On the abundance and source contributions of dicarboxylic acids in size-resolved aerosol particles at continental sites in central Europe

Dicarboxylic acids (DCAs) are among the most abundant organic compounds observed in atmospheric aerosol particles and have been extensively studied at many places around the world. The importance of the various primary sources and secondary formation pathways discussed in the literature is often difficult to assess from field studies, though. In the present study, a large data set of size-resolved DCA concentrations from several inland sites in Germany is combined with results from a recently developed approach of statistical back-trajectory analysis and additional data. Principal component analysis is then used to reveal the most important factors governing the abundance of DCAs in different particle size ranges. The two most important sources revealed are (i) photochemical formation during intense radiation days in polluted air masses, likely occurring in the gas phase on short timescales (gasSOA), and (ii) secondary reactions in anthropogenically influenced air masses, likely occurring in the aqueous phase on longer timescales (aqSOA). While the first source strongly impacts DCA concentrations mainly in small and large particles, the second one enhances accumulation mode DCAs and is responsible for the bulk of the observed concentrations. Primary sources were found to be minor (sea salt, soil resuspension) or non-existent (biomass burning, traffic). The results can be regarded as representative for typical central European continental conditions.


Introduction
Low-molecular-weight dicarboxylic acids (LMW-DCAs) are arguably the most abundant compound group typically identified in atmospheric organic aerosol particles.As such they have a strong impact on the properties of organic particles and can play a role in the hygroscopic growth of particles (Choi and Chan, 2002;Prenni et al., 2003), the activation of cloud condensation nuclei (Pradeep Kumar et al., 2003;Booth et al., 2009), and ice formation (Zobrist et al., 2006).Furthermore, they often serve as tracers for the oxidative processing of water-soluble organic aerosol in the atmosphere (Sorooshian et al., 2010;Ervens et al., 2011).
A variety of both primary and secondary sources of LMW-DCAs have been proposed in the literature.Among the primary ones are the rather minor sources of cigarette smoke (Rogge et al., 1994) and meat charbroiling emissions (Rogge et al., 1991), as well as the potentially more relevant one of fossil fuel combustion (Kawamura and Kaplan, 1987;Fraser et al., 1998).Evidence for the latter one is conflicting, though, with more recent studies neglecting a contribution of traffic to LMW-DCA concentrations in the atmosphere (Rogge et al., 1993a;Huang and Yu, 2007;Stone et al., 2010).DCAs were observed in high concentrations in biomass burning plumes (Jaffrezo et al., 1998;Narukawa et al., 1999;Graham et al., 2002;Gao et al., 2003;Allen et al., 2004;Falkovich et al., 2005;Kundu et al., 2010), although it is unclear to what extent they are directly emitted (Jaffrezo et al., 1998) and/or formed during the processing of the plume (Gao et al., 2003;Allen et al., 2004).In marine-influenced air masses, direct emission of DCAs associated with sea salt particles can be a primary source (Matsumoto et al., 1998), as sea water has been shown to contain non-negligible concentrations of DCAs (Tedetti et al., 2006).
The large variety of secondary sources in diverse atmospheric regimes is a direct result of LMW-DCAs being lategeneration products in the oxidation pathways of primary gaseous precursors.In field studies, it is therefore often difficult to evaluate the contribution of the different sources and formation pathways to the observed concentrations.The aim of the present study was to elucidate the factors governing the abundance of LMW-DCAs at typical European continental sites.Size-resolved concentration data from several campaigns were compiled to ensure robust results from a statistical source apportionment approach (principal component analysis, PCA).An attempt was made to discriminate impacts from natural vegetation versus anthropogenic emissions on the formation of LMW-DCAs by including the results of advanced back-trajectory analysis and further data into the PCA.Some of the data used in this study have already been published within the context of their respective campaigns (Neusüß et al., 2002;Plewka et al., 2004;Gnauk et al., 2005;Müller et al., 2005;Brüggemann et al., 2009).This refers mainly to inorganic ions and organic/elemental carbon (OC/EC) concentrations, which are used as indicators for certain sources in the source apportionment.The DCA data used in this study have not been published before, with the exception of 8 out of 20 samples from Melpitz (Plewka et al., 2004) and 3 out of 10 samples from Goldlauter (Müller et al., 2005;van Pinxteren et al., 2005).Size-resolved particle sampling was done at all sites using five-stage Berner impactors (Hauke, Austria) with 50 % cut-offs at 0.05, 0.14, 0.42, 1.2, 3.5, and 10 µm aerodynamic diameter (D p ) and a flow rate of 75 L min −1 .The impactors were equipped with aluminium foils as impaction substrates, which were heated at 350 • C for at least 2 h to reduce blank levels prior to sampling.During some campaigns, a second impactor was run in parallel and equipped with Tedlar ® foils which were cleaned with 5 % aqueous hydrogen peroxide solution.Sampling durations varied between a few hours and 1 day (typically 12 or 24 h), depending on the aims of the respective campaign.Directly after sampling, the impaction substrates were stored at −20 • C until analysis, which was usually performed within about a year from the time of sampling.

