Top-down solar modulation of climate: evidence for centennial-scale change

During the descent into the recent ‘exceptionally’ low solar minimum, observations have revealed a larger change in solar UV emissions than seen at the same phase of previous solar cycles. This is particularly true at wavelengths responsible for stratospheric ozone production and heating. This implies that ‘top-down’ solar modulation could be a larger factor in long-term tropospheric change than previously believed, many climate models allowing only for the ‘bottom-up’ effect of the less-variable visible and infrared solar emissions. We present evidence for long-term drift in solar UV irradiance, which is not found in its commonly used proxies. In addition, we find that both stratospheric and tropospheric winds and temperatures show stronger regional variations with those solar indices that do show long-term trends. A top-down climate effect that shows long-term drift (and may also be out of phase with the bottom-up solar forcing) would change the spatial response patterns and would mean that climate-chemistry models that have sufficient resolution in the stratosphere would become very important for making accurate regional/seasonal climate predictions. Our results also provide a potential explanation of persistent palaeoclimate results showing solar influence on regional or local climate indicators.


Introduction
Several solar parameters and outputs during the recent sunspot minimum were lower than for the previous minima during the space age, indeed lower that at any time since about 1920 [1]. The decline in UV emission during the descent into this exceptional solar minimum was monitored by the SIM instrument on the SORCE satellite and revealed a larger change [2] than seen by different instruments at the corresponding phase of previous solar cycles [3,4], particularly at the wavelengths responsible for stratospheric ozone production and heating [5] with consequences for the patterns of winds and temperatures throughout the stratosphere [5,6]. However, there is debate about the calibration of the recent data and hence whether UV irradiances in the recent solar minimum were really lower [7].
There is growing evidence that dynamical coupling across the tropopause means that stratospheric changes can influence the underlying troposphere [6,8] and under some circumstances, robust tropospheric responses are indeed predicted by models. Tropospheric jet streams have been predicted to be sensitive to the solar forcing of the stratosphere [9,10]. This could occur through disturbances to the stratospheric polar vortex [11] which are observed to propagate downwards to affect the tropospheric jets [12]. (However, it should be noted that the fact that the disturbances appear first in the stratosphere does not necessarily mean that the stratosphere is driving the troposphere [13].) Alternatively, solarinduced stratospheric changes may influence the refraction of tropospheric eddies [14,15]. Models (e.g., [16]) do predict that perturbations can descend from the stratosphere to the surface by altering the propagation of planetary waves coming up from the surface, an effect that has been reported in observations [17]. Definitive identification of these 'topdown' solar influences on the troposphere is difficult; however, models show that the stratosphere has the potential to play a crucial role in regional climates. For example, Scaife et al [18] have demonstrated that stratospheric trends over recent decades, along with downward links to the surface, are indeed strong enough to explain much of the prominent trend in the North Atlantic Oscillation (NAO) between the 1960s and the 1990s, with implications for regional climate in Europe, particularly in winter. Effects have also been identified in the southern hemisphere [19].
These 'top-down' mechanisms would be effective alongside 'bottom-up' solar heating of the sea surface and the dynamically coupled air-sea interactions [20]. Although differentiating between the effects of variations in the two will often be difficult, recent studies indicate that they are additive, producing amplified SST, precipitation and cloud responses, for example in the tropical Pacific, even for relatively small solar forcing changes [21,22]. The solar output driving the bottom-up effects is the total solar irradiance (TSI) whereas for top-down forcing the spectral solar irradiance in the UV band is relevant. This paper investigates UV spectral variability. The effects on climate of solar variations have been shown to be swamped by anthropogenic effects in recent decades on global scales (e.g., [23]), but they have also been found to be significant in certain regions and seasons [24][25][26]. Significant differences between mean global mean air surface temperature for composite surface temperature maps at low and high sunspot activity have been found [27]. Such a signal can also be seen in both the ECMWF ERA40 and NCEP/NCAR re-analysis datasets, in which observations have been assimilated with model data [28]. This signal is weaker in the ERA40 dataset and varies regionally [27,29]. We here present evidence for long-term drift in UV irradiance, which is not found in its commonly used proxies, and that both stratospheric and tropospheric winds and temperatures show stronger regional responses to those solar indices that show long-term trends.

