Strong enhancements of the upper thermospheric total mass density were observed by the CHAMP satellite at approximately 400 km altitude during three geomagnetic superstorms occurring on 29–30 October 2003, 30–31 October 2003, and 20–22 November 2003. The corresponding density enhancements peaked around 400%, 500%, and 800% of the quiet-time values in both noon and midnight sectors. The disturbances showed strong noon–midnight and hemispheric/seasonal asymmetry. In the noon sector, the average density enhancement was stronger in summer (southern) than in the winter (northern) hemisphere. In the midnight sector, however, no general conclusion can be drawn about the seasonal effect. Stronger density enhancements occurred in the summer hemisphere during the second and third storm events, while in the winter hemisphere during the first storm event. The relative intensity of the disturbance between day and night is strongly dependent on our definition of the intensity. When expressed in absolute terms (storm-quiet), the density enhancement at night was generally less than half of that at day during all three storms. When expressed in percentage terms (), however, the enhancement at night was comparable to or even larger than that at day. The propagation of the disturbance from high to low latitudes during the November storm was faster in summer (southern) than in the winter (northern) hemisphere on both dayside and nightside. In the winter hemisphere, the propagation on the dayside was found to be faster than on the nightside. These variations of the seasonal/hemispheric asymmetry with local time and individual event imply that the many different competing processes involved in a magnetic storm may also vary dramatically with local time and from event to event. The MSIS90 model was unable to reproduce most of the observed features of these three storms.
 The thermospheric density and composition experience dramatic changes around the globe during magnetic storms [e.g., Taeusch et al., 1971; Prölss, 1980, 1981; Forbes et al., 1996]. The thermospheric response is determined by Joule/particle heating, Lorentz force ( × ), thermal expansion, upwelling, and horizontal wind circulation. Wave structures have been identified in the thermospheric disturbances [e.g., Reber et al., 1975; Trinks and Mayr, 1976; Williams et al., 1993; Forbes et al., 1995]. However, the propagation of the disturbance from high to low latitudes is controlled mainly by meridional wind circulation, rather than gravity waves [Johnson, 1960; Richmond, 1979]. Recent simulations by Burns et al.  showed strong dependence of the thermospheric response on the properties of the quiet-time thermosphere. They found that the storm-time enhancement of the thermospheric temperature and composition were larger in winter than in summer, even when they applied the same power input to both hemispheres. The authors attributed this to the relative importance of the magnetic storm-related heating with respect to the background solar heating. Storm-related heating, e.g., Joule heating, is more important in winter, where the solar heating is relatively low. It can therefore produce larger disturbances than in summer. A similar concept was employed to explain the seasonal effect of the equatorward propagation of the disturbance. Here, the direction and strength of the storm-driven wind with respect to the background solar-driven circulation play a dominant role in the seasonal effect of the thermospheric response to magnetic storms at middle and low latitudes [Prölss, 1980, 1981; Fuller-Rowell et al., 1994, 1996]. Simulation results from Fuller-Rowell et al. [1994, 1996] also suggested the density enhancement at high latitudes travels equatorward faster on the nightside than on the dayside.
 Strong differences in the quiet-time thermosphere exist not only between seasons but also between day and night. Within the framework mentioned above, we would expect a larger disturbance on the nightside than on the dayside during magnetic storms. Furthermore, the propagation of the density disturbance on the nightside should also be faster and more efficient than on the dayside. We test these hypotheses by examining the thermospheric density disturbances measured by the CHAMP satellite in the upper atmosphere at ∼400 km altitude. The CHAMP satellite has a near-circular polar orbit. It provided good time and latitudinal coverage during the three magnetic storms, which occurred on 29–30 October 2003, 30–31 October 2003, and 20–22 November 2003. We refer to them as the first, the second, and the third event in the following. The CHAMP orbit was in the noon-midnight sectors during these events (0130–1330 MLT for the first two storms and 2300–1100 MLT for the third one). This fortunate constellation enables us to investigate the day–night dependence of the storm effect (including seasonal/hemispheric dependence) and its variation with individual magnetic storms. Together with other papers in this special issue, the global thermospheric response presented below will help to form a better picture of the physical processes, which occur in the solar-terrestrial chain during these strong magnetic storms.