Sampling
Low-pressure impactors can suffer from sampling artefacts such as evaporation of semi-volatiles from the lowest stages or particle bounce-off.To reduce particle bounce, RHcontrolled sampling was performed during some of the campaigns (Melpitz, Falkenberg, Goldlauter) by installing the impactor downstream of a bundle of 7 3/4 in stainless steel tubes (1.5 m length), which were temperature controlled to keep the RH of the sampling air at approx.60 ± 10 %.Inlet heights were between approx.4 and 20 m above ground for the different campaigns.

Measurements
Total particle mass was determined by weighing the impaction substrates before and after sampling on a UMT-2 microbalance (Mettler-Toledo, Switzerland) with a reading precision of 0.1 µg and a standard deviation of ca. 1 %.For RH equilibration, the substrates were stored in a closed box at constant temperature (20 ± 2 • C) and RH (55 ± 5 %) for usually 48 h (at least 12 h).A fraction of the Tedlar ® foil substrate (or aluminium substrate where only one impactor was run) was extracted in 1-2 mL deionized water by ultrasonication and shaking, while a fraction of the aluminium substrate was used for the determination of organic and elemental carbon (OC/EC).
The aqueous particle extracts were analysed for inorganic cations by capillary electrophoresis (CE) or ion chromatography (IC), while inorganic anions as well as the DCA were always determined by CE.Reproducibility of the CE method for DCA determination is in the range of 5-10 %.More details of the respective methods can be found elsewhere (Neusüß et al., 2000;Brüggemann and Rolle, 1998).The applied extraction and analysis method determines the sum of both the dissociated and free acid form of a respective DCA in the particles.To differentiate between the two forms, parallel extractions in organic solvents together with an appropriate analysis method would be required (Yang and Yu, 2008).
OC/EC determination was done by a two-step thermographic method using a C-mat 5500 carbon analyser (Ströhlein, Germany).The temperature for OC volatilization under nitrogen was 590 • C during the early campaigns (Melpitz and Falkenberg), while for all other samples 650 • C was used.EC was determined at 650 • C under oxygen.More details of the method are given elsewhere (Neusüß et al., 2000;Gnauk et al., 2008).
Mixing ratios of the trace gases ozone (O 3 ), nitrogen monoxide (NO), nitrogen dioxide (NO 2 ), sulfur dioxide (SO 2 ), and carbon monoxide (CO) were measured at all sites except Falkenberg using commercial standard trace gas monitors from different manufacturers using UV absorption (O 3 ), UV fluorescence (SO 2 ) and chemiluminescence (NO x ).The gas monitors were calibrated on a regular basis.Mixing ratios at the different sites are summarized in Fig. S1 of the Supplement.The data were averaged within the sampling intervals and used in the source apportionment approach to reflect local pollution conditions.

Back-trajectory analysis
In this study, a method of back-trajectory calculation and GIS analysis to obtain proxy parameters for the impacts of several land cover classes on the sampled air masses has been applied to the data set of samples from the various campaigns.The method is described in full detail by van Pinxteren et al. (2010) and will only briefly be outlined here.96 h back trajectories were calculated using HYSPLIT v4.9 (Draxler and Hess, 1998) with input fields (1 • resolution) obtained from the FNL archive at NOAA's Air Resource Laboratory (http://ready.arl.noaa.gov/archives.php).The model was run in the ensemble mode with a starting height of 500 m above ground level and offsets of one meteorological grid point in the horizontal dimension and 0.01 sigma units (ca.250 m) in the vertical dimension, yielding an ensemble of 27 trajectories for all possible offsets in the x-, y-, and z-dimension for a given starting time.Trajectory ensembles were calculated for every full hour during a sampling interval and intersected with land cover data from the Global Land Cover 2000 project of the European Commission Joint Research Centre (GLC2000 database, http://bioval.jrc.ec.europa.eu/products/glc2000/glc2000.php)within a database system with GIS functionality (PostgreSQL 9.1 + PostGIS 2.0).The outcome is a dimensionless residence time index (RTI), which -on a relative scale between 0 and 1 -reflects the (weighted) time the sampled air masses have resided above certain land cover categories (water, natural vegetation, agricultural lands, bare areas and urban areas).The land cover classes were obtained by reclassifying the original 22 classes of the GLC2000 data set into the broader categories as given by van Pinxteren et al. (2010).Given that emissions from a given area will be strongly influenced by its land use (e.g.biogenic emissions from areas with natural vegetation, anthropogenic emissions from urban areas), the index can be regarded as a measure for potential impacts of such typical emissions from the different land cover categories to the sampled air mass during their transport to the receptor site.The residence time index is calculated in a time-weighted manner to account for a decreasing impact of areas far away from the receptor site (i.e.backward in trajectory time).
In a similar way, the meteorological output of HYSPLIT can be averaged in a weighted manner "along the trajectories" to yield a proxy for the solar flux during the past 4 days of the sampled air mass.Additionally, the HYSPLIT solar flux as well as the HYSPLIT mixing layer depth (modelled from the input meteorological fields) can be averaged at the location of the receptor site over a given sampling interval and the mean length of the trajectories during a sampling interval can be calculated.These parameters proved helpful in a previous application of the method (van Pinxteren et al., 2010) and are used in the present source apportionment study as well.