Solar UV irradiance
The 10.7 cm solar radio flux, F10.7, originates from about h S = 2100 km above the visible solar surface at the large temperature gradient: this is associated with the strong emission of Lyman-α at 120 nm [30] which produces a temperature plateau that is eroded by ambipolar diffusion. The temporal variation of F10.7 is closely mirrored by the core-to-wing ratio indices of absorption lines, such as Ca-II (wavelength λ ≈ 393 nm) and Mg-II (280 nm) where the wing emission, like the rest of the UV continuum, arises at 0 < h S < 300 km but the absorption determining the line core takes place at about 900 < h S < 2000 km. Thus the indices commonly used to quantify solar UV emission (Ly-α [30], Ca-II [31], Mg-II [32], and F10.7 [30]) are all determined by the upper chromosphere, whereas the continuum UV emission arises from below the minimum of the temperature profile, i.e. below h S ≈ 400 km for λ > 160 nm. The correlations between a given UV spectral band and these indices vary and depend on timescale [33]. Hence good agreement on one timescale does not preclude differences on another. Thus the fact that the commonly used proxies for solar UV emission reproduce the solar cycle and solar rotation variations in the UV emission rather well does not necessarily mean that they will also capture its centennial-scale drift. The large difference between the observed UV emission during recent solar minimum [2] and during previous minima [3,4] can have one, or a combination, of three causes: (a) the new data are in error [7]; (b) the previous data were in error or (c) there has been a real change in the emission, associated with the recent exceptionally low solar minimum. We here investigate possibility (c): because we are concerned with the variation and not the absolute levels, (a) and (b) can be considered together by looking at the intercalibration of the new and the previous data.
We employ the UV spectral irradiance (S) composite (as a function of time t and wavelength λ) by DeLand and Cebula [4] and add the more recent SORCE data [2]. The composite was compiled using absolute calibration but at some λ this leads to discontinuities in the temporal sequences at the joins between the data from different sources. We here use monthly means for each λ (1 nm apart) and make zero-level adjustments at each join. To do this, we employ the high correlations (r SP ) on short (1 year) timescales between S and the Ca-II, Mg-II and F10.7 proxies. Figure 1 demonstrates the procedure for a particularly clear example. For every known change in the source of the data ('join'), monthly means of S (at the λ in question) were taken for one year before and after the join. The data for one year after the join (filled black circles) was compared to that for one year before it (filled grey squares). A least-squares linear regression between the S data after the join and the proxy (F10.7 in the example shown in figure 1) was generated and is shown by the solid line. The best-fit zero-level offset for the data from before the join, S, was then derived such that, when corrected by S (open triangles), the mean square deviation from the regression line is minimized. Note the correction is applied to all data before the join, so that all corrected data become relative to the most recent data. The offset used is the . λ = 121.5nm. date: 30/6/1988 Figure 1. Illustration of the procedure used to intercalibrate data in the composite UV data series. This example is for a wavelength λ = 121.5 nm and uses the relationship to the F10.7 proxy on timescales of order 1 year. For every known join between different data sources (this one is on 30th June, 1988), the monthly means for one year before the join (grey filled squares) were compared to the monthly means for one year after the join (black filled circles). Only corrections for zero-level offsets are made and it is assumed that instrument gain factors are constant. The data from after the join were regressed against the proxy (solid line) and the data from before the join were corrected (giving the open triangles) by a constant offset S which gives the best least-squares fit to the solid line.
best of those derived from the three proxies, as evaluated using the rms deviation from the regression line (for the data after the join plus those before it after correction: i.e., the solid points plus the open triangles in figure 1). Useable corrections were obtained everywhere except for before 1986 at λ > 270 nm: these data were discarded. No assumptions were made about the shape of the spectrum and so no corrections were applied in the λ dimension. Figure 2(b) shows the minimum (in blue) and maximum (in red) of monthly means of S as a function of λ for the interval covered by the composite . Note that these spectra are given on a logarithmic S scale. It can be seen that the solar cycle variation is a very small fraction of the mean level, especially at the larger λ. Part (c) of figure 2 shows normalized variations of S by colour contouring (S − S )/σ S , where S and σ S are the mean and standard deviation of S for the full interval and the λ in question, as a function of time t and wavelength λ. This plot reveals clear solar cycle signals (compare with F10.7 in figure 2(a)) at most λ. It can be seen that the calibration procedure yields some discontinuities in the λ dimension which are likely to be instrumental artefacts. In this paper we integrate the spectra into three bands to average out such errors.