2.1. Geophysical Conditions
 All three storms are classified as superstorms according to the definition given by Gonzalez et al. . The minimum Dst was −363 nT and −401 nT during the first two storms and −473 nT during the third one. The time history of the interplanetary magnetic field (IMF) from the ACE spacecraft, the Dst, AE, and Kp indices are plotted in Figure 1. The two October storms are plotted together in all related figures. The first and third storms started with a storm sudden commencement (SSC) [Araki, 1994] at 0611 UT on 29 October 2003 and at 0803 UT on 20 November 2003, respectively. The minimum IMF Bz was −28 nT, −33 nT, and −50 nT in the main phase of the three storms. Despite the stronger southward IMF Bz component and the larger Dst drop during the third storm, the corresponding Kp index showed smaller values than that during the first two storms. As thermospheric density is strongly related to solar EUV radiation, the corresponding solar flux, F10.7, during these storms is plotted in Figure 2. The day before the storm is also shown for reference. The average F10.7 was ∼270 during the October events, while only ∼170 during the November event. In the following, we will concentrate on the comparison of the storm effects in different local time sectors (noon-midnight) and in different hemispheres (seasons). Detailed relations between the thermosphere response and its solar and geomagnetic drivers are not the focus of the present study and are therefore only briefly mentioned in the context.
2.2. Thermospheric Density Disturbance
 Thermospheric density was derived from Level 2 data of the three-axes accelerometer on board the CHAMP satellite. A detailed description of the method used to derive the density is given by Liu et al. . To investigate density variations due to magnetic storms, the measured density is adjusted to a common altitude of 400 km using scale height calculated from the MSIS90 model, thus correcting for density variations resulting from orbital altitude changes (within ±15 km during October and November 2003). The adjustment procedure is also described by Liu et al. . As we focus mainly on the comparison between different regions observed during the same day, a correction for the solar flux effect is not necessary. Thermospheric density predicted by MSIS90 model is presented as well. The model values were generated with the actual solar radio fluxes (F10.7) and magnetic activities (Ap index) during the storms. In the following description of these three storms, “summer” and “winter” hemispheres correspond to “southern” and “northern” hemispheres, respectively.
2.2.1. Events of 29–31 October Storm
 The thermospheric density at 400 km altitude during 29–31 October 2003 is presented in Figure 3. The top two panels show the CHAMP observations on the dayside and nightside, respectively. The lower two panels show the corresponding densities predicted by the MSIS90 model. This period was characterized by three phases whose Dst minimum occurred around 1000 UT, 2500 UT, and 4700 UT (see Figure 1). Pronounced density enhancements were seen in the CHAMP observations during all three phases around 0700 UT, 2200 UT, and 4600 UT. The density was more than tripled in some regions, reaching values above 30 × 10−12 kg m−3, in comparison with 8 × 10−12 kg m−3 during quiet times. Furthermore, the density enhancements occurred earlier at high latitudes than at lower latitudes. The density variation during the first storm (∼0600–3600 UT) showed clearly the propagation of the disturbance from high to low latitudes, as reported from earlier observations [e.g., Prölss, 1993; Williams et al., 1993; Fuller-Rowell et al., 1994; Forbes et al., 1995].
 The MSIS90 model is obviously unable to reproduce the observed density disturbances (see the lower two panels in Figure 3). First, the occurrence time of the modeled density enhancement was almost 12 hours earlier than the observed enhancement. Second, the density was underestimated by the model at all latitudes except for the southern polar region on the nightside. The underestimation was particularly evident on the dayside, with values less than half of the observed ones. Third, the model could not reflect the structure, particularly the equatorward propagation of the density disturbances, which was seen in the observations. It produced only a smooth density enhancement at high latitudes, with larger values in the summer hemisphere. As an averaged model, MSIS90 is of course not expected to predict transient storm-time disturbances in detail.
