami935]

Yes, I was able to get bootstrap to work with the configuration you gave.

However, is there a way to set the target chain index for bootstrapping, and is there a valid bootstrappable configuration with integer part precision >= 10? Realistically, for my use case, I would like a target chain index >=12 and integer part precision >=10. Thank you.

[aehud]

Increasing the target chain index is easy, by increasing the multiplication depth. This configuration for example has a target chain index of 14:

requirement = pyhelayers.HeConfigRequirement(
    num_slots = 2 ** 16, 
    multiplication_depth = 31, 
    fractional_part_precision = 52, 
    integer_part_precision = 5,
    security_level = 128
)
requirement.bootstrappable = True

integer-part-precision plays a different role in a bootstrappable configuration. Usually, a bootstrap operation is only precise for values in the range [-1,1]. A larger integer_part_precision is used internally to improve precision, and not to support a larger range.

So the easiest solution for you is to check the precision you get for the values you bootstrap, and if it's not good precision, try scaling the values yourself to fit within [-1,1] before bootstrapping. You can try experimenting with various ranges, as the precision might be good also for [-2,2] or more, depending on the distribution.

There's an internal openfhe configuration flag, that helayers currently doesn't expose, called 'bootstrap correction factor', which as I understand is related to this. By default, it is chosen automatically, but if controlled manually it might allow for bootstrapping with a larger range (not sure), at the expense of a lower precision. In some cases you might get better performance (time or precision) by controlling it manually, but the default is supposed to be optimal for most use cases, and you can always normalize yourself as I explained above.