However, there are concerns that the considerably lower values of S seen at some λ during the recent solar minimum, as shown in figure 2, are an instrumental effect: this is because these values all come from the one instrument (SIM on SORCE) and are inconsistent with previously successful models of UV irradiance [7]. To investigate this, we here also consider a change in both the instrument zero-level offset and its sensitivity (gain) between SIM and the previous instruments. We can do this because the SIM instrument was operated at the same time as SUSIM on UARS over the interval May 2003-July 2005. Because this was at a time when S was falling rapidly after the solar maximum, this overlap can be used to evaluate both the gain and offset of SIM compared to SUSIM, and hence the data composite. Figure 3 shows one such a comparison (this example is for a wavelength of λ = 179 nm). The blue curve in the left hand plot gives the monthly means of the SUSIM data, the red curve is the monthly means of the SIM data. The right hand panel in figure 3 is a scatter plot for the interval where these sequences overlap: the correlation coefficient is very high (r = 0.93) but the slope of the regression is not unity, implying SIM may indeed have a different gain factor than SUSIM. Using the bestfit least-squares ordinary regression, the SIM data can be made consistent with the SUSIM data, giving the green curve in the left hand plot. It can be seen that the recent solar minimum is not as deep for the re-calibrated green curve as it was for the uncalibrated red curve (but note that for this λ it is still slightly deeper than for the previous minimum seen in the SUSIM data). For some λ this calibration makes the recent minimum even deeper, but most cases are as in figure 3.
Parts (a)-(c) of figure 4 show the variations of monthly means of the average S over λ bands of 120-200 nm, 200-270 nm and 270-400 nm (referred to here as S FUV , S MUV and S NUV , respectively). Consider first the black curves, which are taken from figure 2 directly-i.e. the irradiance data have been corrected only for zero-level offset changes (using the procedure demonstrated by figure 1). Figure 4(a) shows that the three minima in S FUV show a consistent but very slight decline, whereas S MUV shows a slight decline between the first two and then a much lower value in the third ( figure 4(b)). A smaller drop (relative to the solar cycle amplitude) is seen for the only two minima in the useable S NUV data ( figure 4(c)). On the other hand, if we first adjust the SIM data for each wavelength using the linear regressions of the type shown in figure 3 (i.e., we correct for both instrument gain change and zero-level offset at this particular data join) we obtain the mauve lines (which prior to the start of the SIM data are identical the black line). In all three bands the composite data now show a slight decline from one minimum to the next, but the third minimum is less deep than for the black line in all three cases (considerably so for S MUV ).
Over the same interval as covered by these S FUV , S MUV and S NUV composite sequences, the solar-minimum Total Solar Irradiance (TSI, absolute values of which are dominated by the visible and near IR emissions but the solar cycle variation of which is dominated by the UV [34,35]) has been reported to have fallen to slightly lower values [1,36], suggesting a slight long-term trend mirroring that in F S , the open solar flux [37][38][39] and consistent with reconstructions of the TSI variation over past centuries [40,41]. The inferred long-term variations in F10. 7   and the black line shows an extrapolation using the highly correlated observations of sunspot numbers, R (correlation coefficient in annual means r = 0.96). The Ca-II lines are attenuated in the atmosphere but, unlike the Mg-II lines, can be observed from the ground: historic measurements have been reported as showing a variation that very closely matches that shown in figure 5(a) [42]; however, there is debate about the calibration of these data [31,43]. Figure 5 [37] and the black line shows annual means derived from geomagnetic activity indices [38,39]. There is a major difference between the variations shown in figure 5 in that F10.7 (and Ca-II) show repeated returns to almost the same level at each solar minimum whereas the heliospheric indices F S and (and hence cosmic ray fluxes and cosmogenic isotope abundances) show centennial drift between solar minima.
