kcf-jackson/adtools

An R package for Automatic Differentiation

R package ‘ADtools’

Implements the forward-mode auto-differentiation for multivariate functions using the matrix-calculus notation from Magnus and Neudecker (1988). Two key features of the package are: (i) the package incorporates various optimisaton strategies to improve performance; this includes applying memoisation to cut down object construction time, using sparse matrix representation to save derivative calculation, and creating specialised matrix operations with Rcpp to reduce computation time; (ii) the package supports differentiating random variable with respect to their parameters, targetting MCMC (and in general simulation-based) applications.

Installation

devtools::install_github("kcf-jackson/ADtools")

---

Notation

Given a function , where , the Jacobina matrix of w.r.t. is given by

---

Example 1. Matrix multiplication

Function definition

Consider where is a matrix, and is a vector.

library(ADtools)
f <- function(X, y) X %*% y
X <- randn(2, 2)
y <- matrix(c(1, 1))
print(list(X = X, y = y, f = f(X, y)))

## $X
##            [,1]       [,2]
## [1,] 0.41331040  0.7085659
## [2,] 0.01066195 -1.2300747
## 
## $y
##      [,1]
## [1,]    1
## [2,]    1
## 
## $f
##           [,1]
## [1,]  1.121876
## [2,] -1.219413

Auto-differentiation

Since has dimension (2, 2) and has dimension (2, 1), the input space has dimension , and the output has dimension , i.e. maps to and the Jacobian of should be .

# Full Jacobian matrix
f_AD <- auto_diff(f, at = list(X = X, y = y))
f_AD@dx   # returns a Jacobian matrix

##            d_X1 d_X2 d_X3 d_X4       d_y1       d_y2
## d_output_1    1    0    1    0 0.41331040  0.7085659
## d_output_2    0    1    0    1 0.01066195 -1.2300747

auto_diff also supports computing a partial Jacobian matrix. For instance, suppose we are only interested in the derivative w.r.t. y, then we can run

f_AD <- auto_diff(f, at = list(X = X, y = y), wrt = "y")
f_AD@dx   # returns a partial Jacobian matrix

##                  d_y1       d_y2
## d_output_1 0.41331040  0.7085659
## d_output_2 0.01066195 -1.2300747

Finite-differencing

It is good practice to always check the result with finite-differencing. This can be done by calling finite_diff which has the same interface as auto_diff.

f_FD <- finite_diff(f, at = list(X = X, y = y))
f_FD

##            d_X1 d_X2 d_X3 d_X4       d_y1       d_y2
## d_output_1    1    0    1    0 0.41331039  0.7085659
## d_output_2    0    1    0    1 0.01066194 -1.2300747

---

Example 2. Estimating a linear regression model

Simulate data from

set.seed(123)
n <- 1000
p <- 3
X <- randn(n, p)
beta <- randn(p, 1)
y <- X %*% beta + rnorm(n)

Inference with gradient descent

gradient_descent <- function(f, vary, fix, learning_rate = 0.01, tol = 1e-6, show = F) {
  repeat {
    df <- auto_diff(f, at = append(vary, fix), wrt = names(vary))
    if (show) print(df@x)
    delta <- learning_rate * as.numeric(df@dx)
    vary <- relist(unlist(vary) - delta, vary)
    if (max(abs(delta)) < tol) break
  }
  vary
}

lm_loss <- function(y, X, beta) sum((y - X %*% beta)^2)

# Estimate
gradient_descent(
  f = lm_loss, vary = list(beta = rnorm(p, 1)), fix = list(y = y, X = X),  learning_rate = 1e-4
) 

## $beta
## [1] -0.1417494 -0.3345771 -1.4484226

# Truth
t(beta)

##            [,1]       [,2]      [,3]
## [1,] -0.1503075 -0.3277571 -1.448165

---

Example 3. Sensitivity analysis of MCMC algorithms

Simulate data from

set.seed(123)
n <- 30  # small data
p <- 10
X <- randn(n, p)
beta <- randn(p, 1)
y <- X %*% beta + rnorm(n)

