1 2 3 4 5 The supplementary materials contain 7 figures and a discussion on the relationship between residual cross-correlations and those of the innovations. 6 preserved by the residuals. 7 Conditions under which the cross-correlations between the innovations are asymptotically Let ππ,π‘ be the downstream measurement of the πth variable at time π‘, πΏπ,π‘ the (π + 1)- 8 dimensional vector covariates that include the corresponding upstream measurement ππ,π‘ and its 9 lags, i.e., πΏπ,π‘ = (ππ,π‘ , ππ,π‘−1 , … , ππ,π‘−π ) , and that the following stochastic regression models β€ 10 hold (with the intercept and outliers omitted for simplicity): 11 ππ,π‘ = π·β€ π ππ,π‘ + ππ,π‘ , 12 where π = 1, … , π + 1, π·π is a (π + 1)-dimensional coefficient vector and the regression errors 13 ππ,π‘ follow an order π autoregressive process, i.e. 14 ππ,π‘ = ππ,1 ππ,π‘−1 + β― + ππ,π ππ,π‘−π + ππ,π‘ , 15 with the π’s being independent and identically distributed random variables of zero mean and 16 positive variance ππ2 , and they are independent of past π’s and current and past π’s. The π’s are 17 known as the innovations, and ππ,π‘ may be used as the proxy for the measurement of the πth 18 variable due to the processes occurring within the reservoir in the π‘th period. The main object of 19 study consists of, say, regressing ππ+1,π‘ on ππ,π‘ , π = 1, … , π. Since the π’s are unobservable, they 20 will be replaced by the residuals, denoted by πΜπ,π‘ , obtained from fitting the model defined by 21 (A1) and (A2), separately for each the π + 1 variables. We shall assume that the unknown 22 parameters can be consistently estimated with the estimation error being of order ππ (1/√π) , 23 Μ π , π = 1, β― , π + 1 and where π is the sample size and the coefficient estimates are denoted by π· 24 πΜπ,β , β = 1, β― , π. For instance, this holds if the model defined by (A1) and (A2) is estimated by (A1) (A2) 1 25 the method of conditional least squares, under some mild regularity conditions (Klimko and 26 Nelson, 1978). We show below that the large-sample joint distribution of the regression 27 coefficient estimates using the residuals is identical to that using the latent innovations under the 28 following additional conditions: 29 (C1) The vector process {(ππ,π‘ , πΏβ€ π,π‘ , π = 1, … , π + 1) , π‘ = β― , −1, 0,1,2, β― } is a stationary, 30 ergodic process with finite second moments. 31 We show the validity of the claim by proving below that 32 ∑ππ‘=π+1 πΜπ,π‘ πΜπ,π‘ = ∑ππ‘=π+1 ππ,π‘ ππ,π‘ + ππ (1). 33 The residuals πΜπ,π‘ are computed by the following two equations where 34 Μβ€ ππ,π‘ = π· π πΏπ,π‘ + πΜπ,π‘ , (A4) 35 πΜπ,π‘ = πΜπ,1 πΜπ,π‘−1 + β― + πΜπ,π πΜπ,π‘−π + πΜπ,π‘ , (A5) 36 After some algebra, it can be shown that 37 Μ π )β€ (πΏπ,π‘ − πΜπ,1 πΏπ,π‘−1 − β― − πΜπ,π πΏπ,π‘−π ) + (ππ,1 − πΜπ,1 )ππ,π‘−1 + β― πΜπ,π‘ = ππ,π‘ + (π·π − π· β€ (A3) + (ππ,π − πΜπ,π )ππ,π‘−π . 38 39 Hence, ∑ππ‘=π+1 πΜπ,π‘ πΜπ,π‘ = ∑ππ‘=π+1 ππ,π‘ ππ,π‘ + ππ (1), because, for instance, with 1 ≤ β, π ≤ π , 40 Μ π ) ∑ππ‘=π+1 ππ,π‘ πΏπ,π‘−β = √π(π·π − π· Μ π ) ∑ππ‘=π+1 ππ,π‘ πΏπ,π‘−β /√π = ππ (1), (π·π − π· 41 (ππ,β − πΜπ,β )(ππ,π − πΜπ,π ) ∑ ππ,π‘−β ππ,π‘−π β€ β€ π π‘=π+1 42 = √π(ππ,β − πΜπ,β )√π(ππ,π − πΜπ,π ) ∑ππ‘=π+1 ππ,π‘−β ππ,π‘−π = ππ (1), π 2 43 as ∑ππ‘=π+1 ππ,π‘ πΏπ,π‘−β /√π is asymptotically normally distributed by the martingale central limit 44 theorem (Billingsley 2013, Theorem 18.1), and ∑ππ‘=π+1 ππ,π‘−β ππ,π‘−π /π approaches 45 πΈ(ππ,π‘−β ππ,π‘−π ), with increasing sample size π, in view of (C1). 46 47 48 49 50 51 Supplementary Figure captions: Figure S1. Adjusted measurements as residuals of fitted transfer function models. Figure S1. Model diagnostics of the transfer function model (6) fitted to the square root of N. 52 The top sub-figure is the residual time plot, the middle sub-figure the residual autocorrelation 53 function (ACF) and the bottom sub-figure shows the p-value of Ljung-Box test assessing the 54 whiteness of the residuals based on the first π lags of residual ACF, for π ranging from 13 to 24. 55 All p-values are greater than 5%, suggesting that the fitted transfer function model provides a 56 good fit to the N data. 57 Figure S3. Model diagnostics of the transfer function model (6) fitted to pH (P). Same 58 convention as in Supplementary Figure 1. The p-values are slightly higher than 5%, indicating an 59 adequate fit to the data. 60 Figure S4. Model diagnostics of the transfer function model (6) fitted to total alkalinity (A). 61 Same convention as in Supplementary Figure 1. The Ljung-Box tests suggest the model provides 62 a good fit to the data. 63 Figure S5. Model diagnostics of the transfer function model (6) fitted to total hardness (H). 64 Same convention as in Supplementary Figure 1. The Ljung-Box tests suggest the model provides 65 a good fit to the data. 66 Figure S6. Model diagnostics of the transfer function model (6) fitted to the natural log of TSS. 67 Same convention as in Supplementary Figure 1. The Ljung-Box tests suggest the model provides 68 a good fit to the data. 3 69 Figure S7. Model diagnostics for the final model (8) fitted with all data. 70 71 Figure S1 72 Figure 4 73 74 75 4 76 Figure S2 77 78 5 79 Figure S3 80 81 6 82 Figure S4 83 84 7 85 Figure S5 86 87 8 88 Figure S6 89 90 91 9 92 93 Figure S7 94 10