 As differences exist in the quiet-time density between different regions (particularly between day and night), the density deviation from quiet-time values, rather than the density itself, seems better suited for describing the storm effect. Therefore we will concentrate on the density deviation in the following, taking measurements from the day prior to the storm as a quiet-time reference. There are two ways to define the deviation, one is the absolute difference (storm-quiet), the other is the percentage difference (). At present, there is no general agreement on which expression is more appropriate. Burns et al.  have used the absolute term, with the argument that it is more closely related to the energy input than the percentage term. Fuller-Rowell et al.  used the absolute term as well when examining the neutral wind and composition changes. Composition changes have also been extensively studied using percentage terms [Prölss, 1980; Burns et al., 1995a, 1995b]. To get a complete picture and also to examine the consistency/inconsistency between these two terms, we present both the absolute and the percentage changes for each event. As we will see later, conclusions about the intensity of the disturbance depend at least partly on the term we choose to describe it.
 The density deviation is shown in Figure 4. The upper two panels show absolute changes, and the lower two show percentage changes. First, let us look at the absolute changes. On the dayside, strong density enhancements of 17 × 10−12 kg m−3 occurred at ∼0700 UT in the auroral regions of both hemispheres. The dayside density enhancement reached another maximum of 27 × 10−12 kg m−3 in the summer (southern) hemisphere around 2300 UT, before the occurrence of the second Dst minimum at 2500 UT. The enhancement in the winter hemisphere was only about half of this value. Strong density enhancements occurred again between 4400 and 5200 UT, corresponding to the third Dst minimum, with larger values in the summer than in the winter hemisphere. The density enhancement around 0700 UT was restricted to high latitudes and lasted less than 2 hours, thus indicating its transient nature. It could be linked to the impulsive disturbances following the SSC at 0611 UT.
 On the nightside (see second panel of Figure 4), the density disturbance showed a quite different behavior. The enhancement was generally less than half of the dayside values. In addition, it occurred about 2 hours later than on the dayside, particularly during the first storm. A prominent midnight density bulge appeared near the magnetic equator around 0930 UT. The density enhancement in winter around 2400 UT was larger and extended to lower latitudes than in summer, which was opposite to the dayside case. The hemispheric asymmetry was reversed during the second storm period between 4500 and 5200 UT. Furthermore, the nightside northern hemisphere density during this period showed very smooth and small enhancements, in contrast to a strong and equatorward propagating density bulge on the dayside.
 In the case of percentage changes, as shown in the lower two panels, the basic features remain the same as in absolute changes, with an enhanced seasonal asymmetry on the nightside. The intensity of the density enhancement on the nightside was comparable or even larger than that on the dayside. This was particularly true in the winter hemisphere during the first storm and in the summer hemisphere during the second storm, where density enhancements reached 400% and 500%, respectively. It should also be noted that when expressed in percentage term, the dayside density enhancement after 2400 UT was almost double of that before 2400 UT at midlatitudes in the winter hemisphere. However, this feature was nearly absent in the absolute changes (see top panel of Figure 4). Therefore we assert that the choice of the term affects not only the relative intensity of the storm-time disturbance between different regions but also the time history of the disturbance. Care should be taken when comparing storm effects obtained using different terms.