Estimating a Bayesian linear regression model

Inference using Gibbs sampler

gibbs_gaussian <- function(X, y, b_0, B_0, alpha_0, delta_0, num_steps = 1e4) {
  # Initialisation
  init_sigma <- 1 / sqrt(rgamma0(1, alpha_0 / 2, scale = 2 / delta_0))
  
  n <- length(y)
  alpha_1 <- alpha_0 + n
  sigma_g <- init_sigma
  inv_B_0 <- solve(B_0)
  inv_B_0_times_b_0 <- inv_B_0 %*% b_0
  XTX <- crossprod(X)
  XTy <- crossprod(X, y)
  beta_res <- vector("list", num_steps)
  sigma_res <- vector("list", num_steps)

  pb <- txtProgressBar(1, num_steps, style = 3)
  for (i in 1:num_steps) {
    # Update beta
    B_g <- solve(sigma_g^(-2) * XTX + inv_B_0)
    b_g <- B_g %*% (sigma_g^(-2) * XTy + inv_B_0_times_b_0)
    beta_g <- t(rmvnorm0(1, b_g, B_g))

    # Update sigma
    delta_g <- delta_0 + sum((y - X %*% beta_g)^2)
    sigma_g <- 1 / sqrt(rgamma0(1, alpha_1 / 2, scale = 2 / delta_g))

    # Keep track
    beta_res[[i]] <- beta_g
    sigma_res[[i]] <- sigma_g
    setTxtProgressBar(pb, i)
  }

  list(sigma = sigma_res, beta = beta_res)
}

Auto-differentiation

gibbs_deriv <- auto_diff(
  gibbs_gaussian,
  at = list(
    b_0 = numeric(p), B_0 = diag(p), alpha_0 = 4, delta_0 = 4,
    X = X, y = y, num_steps = 5000
  ),
  wrt = c("b_0", "B_0", "alpha_0", "delta_0")
)

Computing the sensitivity of the posterior mean of w.r.t. all the prior hyperparameters

library(magrittr)
library(knitr)
library(kableExtra)

matrix_ls_to_array <- function(x) {
  structure(unlist(x), dim = c(dim(x[[1]]), length(x)), dimnames = dimnames(x[[1]]))
}

tidy_mcmc <- function(mcmc_res, var0) {
  mcmc_res[[var0]] %>% 
    purrr::map(~.x@dx) %>% 
    matrix_ls_to_array()
}

tidy_table <- function(x) {
  x %>% kable() %>% kable_styling() %>% scroll_box(width = "100%")
}

posterior_Jacobian <- apply(tidy_mcmc(gibbs_deriv, "beta"), c(1,2), mean) 
tidy_table(posterior_Jacobian)