Principal component analysis
PCA was performed on the data set of particulate concentrations, trace gas mixing ratios and trajectory parameters using the open source statistical software package R (http: //www.r-project.org).All concentration data (both particulate and gas data) as well as the two solar flux and the mixing height variables were log10-transformed to reduce the skewness in the variables.The skewness results from the approximate lognormal distribution of these variables and can potentially distort the results of a PCA.After log transformation, the data was mean-centred, scaled to unit variance and the correlation matrix was used for calculation of the principal components, which is necessary if variables on very different scales are to be used.As PCA requires a complete set of data, concentrations below the detection limit were set to half the detection limit, while missing data (e.g.due to instrumental problems or samples not analysed for the full suite of compounds) was replaced by the mean value of the respective parameter.It has to be noted that during the Melpitz campaign only a fraction of the samples (6 out of 20) were analysed for the full suite of DCAs, while oxalate was analysed from all of these samples.By replacing the missing data with their mean values they are set to the non-informative point at the origin of the mean-centred data and thus do not influence the outcome of the PCA.The number of factors to extract was defined by examining the scree plot of eigenvalues vs. number of PCs (Cattell, 1966) and by comparing the magnitudes of the obtained eigenvalues with those obtained from a random data matrix of the same size as the original (so-called parallel analysis, Horn, 1965).Varimax rotation was applied to the extracted principal components to result in rotated components with easier-to-interpret component loadings.

PM 10 concentrations and size distributions of DCAs
In Table 1 the concentrations of DCAs obtained in this work are summarized.Literature data from comparable European rural and urban sites are given for comparison in Table 1.Generally, the mean DCA concentrations obtained in this study do not differ by more than a factor of about 2 between the five different sites.The urban sites do not necessarily show higher concentrations than the rural sites.Highest concentrations are always observed for oxalic followed by malonic acid.Succinic, glutaric (including two unresolved isomers, Neusüß et al., 2000), and malic acid often show comparable concentration ranges, with the exception of the summertime Falkenberg data, where glutaric and malic acid were considerably higher than succinic acid.Tartronic and tartaric acid are observed at rather low concentrations of only a few ng m −3 , if at all.A notable exception is again the Falken-berg data, where tartronic acid was observed at rather high concentrations.
Compared to the literature data, this work's DCA concentrations at urban sites usually agree within a factor of 2 with similar sites in Europe.Oxalic acid determined by Röhrl and Lammel (2001) during summer in Leipzig and Eichstädt somewhat exceeds the concentration observed at the other sites, which might be due to the sampling of total suspended matter (TSP) and/or -as indicated by the authors -possible positive sampling artefacts for oxalic acid.Concentrations at rural sites agree to a similar extent, with the exceptions of rather high TSP oxalic and malonic acid in Falkenberg (Röhrl and Lammel, 2001) and PM 2.5 oxalic acid at K-puszta in the central plain of Hungary (Legrand et al., 2007).Unusually low concentrations of C3-C5 DCAs were recently reported for a rural site in Sweden (Hyder et al., 2012), possibly related to the choice of extraction solvent and/or derivatization agent for a gas chromatography/mass spectrometry (GC/MS) analysis applied in this study.Yang and Yu (2008) demonstrated that the combination of solvent extraction and trimethylsilylation GC/MS recovers only a fraction of total malonic acid (including malonate) concentrations in PM 2.5 samples, while -as reported by Hyder et al. (2012) as welloxalic acid was not detected at all.
The size distributions of DCAs are given in Fig. 1 as averages over all sites.Highest concentrations for all DCAs were observed in accumulation mode particles (D p = 0.14-1.2µm) with significantly lower concentrations in ultrafine (D p = 0.05-0.14µm) and coarse (D p = 1.2-10 µm) particles.Such size distributions are typical for European continental DCA measurements and have been observed in similar ways before (Mészáros et al., 1997;Kerminen et al., 2000;Müller et al., 2005;Legrand et al., 2007).