The lower panels of figure 4 show the best ordinary linear regression fits of monthly means of F10.7, F S , and cosmic ray fluxes (we here use the McMurdo count rate M shown in orange in figure 5(b)) to the S FUV , S MUV and S NUV composites. We include the gain correction to the SIM data. The legend gives the correlation coefficients r and their significance levels. The good agreement on short timescales means that r for F10.7 is high in all 3 bands but the inferred long-term trends in S MUV and S NUV are better fitted by M and F S . Note that correlations for F S are consistently lower because the data sequence is much noisier on these monthly timescales.
Chromospheric UV emission is closely linked to the local magnetic field [46] rather than that at the top of the solar corona (i.e., the open solar flux, F S ). Because inward and outward magnetic field within pixels cancel, solar magnetographs reveal logarithmically greater field strengths (in the photosphere, just below the chromosphere) as resolution is increased [47,48]. Much of this 'hidden' magnetic field seems not to show a solar cycle dependence, and so that part is thought to arise from turbulent motions in the quiet Sun: however, a second component implies an additional cascade down in scale sizes from the larger flux tubes emerging in sunspot groups [49,50]. We note that the extent to which this hidden flux influences UV emissions also remains controversial; however, the time constant for this cascade appears to be of order 1-2 yrs which, coincidentally, is similar to the larger of two time constants for the magnetic flux in sunspot groups to evolve into F S [51]. This offers one potential explanation of why heliospheric indices such as F S , , M and cosmogenic isotopes might be useful proxies for UV emission even though they relate to the solar magnetic field in a quite different locations (respectively, the top and base of the solar atmosphere): a correlation could arise because both UV emission and open solar flux would be enhanced by increases in sunspot activity but they decay away with similar time constants.

Atmospheric response
The indication that key ozone-producing and heating wavelengths in the MUV may show long-term trends, similar to those in heliospheric proxies such as F S , M and , depends critically on the intercalibration of the various radiometric records which remains highly uncertain. In this section, we take a different approach and search for the effects in Earth's atmosphere. We use the European Centre for Medium-Range Weather Forecasting (ECMWF) re-analysis dataset for 1979-2002, updated to 2008 using ECMWF operational analysis. A standard multiple linear regression is performed on annual means to separate the atmospheric dataset into contributions from various climate indices. This analysis is exactly as used in the papers by Crook and Gray [52] and Frame and Gray [53] and the reader is referred to these papers for full details. We employ a constant term, a linear trend, a stratospheric optical depth index, a multivariate ENSO index, two orthogonal quasibiennial oscillation indices, and a solar variability proxy (one of F10.7, F S , TSI, and M). The results are shown in figure 6 for zonal means of temperature T (top) and of zonal wind U (bottom). Results for TSI (not shown) are quite similar in form, but weaker in amplitude, to those for F10.7. The solar effect on equatorial lower stratospheric temperatures (previously found using F10.7 [11,52,53]) can be seen in figure 6(a) and peaks at k/σ = 0.3 K σ −1 , where k is the regression slope and σ is the standard deviation of the solar parameter. (Note that normalizing k by σ allows for the different units and ranges of the various solar indices so that we can compare the amplitudes of the responses). Figures 6(b) and (c) show stronger effects peaking at 0.5 K σ −1 for M and F S (with greater significance and over a larger region). Modelling [9] and data analysis [10] has indicated that the solar effect on T in the lower equatorial stratosphere can influence the jet streams in the underlying troposphere. This can be seen in figure 6(d) (for F10.7) where the effect on U in the mid-latitude troposphere peaks at about k/σ = −0.3 m s −1 σ −1 . However, the effect is considerably stronger for F S (k/σ peaking at −0.8 m s −1 σ −1 in figure 6(f)) and the statistical significance is greater and over a wider region. The additional response for F S (figure 6(f)) is mainly in the northern hemisphere, being somewhat more asymmetric than that for M (figure 6(e)). Figure 6(c) also shows a small band of significant surface temperature response at middle northern latitudes where recent studies have revealed solar signatures in regional climate, particularly in the European sector [24][25][26].
