Palm distribution
A Palm distribution describes a stationary point process from a typical event placed at the origin rather than from a typical time or location.
For finite positive intensity, event-weighted recentering defines the Palm law. In continuous time, saying 'condition on an event at zero' is an interpretation, not ordinary conditioning on a positive-probability event. The ordinary Palm law retains that event; a reduced Palm law removes it. [1]
For a stationary ergodic marked event stream, event averages give the event-centered law. Batches must retain their multiplicity: selecting a nonempty timestamp is not generally selecting an individual event. Palm inversion recovers the time law using the evolution between events, not only the state at event times. [2]
In Hesela's constructed ten-slot cycle, the high state occupies 5/10 slots, 5/6 nonempty timestamps and 9/10 arrivals. These are exact synthetic fractions, not storage measurements. The state is an external label, not inferred queue occupancy. [3]
PASTA is a separate result about pre-arrival state observations with its own assumptions; naming a Palm distribution does not establish Poisson input or equality with time averages. [4]
Relations
Citations
- Last and Penrose, Lectures on the Poisson Process (author version, 21 August 2017), chapter 9, Theorem 9.1 and sections 9.3-9.4
- Sigman and Whitt (2019), Marked point processes in discrete time, sections 4.3-4.4 and 6.1; DOI 10.1007/s11134-019-09612-3
- Hesela: exact event, timestamp and time weighting example
- Hesela: PASTA assumptions and pre-arrival observation boundary
Sources & caveats
- Mathematical source synthesis; no empirical population or replicated device experiment.
- Sigman and Whitt citations refer to the final 2019 version, not the 2018 draft.
- Single-trace convergence requires suitable ergodicity assumptions; stationarity alone is not a finite-sample accuracy guarantee.