Back-trajectory analysis
The residence time indices as main results from the statistical back-trajectory analysis are shown in Fig. 2. As expected from the location of the sites (in-land continental), the sampled air masses had higher (time-weighted) residence times from the two main continental land cover classes of agriculture and natural vegetation as compared to the residence time above marine areas (water).Urban areas make up only a tiny fraction of continental surface area, which is why the values of this index are much smaller.The indices show a large variability, indicating that very different air masses were sampled in the various campaigns.Further output parameters from the back-trajectory analysis as described above are given in Fig. 2 as well.In agreement with the campaign season, solar flux and mixing depth were quite different at the different sites with higher values for the Falkenberg summer measurements, lower ones for the autumn campaigns in Melpitz and Goldlauter and highly variable values for the year-round measurements in Dresden and Leipzig.

Source apportionment
To assess the latent factors controlling the abundance of DCAs in different particle sizes, PCA was performed on the data set of size-resolved DCA concentrations.The following further particle constituents were included in the analysis in order to aid the interpretation of the resolved principal components: EC (primary, anthropogenic combustion or biomass burning), OC (both primary and secondary, both anthropogenic and biogenic), sulfate (mainly secondary, anthropogenic combustion), potassium (primary, biomass burning), sodium (primary, sea salt and/or soil) and calcium (primary, dust/soil).Due to missing concentration data for sodium and calcium on the smallest particles (impactor stages 1 and 2, all data below detection limit), concentrations from impactor stages 3-5 only were included for these species.Additionally, trace gas mixing ratios of NO x , CO and SO 2 (anthropogenic combustion) and O 3 (photochemically formed) were included.All the back-trajectory parameters were included as well with the exception of the residence time indices for bare areas (negligible values for sampled air masses) and water (simply anti-correlated to the continental indices).
In PCA, the number of PCs to extract is a critical (and to some extent subjective) decision which can strongly affect the interpretability of the result.Many different procedures exist to aid this decision (Jolliffe, 2002).One of the most common ones is the scree plot of successive eigenvalues for all calculated PCs, which is shown in Fig. 3.As can be seen, the rate of decline is fast for the first seven PCs and then clearly levels off.The "elbow" of the plot (i.e.PC 7) is often considered to indicate the number of components to extract (Jolliffe, 2002).On the other hand, the eigenvalue of PC 7 is already significantly lower than that obtained from a random matrix of univariate normal data ("simulated data" in Fig. 3) and from random samples (randomized across rows) of the original data ("resampled data" in Fig. 3).Thus, in the final solution the first six PCs were chosen to be subjected to Varimax rotation.The "correctness" of the six-factor solution was, however, verified by varying the number of extracted PCs between six and seven, as will be discussed below.In Table 2 the loadings of the variables on the six rotated components are shown, which basically describe the correlation of the variable with the component.Note that for better readability loadings with |x| < 0.2 were considered insignificant and removed, while loadings with |x| ≥ 0.6 were considered "high" and are printed bold.