	d_b_01	d_b_02	d_b_03	d_b_04	d_b_05	d_b_06	d_b_07	d_b_08	d_b_09	d_b_010	d_B_01	d_B_02	d_B_03	d_B_04	d_B_05	d_B_06	d_B_07	d_B_08	d_B_09	d_B_010	d_B_011	d_B_012	d_B_013	d_B_014	d_B_015	d_B_016	d_B_017	d_B_018	d_B_019	d_B_020	d_B_021	d_B_022	d_B_023	d_B_024	d_B_025	d_B_026	d_B_027	d_B_028	d_B_029	d_B_030	d_B_031	d_B_032	d_B_033	d_B_034	d_B_035	d_B_036	d_B_037	d_B_038	d_B_039	d_B_040	d_B_041	d_B_042	d_B_043	d_B_044	d_B_045	d_B_046	d_B_047	d_B_048	d_B_049	d_B_050	d_B_051	d_B_052	d_B_053	d_B_054	d_B_055	d_B_056	d_B_057	d_B_058	d_B_059	d_B_060	d_B_061	d_B_062	d_B_063	d_B_064	d_B_065	d_B_066	d_B_067	d_B_068	d_B_069	d_B_070	d_B_071	d_B_072	d_B_073	d_B_074	d_B_075	d_B_076	d_B_077	d_B_078	d_B_079	d_B_080	d_B_081	d_B_082	d_B_083	d_B_084	d_B_085	d_B_086	d_B_087	d_B_088	d_B_089	d_B_090	d_B_091	d_B_092	d_B_093	d_B_094	d_B_095	d_B_096	d_B_097	d_B_098	d_B_099	d_B_0100	d_alpha_01	d_delta_01
d_output_1	0.0569887	0.0240169	-0.0085018	-0.0114905	-0.0002056	-0.0099043	0.0069673	-0.0250403	-0.0138359	0.0184303	-0.0441342	-0.0186824	0.0065791	0.0089195	0.0001676	0.0077123	-0.0054299	0.0194540	0.0107661	-0.0143304	-0.0422326	-0.0177185	0.0062827	0.0085277	0.0001388	0.0073738	-0.0051957	0.0186088	0.0102891	-0.0137093	-0.0548737	-0.0230575	0.0082968	0.0110663	0.0001859	0.0095235	-0.0066902	0.0241129	0.0132928	-0.0177620	-0.0718000	-0.0302760	0.0107191	0.0145789	0.0002573	0.0124716	-0.0087673	0.0315428	0.0174258	-0.0232140	-0.0156627	-0.0065633	0.0023708	0.0031359	0.0001308	0.0027179	-0.0019058	0.0068750	0.0037912	-0.0050542	0.0418913	0.0176076	-0.0062615	-0.0084694	-0.0001351	-0.0071959	0.0051419	-0.0184411	-0.0101835	0.0135673	-0.1016804	-0.0428326	0.0151174	0.0204964	0.0003181	0.0176513	-0.0123461	0.0446991	0.0246896	-0.0329021	0.0099753	0.0042146	-0.0014552	-0.0019818	-0.0000670	-0.0017475	0.0011945	-0.0043029	-0.0024432	0.0032439	0.0687012	0.0289268	-0.0102440	-0.0138316	-0.0001937	-0.0119421	0.0084227	-0.0302179	-0.0166070	0.0222462	0.1152496	0.0486374	-0.0171780	-0.0232862	-0.0003697	-0.0199968	0.0140737	-0.0506390	-0.0280183	0.0373722	-0.0018965	0.0016811
d_output_2	0.0240377	0.0904281	0.0127987	-0.0154678	0.0011496	-0.0297209	0.0192483	-0.0162636	-0.0265724	0.0235886	-0.0183982	-0.0700067	-0.0099287	0.0119620	-0.0008713	0.0230590	-0.0149584	0.0126714	0.0206031	-0.0183164	-0.0177858	-0.0670336	-0.0095958	0.0114911	-0.0009177	0.0222430	-0.0144052	0.0121760	0.0198080	-0.0176238	-0.0232262	-0.0869873	-0.0120353	0.0149244	-0.0011250	0.0286433	-0.0185164	0.0156581	0.0255766	-0.0227551	-0.0303404	-0.1139404	-0.0160980	0.0197683	-0.0014266	0.0374066	-0.0242106	0.0204458	0.0334838	-0.0296973	-0.0066046	-0.0247385	-0.0034116	0.0041890	-0.0001068	0.0081420	-0.0052474	0.0044657	0.0072442	-0.0064331	0.0176188	0.0663741	0.0094205	-0.0114027	0.0009022	-0.0216447	0.0141864	-0.0120085	-0.0195519	0.0173548	-0.0428613	-0.1613073	-0.0230086	0.0275584	-0.0021971	0.0529932	-0.0341142	0.0290588	0.0474033	-0.0421258	0.0040008	0.0157923	0.0023707	-0.0025788	0.0001263	-0.0052141	0.0032759	-0.0025446	-0.0046489	0.0041278	0.0289120	0.1089914	0.0154794	-0.0185779	0.0015487	-0.0358662	0.0232800	-0.0196546	-0.0318409	0.0285183	0.0487821	0.1830006	0.0259046	-0.0314158	0.0024403	-0.0600001	0.0388629	-0.0329070	-0.0538610	0.0480035	-0.0051331	0.0045483