2.2.2. Event of 20–22 November Storm
 The density disturbances during 20–22 November 2003 are shown in Figure 5, using the same format of presentation as in Figure 4. The quiet-time reference values for this storm were much smaller than the ones for the first two due to lower F10.7, and the peak density was below 6 × 10−12 kg m−3. Let us first look at the disturbance expressed in absolute terms. On the dayside, strong density enhancements of 16 × 10−12 kg m−3 occurred around 1000 UT in the summer auroral region. It could be related to the SSC at 0800 UT. However, this enhancement was absent in the winter hemisphere. The density enhancement at high latitudes reached values above 25 × 10−12 kg m−3 in both hemispheres around 2000 UT in the main phase of the storm. However, it occurred at lower latitudes in the summer hemisphere than in the winter one. The average density enhancement at middle and low latitudes was higher in summer than in winter. Its maximum value was about 15 × 10−12 kg m−3, occurring in the main phase of the storm. Density enhancements on the nightside occurred later and were also smaller than on the dayside. At high latitudes, the maximum enhancement was about 16 × 10−12 kg m−3, in comparison with 25 × 10−12 kg m−3 on the dayside. The enhancement in summer was clearly larger than in winter. Note that the largest enhancement in the auroral region occurred at about 60°S in summer but at 75°N in winter. The MSIS90 model was unable to reproduce these features, with similar inability as noted for the October storm.
 The density enhancement during this storm was extremely large when expressed in percentage terms (see lower two panels in Figure 5). It reached more than 800% on both day and night sides around 2400 UT. The general features of the storm effect were similar to that in the absolute changes. The hemispheric asymmetry was also strengthened. However, the nightside density enhancement at high latitudes was comparable to the dayside one, rather than half of its values as expressed in absolute changes.
 The maximum density enhancement expressed in percentage terms during this storm (800%) was much larger than that of the October events (400% and 500%), though the absolute enhancement was comparable. This could, of course, be partly due to the higher intensity of this storm (Minimum Dst = −472 nT). However, its lower quiet-time reference density also contributed considerably to this difference.
 We have seen a common phenomenon in these events, that is, the density disturbance occurred first in the auroral region and later at lower latitudes. This suggests the generation of the disturbance at auroral latitudes and a subsequent equatorward propagation. To first order, the propagation time can be estimated by applying a cross-correlation analysis to the time series of the density disturbance at polar and low latitudes. To do this, we first average the density changes within three different latitudinal bands, namely, the northern and southern polar regions and the low-latitude region. The corresponding latitude ranges for averaging are taken as 60°–90°N, 60°–90°S, and 30°S–30°N (geomagnetic). These ranges are defined so as to cover the disturbances in each region as much as possible at each point in time. The absolute changes are used for this analysis, as they are more directly linked to the energy transfer involved in the propagation. In the following, we take the November event as an example. As this event is well isolated, it is suitable for studying the propagation of the storm-time disturbance, without mixing preexisting effects.
 The latitudinally averaged time series are shown in Figure 6, for both dayside and nightside. The corresponding time series obtained from the MSIS90 model, by applying the same averaging, are shown in Figure 6 as dotted lines. They demonstrates again the strong deficiency of the MSIS90 model in describing short-term disturbances. The model shows only density changes due to the day-to-day variation of the F10.7, as seen in Figure 2. The strong storm effect observed by CHAMP is almost completely absent in the modelled density.
 We apply a cross-correlation analysis to the time series (from CHAMP) at polar and equatorial regions. The obtained cross-correlation coefficients are shown in Figure 7. The x-axis denotes the time lag between the disturbance at high and low latitudes. A negative time lag indicates that the disturbance at high latitudes leads that at low latitudes. We see that the correlation coefficient reached values close to 0.9 in both hemispheres, indicating a good correlation between the disturbances at high and low latitudes. The time lag for the maximum correlation coefficient in all cases was negative, implying the propagation of the disturbance from high to low latitudes. The propagation time can be roughly estimated from the time lag at the maximum coefficient. On the dayside, it was about 1.5 hours in the northern/winter hemisphere, and about 1 hour in the southern/summer hemisphere. On the nightside, the propagation took almost 2.5 hours in the winter hemisphere and about one hour in the summer hemisphere. These estimates are quantitatively rather crude due to the large latitude averaging and the sparse sampling (once per 93 min).