PC 1: anthropogenically influenced gasSOA
The first rotated component (PC 1) explains the largest part of the total variance in the data set (19 %) and has significant loadings for all DCAs in basically all particle size ranges.Highest loadings are observed for DCAs in smallest (D p = 0.05-0.14µm) and largest (D p = 3.5-10 µm) particles, while they tend to decrease towards accumulation mode particles (D p = 0.42-1.2µm) (even non-significant for succinic acid).The hydroxy-DCA malic acid shows a somewhat different behaviour with continuously decreasing loadings from smallest towards largest particle sizes.High loadings are also observed for the two solar flux variables and for O 3 .PC 1 likely represents photochemical formation of DCAs in the gas phase (gasSOA) during days with high intensity of solar radiation.The gas phase formation is inferred Residence time index q q q q q q q Mean trajectory length (i) from the typically low ambient relative humidity (RH) and cloud cover during such days, leading to low amounts of liquid water in the atmosphere and (ii) from the strongly increasing loadings towards smaller particle sizes, reflecting the higher surface area available for condensation.Regarding RH, it has to be noted that it is strongly anti-correlated to the solar flux parameters in our data set and its inclusion to the PCA leads to strong negative loadings for both RH along the trajectory as well as RH at the receptor site (data not shown).Regarding particle sizes, the high loading of sulfate in very small particles is consistent with PC 1 representing a gas phase source, as gaseous formation and subsequent condensation of sulfuric acid is a well-known process under the influence of strong solar radiation (Hamed et al., 2010).
The high loadings of DCAs in coarse particle size ranges might be explained by their repartitioning onto crustal material and/or mineral dust particles.Calcium -a typical crust component -has a quite high loading to PC 1 for the largest particle size range as well.Due to their less acidic character (Scheinhardt et al., 2013), the diacids might strongly condense to such particles despite the lower surface area they provide.Such a preferential partitioning of DCA has already been suggested previously (Kerminen et al., 2001).
The different trend of malic acid loadings with respect to coarse particles is consistent with the lower vapour pressures of hydroxyl-DCAs and thus less favourable evaporation from smaller particles.Heterogeneous photochemical reactions on dust could be another plausible explanation for the observed coarse mode loadings.Regarding the timescale of the photochemical production of DCA in PC 1, the very similar magnitude of the two solar flux loadings indicate a rather local and thus quick formation.If major fractions of particulate DCAs would have been formed during previous days already and just been transported to the site, one could expect the solar flux along the back-trajectory loading to be higher than the local solar flux loading.Even though this might not be a very strong indication, it is consistent with a study on highly time-resolved oxalate concentrations, showing a strong diurnal profile during summer and thus a rather fast formation (Fisseha et al., 2006).Based on these indications, the timescale of gas phase DCA formation might be estimated to lie in the range of minutes to hours, depending on the photochemical conditions.A real quantitative discussion of formation kinetics is, however, not feasible from the data of the present study.Regarding the type of emissions, which might impact DCA formation in PC 1, the negative loading of mean trajectory length to PC 1 indicates comparatively short back-trajectories, thus rather high residence times above continental areas as compared to marine areas (where longer trajectories spend much of their travelling time for the given sampling sites of this study).It has to be noted, though, that all back trajectories of sampled air masses are well above 1000 km in mean trajectory length (Fig. 2).The negative loading of this parameter does therefore not indicate a local influence of emissions.In fact, local emissions cannot be resolved by the coarse resolution of trajectory calculation.Among the residence time indices, urban areas have the highest loading, followed by a minor loading of agricultural areas, and a non-significant loading of natural vegetation, indicating a significant influence of anthropogenic emissions to this factor.Overall, PC 1 likely represents secondary organic aerosol formed in the gas phase of polluted air masses (gasSOA) as an important source of DCAs.

PC 2: anthropogenically influenced aqSOA
The second rotated component (PC 2) explains another 15 % of the total variance.Moderate to high loadings on PC 2 were found for all DCAs in the particle size range D p = 0.14-3.5 µm (impactor stages 2 to 4) with maximum loadings in accumulation mode particles with D p = 0.  masses with long residence times above continental areas and a strong input of anthropogenic emissions including biomass burning.The high loading of agricultural areas might reflect the impact of sources located outside urban areas, such as industrial facilities, roads and power plants.The moderate loading of SO 2 , whose main source in Europe is energy production (European Environment Agency, 2010), seems to support especially the latter suggestion.The solar flux parameters do not show significant loadings on PC 2, indicating that the formation of DCAs in this factor is not primarily driven by strong radiation.Together with the regional character of the source (highest loadings in accumulation mode particles), it can be assumed that DCA production takes place on a slower timescale (hours to days) as compared to PC 1.In principle, contributions from primary sources to PC 2 cannot be excluded, as EC and -to a lesser extent -potassium correlate with the factor as well.However, the two primary sources which are most often discussed in the context of DCA emissions (traffic and biomass burning) are resolved by the model in PCs 3 and 4. PC 2 might thus represent the SOA from traffic and/or biomass burning, while PCs 3 and 4 represent their primary components.The similarity between sulfate and DCA loadings on PC 2 implies a high correlation of their concentrations in this source.Good correlation of sulfate with oxalate (and sometimes other DCAs) has often been reported from field studies (Yu et al., 2005;Sorooshian et al., 2006;Hsieh et al., 2007Hsieh et al., , 2008;;Miyazaki et al., 2009) and can be taken as an indication of aqueous phase formation of DCAs (Yu et al., 2005;Sorooshian et al., 2006) due to aqueous SO 2 oxidation being the dominant source of sulfate in the atmosphere (Yu et al., 2005).From size-resolved DCA measurements, a dominant "droplet mode" with D p ca. 0.7-0.9µm as opposed to a less abundant "condensation mode" www.atmos-chem-phys.net/14/3913/2014/Atmos.Chem.Phys., 14, 3913-3928, 2014 with D p ca. 0.2-0.4µm can be taken as evidence for aqueous phase formation as well (Yao et al., 2003;Hsieh et al., 2009;Lan et al., 2011), in analogy to earlier discussions on the abundance of sulfate in these two particle modes (Hering and Friedlander, 1982;John et al., 1990;Meng and Seinfeld, 1994).Both lines of thought (correlation with aqueous phase formed sulfate and aqueous phase production leading to a "droplet mode") are supported to some extent by PC 2 as well (high loadings of DCAs and sulfate as well as highest loadings in "droplet mode" size range).RH during sampling did not help to further elucidate the role of aqueous chemistry in our data set, as PCA runs with RH included resulted in insignificant loadings to PC 2.Even though more direct evidence of aqueous phase formation would be desirable (e.g. the concentrations of aqueous phase oxidants or the amount of available particle and/or cloud water), it can be concluded that aqueous phase formation of DCAs likely plays an important role in the source resolved by PC 2 and it can thus be attributed to aqueous phase SOA (aqASOA) with a significant anthropogenic influence on either the precursors and/or the oxidant chemistry.