d_output_3	-0.0085091	0.0128582	0.0606947	0.0018708	0.0093748	-0.0038372	-0.0124373	0.0108905	-0.0013585	0.0035816	0.0066996	-0.0100238	-0.0470345	-0.0014486	-0.0072611	0.0030017	0.0096190	-0.0084236	0.0010817	-0.0028071	0.0064625	-0.0094615	-0.0454735	-0.0014125	-0.0070568	0.0028694	0.0093231	-0.0081605	0.0009901	-0.0026940	0.0081889	-0.0123584	-0.0585021	-0.0017899	-0.0090490	0.0037064	0.0120127	-0.0105169	0.0013044	-0.0034779	0.0106775	-0.0162045	-0.0763825	-0.0022640	-0.0117940	0.0048318	0.0156772	-0.0137081	0.0017181	-0.0045021	0.0023308	-0.0035036	-0.0166668	-0.0005309	-0.0025077	0.0010566	0.0034371	-0.0029829	0.0003518	-0.0009703	-0.0063362	0.0093897	0.0447738	0.0013822	0.0069415	-0.0027284	-0.0091884	0.0080518	-0.0009880	0.0026380	0.0152430	-0.0228882	-0.1084255	-0.0033729	-0.0168030	0.0068048	0.0223176	-0.0194609	0.0023904	-0.0063921	-0.0015991	0.0022439	0.0107478	0.0003773	0.0016300	-0.0006714	-0.0022394	0.0020392	-0.0002404	0.0006235	-0.0103342	0.0154750	0.0733211	0.0022912	0.0113836	-0.0046196	-0.0150125	0.0131633	-0.0015559	0.0043453	-0.0171223	0.0260386	0.1225932	0.0037405	0.0189755	-0.0077206	-0.0251427	0.0219994	-0.0027912	0.0073362	-0.0016336	0.0014212
d_output_4	-0.0115148	-0.0155222	0.0018741	0.0538145	-0.0020646	0.0030748	-0.0087192	0.0141244	0.0141004	-0.0121362	0.0090350	0.0119581	-0.0015009	-0.0416657	0.0016067	-0.0023328	0.0067214	-0.0109202	-0.0108953	0.0093877	0.0087038	0.0117329	-0.0014675	-0.0402955	0.0015294	-0.0022764	0.0065180	-0.0105615	-0.0105698	0.0090910	0.0111127	0.0150710	-0.0017092	-0.0520296	0.0019920	-0.0029658	0.0084436	-0.0136627	-0.0136540	0.0117369	0.0144505	0.0195160	-0.0023350	-0.0675164	0.0026036	-0.0038820	0.0109700	-0.0177700	-0.0177239	0.0152723	0.0031468	0.0042670	-0.0004861	-0.0148352	0.0006424	-0.0008148	0.0024087	-0.0038815	-0.0038776	0.0033266	-0.0085681	-0.0115627	0.0013937	0.0396951	-0.0015198	0.0023721	-0.0064352	0.0104493	0.0104159	-0.0090011	0.0206229	0.0278116	-0.0034118	-0.0961119	0.0036398	-0.0055483	0.0156814	-0.0252497	-0.0252152	0.0216987	-0.0021391	-0.0027703	0.0003903	0.0095449	-0.0003900	0.0005380	-0.0015780	0.0026081	0.0025117	-0.0021678	-0.0139711	-0.0188096	0.0022905	0.0650464	-0.0024392	0.0037289	-0.0105230	0.0170629	0.0171256	-0.0146613	-0.0232066	-0.0313395	0.0037780	0.1086399	-0.0041298	0.0062744	-0.0176437	0.0285431	0.0284539	-0.0244253	-0.0016562	0.0014197
d_output_5	-0.0002015	0.0011602	0.0093305	-0.0021301	0.0416820	0.0063474	-0.0124201	-0.0084977	0.0127248	0.0140695	0.0000392	-0.0009037	-0.0072236	0.0016548	-0.0322917	-0.0049382	0.0096484	0.0065554	-0.0098702	-0.0108998	0.0000799	-0.0010397	-0.0070015	0.0015956	-0.0311980	-0.0047749	0.0093385	0.0063402	-0.0095232	-0.0105459	0.0002337	-0.0011615	-0.0091311	0.0019994	-0.0403015	-0.0061458	0.0119972	0.0082189	-0.0123143	-0.0135727	0.0002891	-0.0014201	-0.0117045	0.0025540	-0.0522549	-0.0079484	0.0155687	0.0106572	-0.0159636	-0.0176338	0.0000658	-0.0003473	-0.0025924	0.0005953	-0.0115158	-0.0017385	0.0033803	0.0023231	-0.0034806	-0.0038647	-0.0000968	0.0009752	0.0069060	-0.0015688	0.0307289	0.0045695	-0.0091735	-0.0062613	0.0093726	0.0104121	0.0003115	-0.0021644	-0.0166177	0.0038501	-0.0743845	-0.0112928	0.0220746	0.0151891	-0.0226929	-0.0251613	0.0000843	0.0002798	0.0016057	-0.0004489	0.0073900	0.0011070	-0.0021410	-0.0016478	0.0022166	0.0025215	-0.0001848	0.0014832	0.0112817	-0.0026156	0.0503344	0.0076649	-0.0150416	-0.0102730	0.0152763	0.0170167	-0.0004861	0.0022518	0.0188271	-0.0042536	0.0841183	0.0127844	-0.0250641	-0.0171571	0.0257523	0.0282801	0.0021382	-0.0018787