We have checked that the larger responses seen in figure 6 for F S and M do not arise from an interplay with the linear trend. Plots equivalent to figure 6 for the best-fit trend term using all three solar indices are virtually identical. (We do not reproduce those plots here as the trend magnitudes are not thought to be reliable because of the way the ERA reanalysis dataset is constructed.) For the analysis using F10.7, Frame and Gray [53] investigated how the percentage variances were apportioned between the different terms and the errors introduced (e.g., by small changes in the data period employed) and their results are not significantly changed by the use of the other solar indices. Also, the regression analysis included an AR3 autoregressive noise model [53] and the residual was checked to ensure it is acceptably small, (i.e., that the residual is random noise and there is no significant remaining unexplained variations).

Discussion and conclusions
The work presented here is consistent with the interpretation of a recently reported effect [25] of solar variability on the North Atlantic Oscillation (NAO) and European winter temperatures over the interval 1659-2010 in terms of top-down modulation of the blocking phenomenon [52,53]. In fact, Woollings et al [26] show that the solar response pattern is, despite being similar in form to that of the NAO, significantly different in that it reaches further east. These authors also show that open solar flux has a much stronger control over blocking events in this sector than the previously reported effect of F10.7 [55].
There is seasonality in the solar responses reported here. This is expected as modulation of upwards-propagating planetary waves in wintertime, and the associated stratospheretroposphere interaction, is most widely believed to be the key mechanism [8,11]. In addition, the tropospheric signature is a response of the eddy-driven jet streams, and these are at their strongest and most responsive in winter. While the results are presented here as annual means, the regression analysis was actually carried out on monthly mean data and thus takes this seasonality into account. The seasonal evolution of the F10.7 cm flux regression was described in detail by Frame and Gray [53] and this was not significantly affected by using either the open solar flux F S nor the cosmic ray flux, M, instead of F10.7.
The relationship between the open solar flux F S and UV spectral irradiance changes inferred here provides an explanation of why the open solar flux was a valuable proxy for solar activity in the recent studies by Lockwood et al [25] and Woollings et al [26]. In addition, F S is correlated with other parameters such as the level of geomagnetic activity and solar wind dynamic pressure and a relationship with UV variability may be responsible for persistent reports of a relationship between these phenomena and the NAO [56][57][58][59]. This is a much more satisfactory explanation than any causal link, for which the unexplained amplification required would be massive (the energy density of the solar wind being smaller than the TSI by a factor of order 10 5 and the currents that it generates, and that cause geomagnetic activity and heating, are in the upper atmosphere where the mass is of order 10 −3 of the total atmospheric mass).
Furthermore, F S is highly anticorrelated with cosmic ray fluxes reaching Earth on a wide range of timescales [60]. This makes any relationship of F S with UV emissions and top-down solar forcing a relevant factor in the considerable debate about reported correlations between climate (in particular global or regional cloud cover) and GCRs. Indeed some authors have noted that apparent solar cycle variations in cloud cover are in better agreement with the UV irradiance variation than that of GCRs [61][62][63], although we note that other studies that have tried to discriminate between electromagnetic irradiance and direct cosmic ray effects do find some evidence for the latter [64,65]. A great many palaeoclimate studies have found links between regional or local climate indicators and cosmogenic isotopes and so it has been argued that this is either evidence for a direct cosmic ray effect or that the cosmic rays are proxy indicators for the correlated irradiance variability [66]. In particular, the much cited and much debated paper by Bond et al [67] revealed a persistent correlation during the Holocene between ice-rafted debris in the North Atlantic region and cosmogenic isotopes. The results presented here and in the recent paper by Woollings et al [26] suggest that solar UV variability and top-down solar forcing can introduce regional changes in the troposphere in the region studied by Bond et al and that these may vary on centennial timescales, as recently reported in the Central England temperature records by Lockwood et al [25]. The implication is that seasonal/regional evaluation of past and future climate change will be improved by models with adequate resolution in the stratosphere to reproduce top-down solar forcing.