PC 3: traffic
The third rotated component (PC 3) explains 10 % of the variance and has high positive loadings for NO x and CO.EC shows a high loading in ultrafine particles (D p = 0.05-0.14)and moderate and continuously decreasing loadings in larger particle size fractions, while the solar flux parameters and the mixing depth at the receptor site have negative loadings.PC 3 can clearly be attributed to a traffic source, presumably in wintertime conditions, where pollution can build up quickly due to less efficient boundary layer mixing processes.All significant DCA loadings on PC 3 are negative, thus strongly implying that primary traffic emissions are not a significant source of DCAs in this data set.

PC 4: biomass burning
The fourth rotated component (PC 4) explains another 9 % of the total data variance and contains mainly potassium with high loadings in all particle size ranges.In addition, sulfate, sodium and calcium show significant loadings within some size ranges and SO 2 has a rather strong loading as well.As fine-particle potassium is a classical tracer of biomass burning (Khalil and Rasmussen, 2003), PC 4 is assumed to represent a primary biomass burning source.No DCA variable loads significantly on this factor, which is why biomass burning is unlikely to be a primary source of DCAs in the present study.

PC 5: sea salt
The fifth rotated component (PC 5) explains 8 % of the total variance.High loadings of coarse mode sodium together with moderate loadings of coarse mode sulfate are observed.The mean trajectory length shows a moderate positive loading, while the residence time indices for natural vegetation and agriculture are anticorrelated to the factor.PC 5 can be attributed to (processed) sea salt.Among the DCAs, only oxalate in coarse mode particles shows small loadings, while other DCAs have either insignificant or negative loadings.Sea salt might thus contribute to some small extent to coarse mode oxalate concentrations at continental sites in Germany.
A sea salt mode of DCA is usually observed in marine regions (Kerminen et al., 1999(Kerminen et al., , 2000;;Neusüß et al., 2000;Mochida et al., 2003b).Whether the oxalate measured at the inland sites of this study indeed originates from sea salt or from the processing of these particles on their way to the receptor sites (e.g. by condensation to less acidic sea salt particles) cannot be concluded from the data, though.

PC 6: soil
The sixth rotated component (PC 6) explains 7 % of the data variance and contains moderate to high loadings of OC in all particle size ranges and of sulfate and calcium in coarse mode particles.EC shows significant loadings in small and large particles, while minor loadings are observed for a few other variables.Based on this loadings pattern the attribution of a source to this factor is less evident than for the previous components.Mainly due to the loadings of calcium, sulfate and OC, PC 6 is tentatively interpreted as a primary soil source.The EC loadings might then result from complex high-molecular-weight soil organic matter (humic substances), which does not likely fully evaporate at 650 • C and is thus incorrectly attributed to the EC fraction of total particulate carbon in our method of OC/EC determination.Among the DCAs, minor loadings of coarse mode oxalate and malonate are observed, while coarse mode succinic acid shows moderate loadings to this factor.Organic acids, including the DCAs of this study, are known to be important constituents of soil organic matter (Jones, 1998), thus resuspended soil in coarse mode particles can be assumed to contribute to coarse mode DCA concentrations.
As discussed above, a seven-factor solution for the PCA could in principle be justified from the scree plot as well.The main difference to the six-factor solution is that the loadings of the original PC 6 are then distributed between the new PC 6 and PC 7, i.e.PC 6 seems to be split into two factors.One factor tends to hold the loadings on impactor stage 4 (D p =1.2-3.5 µm), while the other tends to take over the stage-5 (D p = 3.5-10 µm) loadings to some extent.The variables for which significant loadings are observed for the new PC 6 and PC 7 are the same as for the original PC 6 discussed here.The interpretability of the new PC 6 and PC 7 is even D. van Pinxteren et al.: On the abundance and source contributions of dicarboxylic acids more ambiguous than for the original PC 6.This is why the seven-factor solution has been considered to resolve two artificial sources instead of one real one and has been discarded.