d_output_6	-0.0098928	-0.0297162	-0.0037695	0.0030480	0.0063663	0.0552486	-0.0143575	0.0052084	0.0107802	0.0029326	0.0074214	0.0230069	0.0029022	-0.0022937	-0.0049406	-0.0428277	0.0111680	-0.0040983	-0.0083541	-0.0022415	0.0072248	0.0217703	0.0027583	-0.0020931	-0.0046577	-0.0413635	0.0107642	-0.0039331	-0.0079740	-0.0022010	0.0096220	0.0285772	0.0033567	-0.0029425	-0.0061197	-0.0534077	0.0138521	-0.0050280	-0.0103800	-0.0028076	0.0124796	0.0373442	0.0047290	-0.0041209	-0.0080016	-0.0692467	0.0179884	-0.0065145	-0.0135446	-0.0036905	0.0027195	0.0080265	0.0009267	-0.0007802	-0.0019473	-0.0150756	0.0038894	-0.0014342	-0.0029022	-0.0008212	-0.0071938	-0.0216834	-0.0027394	0.0022030	0.0046396	0.0405590	-0.0106023	0.0038578	0.0079088	0.0022191	0.0175856	0.0529079	0.0068558	-0.0053461	-0.0112332	-0.0985946	0.0253945	-0.0093039	-0.0191869	-0.0052745	-0.0015026	-0.0050982	-0.0007503	0.0003666	0.0011963	0.0097106	-0.0024161	0.0006043	0.0018568	0.0005551	-0.0118376	-0.0357508	-0.0045558	0.0035470	0.0075299	0.0667830	-0.0173927	0.0063052	0.0127777	0.0035293	-0.0201816	-0.0602960	-0.0076860	0.0063388	0.0127583	0.1115422	-0.0289714	0.0105507	0.0219431	0.0056070	0.0048153	-0.0042268
d_output_7	0.0069446	0.0192320	-0.0124953	-0.0086532	-0.0124090	-0.0143473	0.0558417	-0.0146184	-0.0051602	-0.0035448	-0.0051272	-0.0148759	0.0097013	0.0066371	0.0096298	0.0111660	-0.0432957	0.0113787	0.0040087	0.0027347	-0.0050374	-0.0139463	0.0094648	0.0063540	0.0092342	0.0107642	-0.0418240	0.0109867	0.0038158	0.0026658	-0.0067943	-0.0184984	0.0123624	0.0084440	0.0120081	0.0139189	-0.0540180	0.0141399	0.0049925	0.0034015	-0.0087916	-0.0242142	0.0156719	0.0111644	0.0156054	0.0179926	-0.0700294	0.0183071	0.0065053	0.0044500	-0.0019214	-0.0051924	0.0035041	0.0023222	0.0036054	0.0039501	-0.0152366	0.0040064	0.0013745	0.0009905	0.0050195	0.0139367	-0.0093230	-0.0063717	-0.0091359	-0.0103301	0.0413188	-0.0108259	-0.0037585	-0.0026864	-0.0123397	-0.0342099	0.0222600	0.0153796	0.0220557	0.0255470	-0.0996328	0.0261613	0.0091591	0.0063636	0.0009729	0.0032416	-0.0021301	-0.0013462	-0.0022680	-0.0025112	0.0097469	-0.0022555	-0.0008595	-0.0006834	0.0082659	0.0230507	-0.0151265	-0.0103304	-0.0148400	-0.0173030	0.0675348	-0.0176816	-0.0059760	-0.0042830	0.0142285	0.0391065	-0.0251826	-0.0176514	-0.0249597	-0.0289096	0.1127846	-0.0295681	-0.0105792	-0.0068367	-0.0050240	0.0044264
