IIUC, if observations are multivariate normal (with possibly some additional hypotheses)-- which is consistent with many proposed
stationary stochastic processes (iid, AR (1), MA (1), ARMA (p, q), FARIMA (p, d, q), etc.)-- then the first and second differences will also be multivariate normal (note that the reverse is not generally true).
But unfortunately the MBH98 algorithm manages to find the hockey stick at the end of the signal notwithstanding it being
a stationary stochastic series.
From
a stationary stochastic process the MBH98 algorithm detects a non-stationary signal that is then attributed to CO2 forcing (the «hockey stick»).
«From
a stationary stochastic process the MBH98 algorithm detects a non-stationary signal: just how does it do that and still get described as robust?»
You are referring to a trend over half a century but we are talking about tests using
stationary stochastic simulations.
MM2005 model
a stationary stochastic process but the MBH98 algorithm somehow detects what you describe as a non-stationary forcing attributed to CO2.
The climate system is not
a stationary stochastic process.
Not exact matches
If one assumes that a
stochastic process is
stationary (Hurst's Hunit root test), which would be consistent with DrK's Hurst - Kolmogorov pragmaticity, then the process has a finite mean.
If we continue this type of thinking (means varying at a cascade of time scales, in an unpredictable manner), the eventual result is a
stationary (yes,
stationary)
stochastic process with LTP.
Of course any finite realization of a
stochastic process,
stationary or not, has a finite mean (you can always compute the average of a bunch of numbers).
Nevertheless, the salutary aspect of the GISP 2 data is the clear indication it provides of a gentle, truly secular cooling trend since the Holocene optimum, overlain by weakly
stationary, strongly structured, quasi-Gaussian
stochastic variations whose ordinate distribution and power - spectrum both diverge from anything resembling a Poisson process of abrupt jumps.
Forecasting confidence intervals for trend
stationary and
stochastic trend specification over 1885 - 2008, with parameter estimates based on 1881 - 1935 sample, here, with accompanying figures here.