Discussion of main DCA sources
The results of the PCA show that the major part of the variance of DCAs in different particle size ranges can be attributed to two main sources: (1) fast photochemical formation during intense radiation days, likely due to gas phase oxidation of precursors in anthropogenically influenced air masses (gasSOA), and (2) slower, but presumably more continuous, secondary formation in anthropogenically polluted air masses, possibly with an important contribution of aqueous phase oxidation processes (aqSOA).Much smaller contributions to the abundance of LMW-DCAs can be expected from primary biomass burning emissions, (processed) sea salt and primary soil emissions.
While the importance of secondary processes in the formation of LMW-DCAs is well established, they are typically assumed to be products of aqueous phase reactions (Kanakidou et al., 2005;Ervens et al., 2011).From the results of this study, however, a significant contribution from gas phase reactions seems possible (PC 1) under certain conditions.For the 10 % of samples which have the highest scores on PC 1 (i.e. the days with presumably the highest impact of gasSOA) the sum of impactor stage 1 and 2 (D p = 0.05-0.42µm) concentrations for a given DCA typically makes up 45-55 % of its total PM 1 and 30-55 % of its total PM 10 concentration.Making the (very crude) assumption that stage 1 and stage 2 concentrations represent DCAs formed solely from gas phase oxidation processes, these results would indicate that roughly up to one third to one half of total DCA concentrations can be attributed to gas phase sources under appropriate conditions.Given, however, that condensed phase reactions likely contribute to DCA concentrations in these particle sizes (impactor stages 1 and 2) as well, the contribution of gas phase sources to total DCA concentrations will likely be lower.To the best of the authors' knowledge, the formation of LMW-DCAs is at present not included in comprehensive atmospheric gas phase chemistry models.For example, none of the DCAs determined in this study are incorporated in the Master Chemical Mechanism v3.2 (http://mcm.leeds.ac.uk/MCM).Two possibilities of gas phase formation of organic acids are well known, though: (i) ozonolysis of unsaturated compounds (see Orzechowska and Paulson (2005) and references therein) and (ii) the reactions of acylperoxy radicals with hydroperoxyl (HO 2 ) radicals (see Calvert et al., 2011 and references therein).Ozonolysis of semi-volatile unsaturated fatty acids has since long been suggested as a possible source of LMW-DCAs in the atmosphere (Kawamura et al., 1996a) and DCAs have frequently been observed in the aerosol phase in chamber studies of unsaturated hydrocarbons (Hatakeyama et al., 1985(Hatakeyama et al., , 1987;;Kalberer et al., 2000;Fisseha et al., 2004;Gao et al., 2004;Hamilton et al., 2006;Sato et al., 2007;Sorooshian et al., 2007;Borras and Tortajada-Genaro, 2012).Depending on the position of the C-C double bond in the monocarboxylic acid precursor, the C2 to C5 DCAs observed in this study could easily result from such reactions.Ozone has recently been suggested to be an important driver in the formation of carboxylic acid containing SOA (Liu et al., 2011) and it is correlated with PC 1 in our study as well.The oxidation of ω-oxo-monocarboxylic acids to DCAs has also been proposed in the past (Kawamura et al., 1996a).This reaction would proceed via abstraction of the aldehydic H atom (e.g. by hydroxyl radicals (OH)), addition of molecular oxygen and reaction of the formed acylperoxy radical with HO 2 to result in the corresponding DCA.The relevance of these qualitatively feasible formation pathways, however, depends on their competition with the manifold of other reaction channels in the case of ozonolysis (Orzechowska and Paulson, 2005) and -for continental conditions -mainly with the reaction of acylperoxy radicals with NO, leading to chain-shortened aldehydes (Calvert et al., 2011).In general, these pathways can be expected to be of minor importance only, although kinetic data and product studies of relevant precursors (unsaturated or oxygenated acids) are very limited (Calvert et al., 2011).If the concentrations of precursors (and oxidants) are high enough, however, they might significantly contribute to the observed LMW-DCA concentrations even in low yields.
Another possible gas phase formation pathway of carboxylic acids could be the reaction of enols with OH.Recently, the common assumption of keto-enol tautomerization of gas phase aldehydes and ketones being negligible in the atmosphere has been challenged by both theoretical and experimental studies (da Silva, 2010; Andrews et al., 2012;Clubb et al., 2012).It was suggested that a significant proportion of carbonyl compounds might exist in their (highly reactive) enol form.Addition of OH to the C-C double bond then results in a carboxylic acid to be formed, which could account for the discrepancies between typical modelled versus measured budgets of organic acids in the atmosphere (Archibald et al., 2007).This could hypothetically be another pathway for the gas phase oxidation of keto-acids to LMW-DCAs.Naturally, much more thermodynamic and kinetic data and implementation into models is needed to assess the relevance of such a formation pathway.
Despite a potential relevance of gas phase processes (PC 1), in terms of total DCA concentrations, PC 2 likely represents the most important source, as highest concentrations of DCAs are typically observed on impactor stages 2 and 3 (D p = 0.14-1.2µm, see Fig. 1).This is in line with the acknowledged importance of aqueous phase oxidation processes on the formation of LMW-DCAs in the atmosphere (Tilgner and Herrmann, 2010;Ervens et al., 2011).Water-soluble products of volatile organic compound (VOC) oxidation can readily be taken up by deliquesced particles and cloud droplets.Especially by the hydration of aldehyde groups to gem-diols, this uptake opens a reaction channel towards carboxylic acids which is not available in the gas phase.
As outlined above (Introduction), the precursors of DCAs are numerous and can originate from both continental and marine as well as from anthropogenic and biogenic sources.For the sampling sites of the present study, the results of back-trajectory analysis suggest that continental sources dominate (negative correlation with trajectory length) and indicate an influence of anthropogenic activities on the DCA concentrations for both PC 1 (correlation with urban areas RTI) and PC 2 (correlation with agriculture and urban areas RTI).It has to be noted that correlations with primary anthropogenic trace gases are weak or even absent in accordance with the secondary nature of the two main sources, which implies highest concentrations away from direct sources (dispersion/dilution of trace gases).These results might indicate the importance of anthropogenic precursors on DCA concentrations.There is, however, evidence from several field studies on SOA formation that despite a strong anthropogenic influence to the studied air masses (typically deduced from correlations with anthropogenic tracers), a dominant fraction of carbonaceous material is composed of modern carbon (in contrast to fossil carbon) and thus originates from biogenic sources (Weber et al., 2007;Worton et al., 2011;El Haddad et al., 2013).The coupling of an anthropogenic and biogenic component in sampled air masses has been suggested to potentially enhance biogenic SOA formation, e.g. through higher oxidant concentrations due to higher NO x and/or higher partitioning of biogenic SOA due to higher concentrations of (primary or secondary) anthropogenic organic aerosol (Carlton et al., 2010).Thus, in the absence of additional data, firm conclusions on the nature of DCA precursor VOCs cannot easily be drawn.