d_output_8	-0.0250804	-0.0161938	0.0109440	0.0141087	-0.0084907	0.0051460	-0.0146053	0.0673064	0.0047617	-0.0103918	0.0193901	0.0123672	-0.0085698	-0.0107848	0.0066224	-0.0038823	0.0111201	-0.0522336	-0.0036282	0.0080575	0.0187944	0.0119589	-0.0082823	-0.0105283	0.0064342	-0.0038049	0.0108549	-0.0505127	-0.0035369	0.0078282	0.0242659	0.0156709	-0.0106651	-0.0137003	0.0081853	-0.0050257	0.0142004	-0.0650318	-0.0046159	0.0100563	0.0314683	0.0204053	-0.0137138	-0.0178233	0.0105922	-0.0065138	0.0184283	-0.0843316	-0.0060171	0.0130294	0.0068385	0.0043584	-0.0030011	-0.0038238	0.0022621	-0.0013739	0.0038935	-0.0183673	-0.0012828	0.0028227	-0.0185725	-0.0119098	0.0081547	0.0104460	-0.0063207	0.0037169	-0.0108092	0.0498851	0.0035081	-0.0076994	0.0449804	0.0289774	-0.0196031	-0.0252719	0.0153021	-0.0091713	0.0260506	-0.1207382	-0.0085237	0.0186647	-0.0042959	-0.0026810	0.0019020	0.0023368	-0.0014871	0.0008146	-0.0023603	0.0116123	0.0007876	-0.0018047	-0.0303050	-0.0194974	0.0132639	0.0170032	-0.0103353	0.0061868	-0.0176369	0.0814000	0.0056628	-0.0125827	-0.0507465	-0.0328884	0.0220748	0.0286126	-0.0171680	0.0104383	-0.0296133	0.1359903	0.0097023	-0.0211098	0.0015298	-0.0013439
d_output_9	-0.0138377	-0.0266094	-0.0013464	0.0141003	0.0127371	0.0108331	-0.0052421	0.0048221	0.0530128	-0.0142592	0.0106016	0.0204647	0.0009927	-0.0108144	-0.0098145	-0.0083304	0.0039921	-0.0038836	-0.0410147	0.0110594	0.0103596	0.0196407	0.0009328	-0.0104557	-0.0094673	-0.0080459	0.0037988	-0.0038664	-0.0396832	0.0107345	0.0134222	0.0256635	0.0011410	-0.0136421	-0.0123035	-0.0104722	0.0050322	-0.0046676	-0.0512514	0.0138072	0.0173695	0.0334742	0.0017008	-0.0178857	-0.0160227	-0.0136459	0.0066662	-0.0059179	-0.0664936	0.0178668	0.0037365	0.0072135	0.0003140	-0.0037982	-0.0036054	-0.0029823	0.0015032	-0.0012083	-0.0144189	0.0038367	-0.0102321	-0.0195538	-0.0009581	0.0104256	0.0093853	0.0078496	-0.0038132	0.0036739	0.0392466	-0.0105759	0.0247708	0.0475729	0.0024715	-0.0252144	-0.0227486	-0.0193319	0.0092072	-0.0086815	-0.0949740	0.0255564	-0.0021981	-0.0043490	-0.0002609	0.0022409	0.0021664	0.0017800	-0.0007949	0.0006660	0.0088741	-0.0023693	-0.0167036	-0.0320287	-0.0016078	0.0169591	0.0152770	0.0130280	-0.0062642	0.0059658	0.0639315	-0.0172731	-0.0280732	-0.0539748	-0.0027756	0.0286292	0.0256939	0.0218936	-0.0106074	0.0097172	0.1072256	-0.0290052	0.0027076	-0.0023674
d_output_10	0.0183998	0.0235696	0.0035820	-0.0121090	0.0140323	0.0028750	-0.0034804	-0.0103564	-0.0142733	0.0604927	-0.0145180	-0.0181169	-0.0026837	0.0092915	-0.0108873	-0.0023307	0.0028632	0.0081400	0.0110213	-0.0467783	-0.0140121	-0.0179260	-0.0025975	0.0089999	-0.0105325	-0.0022815	0.0028046	0.0079632	0.0107114	-0.0453199	-0.0177544	-0.0229244	-0.0036911	0.0116933	-0.0135718	-0.0027779	0.0033604	0.0100647	0.0138165	-0.0584723	-0.0229918	-0.0295581	-0.0045324	0.0150177	-0.0176017	-0.0035554	0.0042804	0.0129093	0.0178981	-0.0758737	-0.0050233	-0.0064541	-0.0010407	0.0033427	-0.0040061	-0.0008135	0.0009901	0.0029355	0.0039354	-0.0164856	0.0137477	0.0177026	0.0026546	-0.0089431	0.0103514	0.0019110	-0.0025516	-0.0076565	-0.0106108	0.0448584	-0.0330814	-0.0424295	-0.0063049	0.0217503	-0.0250430	-0.0050663	0.0060199	0.0185659	0.0256592	-0.1084753	0.0033399	0.0040222	0.0004825	-0.0021145	0.0024388	0.0005097	-0.0005756	-0.0020969	-0.0024352	0.0102202	0.0223668	0.0286054	0.0042865	-0.0146718	0.0168359	0.0034689	-0.0042721	-0.0125683	-0.0174525	0.0730667	0.0369782	0.0475096	0.0072398	-0.0244064	0.0282180	0.0056742	-0.0068803	-0.0207674	-0.0287255	0.1219028	0.0038724	-0.0033642