Custom operators and observables

NetKet exposes several extension paths for custom observables and operators. They differ in how much you implement yourself, whether gradients come for free, and whether expect_to_precision() works automatically. All three rely on NetKet’s multiple-dispatch extension mechanism; if you need the background first, start with Overriding defaults in NetKet and in particular NetKet Architecture: Multiple Dispatch.

Use the table below as the entry point, then jump to the tutorial section that matches the interface you want to implement.

Path	What you implement	expect()	expect_and_grad()	expect_to_precision()	Best when	Start here
Scalar operator kernel interface	Subclass AbstractOperator, then define get_local_kernel_arguments() and get_local_kernel().	Automatic	Automatic	Automatic for scalar local estimators	The observable is a standard operator whose expectation is the mean of one local-estimator channel.	Example: the lean operator-kernel interface
Explicit local_estimators() dispatch	Register local_estimators() and return either LocalEstimators or LocalEstimatorsBatch.	Automatic	No	Automatic	You need multi-channel local estimators, delta-method error propagation, or expect_to_precision() without gradients.	Example: V-score custom operator
Explicit observable dispatch	Register expect() for your custom observable, and optionally expect_and_grad().	Yes	Only if you define it	No	You want full control over the computation, or the observable does not fit NetKet’s local-estimator interfaces.	Example: explicit observable dispatch

Recommended starting points

• If you need gradients and your observable is a standard scalar operator, start with Example: the lean operator-kernel interface.

• If you do not need gradients and you want expect_to_precision(), start with Example: V-score custom operator.

• If neither interface fits, fall back to Example: explicit observable dispatch and implement the exact expectation logic yourself.

• Defining custom expect and expect_and_grad
  • Example: explicit observable dispatch
  • Example: a fully custom expect implementation
  • Example: the lean operator-kernel interface
  • Comparison of the two approaches
• Defining custom local estimators (new)
  • What local_estimators() returns
  • Nonlinear observables and the delta method
  • Online accumulation with accumulate
  • Example: V-score custom operator
  • Example: using LocalEstimatorsBatch without MCState