Conclusions
A relatively large set of size-resolved samples of aerosol particles has been used to study the abundance of LMW-DCA in different particle size ranges at several continental sites in Germany.PCA was used to qualitatively distinguish between different sources and formation pathways.Together with additional data from back-trajectory analysis and trace gas measurements, two main sources of DCA could be identified: photochemical formation in the gas phase, strongly affecting concentrations of small and large particles, and secondary formation via aqueous phase processes, leading to increased concentrations in accumulation mode particles mainly.For both sources, an anthropogenic influence is apparent from the back-trajectory analysis, either by providing the precursors for DCA formation, or by triggering enhanced biogenic SOA formation, e.g. through increased oxidant levels.The relative importance of these sources will be different in different seasons with photochemical gas phase formation being most important in summer.How it differs between different sites likely depends on the nature of precursors and their possibly complex interactions and will need to be addressed in future studies.Compared to the secondary sources, primary emissions of DCAs from traffic, biomass burning, sea salt and soil are considered to contribute much less or even nothing at all to the concentrations of DCAs at the sites studied.The results of this study suggest that beyond the well-known importance of aqueous phase formation, gas phase formation of LMW-DCAs can be important under conditions of high photochemistry in anthropogenically influenced air masses as well.

Fig. 1 .
Fig. 1.Mean mass size distributions of DCAs at all sites.Error bars indicate the 15th and 85th percentiles as a measure of data variability.

Fig. 2 .
Fig. 2. Box plots of residence time indices and further parameters of back trajectory analysis at different sampling sites.Bold line represents median, boxes lower and upper quartiles, whiskers 1.5 times the interquartile range, and dots indicate values outside this range.

Fig. 3 .
Fig. 3. Scree plot of PCA including results from parallel analysis.

Table 1 .
Concentrations of DCAs determined in this work and literature data from comparable European rural and urban sites.All data given in ng m −3 as mean ± standard deviation (min-max).For Melpitz, Falkenberg and Goldlauter data mean and standard deviation were calculated in a weighted manner to account for the variable sampling times applied at these sites.
Abbreviations: na: not available, bdl: below detection limit, a sum of three C5 DCA isomers, b mean conc.for summer and winter, respectively, c estimated from graphical data representation, d sum of front-and backup filters e median conc.#: number of samples.