diff --git a/DESCRIPTION b/DESCRIPTION
index 4cb21d8..28c6f0f 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -25,7 +25,7 @@ License: GPL-3 + file LICENSE
 URL: https://github.com/christophrust/dpca
 Encoding: UTF-8
 LazyData: true
-RoxygenNote: 7.3.1
+RoxygenNote: 7.3.2
 Suggests:
     testthat (>= 3.0.0),
     freqdom (>= 2.0.3),
diff --git a/R/dpca.R b/R/dpca.R
index 9e68c9e..57ed4f8 100644
--- a/R/dpca.R
+++ b/R/dpca.R
@@ -20,8 +20,8 @@
 #' spectral density is evaluated.
 #'
 #' @param bandwidth Single integer, giving the width of the
-#' lag window estimator. If unspecified, the cube root of the number of time
-#' observations is used as default.
+#' lag window estimator. If unspecified, 0.75 times the square root of the number
+#' of time observations is used as default.
 #'
 #' @param weights Kernel used for the lag window estimation of spectrum.
 #'
@@ -30,6 +30,7 @@
 #'
 #' @param q_max Maximum numer of dynamic factors considered in the data-driven
 #' selection.
+#'
 #' @param n_path Integer vector specifying which (nested) subsets of the
 #' cross section are used in the Hallin & Liska procedure. If unspecified,
 #' a regular sequence of length \code{20} from \code{n/2} to \code{n} is used.
@@ -110,23 +111,24 @@
 dpca <- function(
     x,
     q,
-    freqs = -20:20 / 20 * pi,
-    bandwidth = floor(ncol(x)^(1 / 3)),
+    freqs = -bandwidth:bandwidth / bandwidth * pi,
+    bandwidth = floor(0.75 * sqrt(tx)),
     weights = c(
       "bartlett", "trunc", "tukey", "parzen",
       "bohman", "daniell", "parzen_cogburn_davis"
     ),
     qsel_crit = c("IC1", "IC2"),
     q_max = 15,
-    n_path = floor(seq(nrow(x) / 2, nrow(x), nrow(x) / 20)),
+    n_path = floor(seq(nx / 2, nx, nx / 20)),
     penalties = (
-      bandwidth^(-2) + sqrt(bandwidth / ncol(x)) + 1 / n_path
-    ) * log(pmin(n_path, bandwidth^2, sqrt(ncol(x) / bandwidth))),
+      bandwidth^(-2) + sqrt(bandwidth / tx) + 1 / n_path
+    ) * log(pmin(n_path, bandwidth^2, sqrt(tx / bandwidth))),
     penalty_scales = seq(0, 2, by = 0.01)) {
   if (length(weights) > 1) {
     weights <- "bartlett"
   }
 
+  ## our internal object is n times t
   x <- if (is.ts(x) || "zoo" %in% class(x)) {
     t(x)
   } else if (is.matrix(x)) {
@@ -134,6 +136,8 @@ dpca <- function(
   } else {
     stop("x must either a \"ts\" or \"zoo\" object or a matrix!")
   }
+  nx <- nrow(x)
+  tx <- ncol(x)
 
   if (!missing(q) && (length(q) > 1 || floor(abs(q)) != q)) {
     stop("\"q\" has to be a single positive integer!")
@@ -186,7 +190,7 @@ dpca <- function(
     as.integer(q_max),
     select_q,
     as.integer(n_path),
-    as.integer(rep(ncol(x), length(n_path))),
+    as.integer(rep(tx, length(n_path))),
     as.numeric(penalties),
     as.numeric(penalty_scales),
     PACKAGE = "dpca"
diff --git a/man/dpca.Rd b/man/dpca.Rd
index 6c73944..af49575 100644
--- a/man/dpca.Rd
+++ b/man/dpca.Rd
@@ -7,15 +7,15 @@
 dpca(
   x,
   q,
-  freqs = -20:20/20 * pi,
-  bandwidth = floor(ncol(x)^(1/3)),
+  freqs = -bandwidth:bandwidth/bandwidth * pi,
+  bandwidth = floor(0.75 * sqrt(tx)),
   weights = c("bartlett", "trunc", "tukey", "parzen", "bohman", "daniell",
     "parzen_cogburn_davis"),
   qsel_crit = c("IC1", "IC2"),
   q_max = 15,
-  n_path = floor(seq(nrow(x)/2, nrow(x), nrow(x)/20)),
-  penalties = (bandwidth^(-2) + sqrt(bandwidth/ncol(x)) + 1/n_path) * log(pmin(n_path,
-    bandwidth^2, sqrt(ncol(x)/bandwidth))),
+  n_path = floor(seq(nx/2, nx, nx/20)),
+  penalties = (bandwidth^(-2) + sqrt(bandwidth/tx) + 1/n_path) * log(pmin(n_path,
+    bandwidth^2, sqrt(tx/bandwidth))),
   penalty_scales = seq(0, 2, by = 0.01)
 )
 }
@@ -31,8 +31,8 @@ it is chosed data-driven by using the criterion of Hallin & Liska (2007).}
 spectral density is evaluated.}
 
 \item{bandwidth}{Single integer, giving the width of the
-lag window estimator. If unspecified, the cube root of the number of time
-observations is used as default.}
+lag window estimator. If unspecified, 0.75 times the square root of the number
+of time observations is used as default.}
 
 \item{weights}{Kernel used for the lag window estimation of spectrum.}