Recent investigations have considered whether it is possible to achiev
e early detection of greenhouse-gas-induced climate change by observin
g changes in ocean variables. In this study we use model data to asses
s some of the uncertainties involved in estimating when we could expec
t to detect ocean greenhouse warming signals. We distinguish between d
etection periods and detection times. As defined here, detection perio
d is the length of a climate time series required in order to detect,
at some prescribed significance level, a given linear trend in the pre
sence of the natural climate variability. Detection period is defined
in model years and is independent of reference time and the real time
evolution of the signal. Detection time is computed for an actual time
-evolving signal from a greenhouse warming experiment and depends on t
he experiment's start date. Two sources of uncertainty are considered:
those associated with the level of natural variability or noise, and
those associated with the time-evolving signals. We analyze the ocean
signal and noise for spatially averaged ocean circulation indices such
as heat and fresh water fluxes, rate of deep water formation, salinit
y, temperature, transport of mass, and ice volume. The signals for the
se quantities are taken from recent time-dependent greenhouse warming
experiments performed by the Max Planck Institute for Meteorology in H
amburg with a coupled ocean-atmosphere general circulation model. The
time-dependent greenhouse gas increase in these experiments was specif
ied in accordance with scenario A of the Intergovernmental Panel on Cl
imate Change. The natural variability noise is derived from a 300-year
control run performed with the same coupled atmosphere-ocean model an
d from two long (>3000 years) stochastic forcing experiments in which
an uncoupled ocean model was forced by white noise surface flux variat
ions. In the first experiment the stochastic forcing was restricted to
the fresh water fluxes, while in the second experiment the ocean mode
l was additionally forced by variations in wind stress and heat fluxes
. The mean states and ocean variability are very different in the thre
e natural variability integrations. A suite of greenhouse warming simu
lations with identical forcing but different initial conditions reveal
s that the signal estimated from these experiments may evolve in notic
eably different ways for some ocean variables. The combined signal and
noise uncertainties translate into large uncertainties in estimates o
f detection time. Nevertheless, we find that ocean variables that are
highly sensitive indicators of surface conditions, such as convective
overturning in the North Atlantic, have shorter signal detection times
(35-65 years)than deep-ocean indicators (greater than or equal to 100
years). We investigate also whether the use of a multivariate detecti
on vector increases the probability of early detection. We find that t
his can yield detection times of 35-60 years (relative to a 1985 refer
ence date) if signal and noise are projected onto a common ''fingerpri
nt'' which describes the expected signal direction. Optimization of th
e signal-to-noise ratio by (spatial) rotation of the fingerprint in th
e direction of low-noise components of the stochastic forcing experime
nts noticeably reduces the detection time (to 10-45 years). However, r
otation in space alone does not guarantee an improvement of the signal
-to-noise ratio for a time-dependent signal. This requires an ''optima
l fingerprint'' strategy in which the detection pattern (fingerprint)
is rotated in both space and time.