3. Discussion and Conclusion
 Our analysis of three storms has shown quite different storm effects in the noon and midnight sectors. First, we expected to see a larger density enhancement on the nightside than on the dayside based on the framework mentioned in the introduction. The examination of three storms, however, shows that this is not the case. The absolute density changes on the nightside were only about half of the values on the dayside. Even the percentage change was not obviously higher than that on the dayside, though it was comparable. Thus in spite of the close link between thermospheric total mass density and composition, the idea which explains stronger composition changes under lower “quiet-time” background [Burns et al., 2004] cannot be generalized to the total thermospheric mass density.
 Second, the hemispheric/seasonal asymmetry in the noon sector is different from that in the midnight sector during three storms. In the noon sector, the density enhancement in the winter (northern) polar region occurred at somewhat higher latitudes than in the summer (southern) polar region during all three storms. This asymmetry is even stronger when plotted in geographic coordinates (not shown). The average density enhancement in the summer hemisphere was larger and penetrated to lower latitudes than in the winter hemisphere. These features are consistent with the simulation results of Fuller-Rowell et al. . They attributed this seasonal effect to the prevailing summer-to-winter circulation. On the nightside, a hemispheric/seasonal asymmetry between locations of the enhancement was not obvious in polar regions during all three storms. However, it varied with individual events at middle and low latitudes. The density enhancement was stronger and penetrated to lower latitudes in summer than in winter during the second (30–31 October) and the third (20–23 November) storms, but the first storm (29–30 October) showed the opposite feature. During this storm, the winter hemisphere experienced a much stronger disturbance than the summer hemisphere (see Figure 4). As the October events occurred closer to the September equinox than to the December solstice, the difference between summer and winter hemispheres may not be as obvious as that for the November event. This might have played a role in the difference between the October and November events. However, even the two October events, which occurred on three adjacent days, exhibited quite different hemispheric behavior. We therefore see that the hemispheric/seasonal asymmetry of the thermospheric density disturbance varies from event to event, at least in the midnight sector. Here, the summer hemisphere does not always experience stronger disturbances than its winter counterpart. This conclusion applies to density changes described both in percentage and in absolute terms during these three storms.
 Third, the equatorward propagation of the density disturbance in the noon sector is different from that in the midnight sector. In spite of its great simplification, the cross-correlation analysis of the November 2003 event yielded consistent results with previous studies [Fuller-Rowell et al., 1996], concerning the seasonal effect of the propagation. For instance, the equatorward propagation is faster in the summer hemisphere than in the winter hemisphere. However, it also showed for this event that in the winter hemisphere, the propagation on the dayside appeared to be faster than on the nightside. This is not expected from simulations of Fuller-Rowell et al. , which predict faster and more efficient equatorward propagation on the nightside. Their explanation was that on the nightside, the storm-time density bulge at high-latitudes encounters equatorward winds, while on the dayside, it encounters poleward winds. One possible explanation for our observations is that the equatorward wind produced by the storm-time density gradient at high latitudes was much larger (double) on the dayside than on the nightside. It could thus overcome the poleward wind on the dayside, leading to a faster propagation than on the nightside.
 In conclusion, the thermospheric density response to magnetic storms is a multifaceted phenomenon. It exhibits strong day-night and hemispheric/seasonal asymmetries. Furthermore, these asymmetries may vary from storm to storm. The direct geomagnetic driver of these effects is yet to be identified. Some of the observed features are possibly due to spatially asymmetric energy input at high latitudes, which could result from activities in the cusp region and substorms. Asymmetry between hemispheres may exist in the energy input as well. Further studies comparing observations and model results with Joule heating from the AMIE technique as input may provide us some clues.
 The work of Huixin Liu was supported by the Alexander von Humboldt Foundation and the International Communication Foundation. We thank Gang Lu for providing us the AE and Dst data and W. Köhler for preprocessing the satellite accelerometer data. We also appreciate the operational support of the CHAMP mission by the German Aerospace Center (DLR) and the financial support for the data processing by the Federal Ministry of Education and Research (BMBF).
 Arthur Richmond thanks Stephan Buchert and Frank Marcos for their assistance in evaluating this paper.