diff --git a/README.org b/README.org
index f5f1930..edfa733 100644
--- a/README.org
+++ b/README.org
@@ -1,10 +1,11 @@
-Figures for Labeled Optimal PARTitioning paper
+Figures for Labeled Optimal PARTitioning paper 
 
 - Figure 1: example data and cost computation. [[file:figure-signal-cost.R][R script]], [[file:figure-signal-cost-standAlone.pdf][standAlone
   pdf]], [[file:figure-signal-cost.tex][tex for inclusion in paper]].
 - Figure 2: timings. [[file:figure-timings.R][R script]], tex for inclusion in paper: [[file:figure-timings.tex][time vs
   number of data]], [[file:figure-timings-labels.tex][time vs number of labels]].
-- Figure 3: label error. [[file:figure-label-errors.R][R script]], [[file:figure-label-errors.pdf][OPART pdf]], [[file:figure-label-errors-SegAnnot.pdf][SegAnnot pdf]].
+- Figure 3: best case label error. [[file:figure-label-errors.R][R script]], [[file:figure-label-errors.pdf][OPART pdf]], [[file:figure-label-errors-SegAnnot.pdf][SegAnnot pdf]].
+- Figure 4: cross-validation label error. [[file:figure-cv.R][R script]], [[file:figure-cv.pdf][pdf]].
 
 Reproducibility: type "make" in the shell.
 
diff --git a/figure-cv.R b/figure-cv.R
new file mode 100644
index 0000000..a050530
--- /dev/null
+++ b/figure-cv.R
@@ -0,0 +1,61 @@
+source("packages.R")
+
+err.dt <- data.table(
+  csv=Sys.glob("figure-label-errors-data/*.csv")
+)[, data.table::fread(
+  csv,
+  colClasses=list(character=5)
+), by=csv]
+err.dt[model.name=="LOPART" & set=="train", table(errors)]
+err.dt[model.name=="LOPART" & set=="train" & 0<errors, .(
+  csv, test.fold, set, penalty, fp, fn)]
+
+err.train <- err.dt[model.name=="OPART" & set=="train", .(
+  train.errors=sum(errors)
+), by=.(test.fold, penalty)]
+best.penalty <- err.train[, .SD[which.min(train.errors)], by=test.fold]
+err.test <- err.dt[set=="test"]
+err.pred <- err.test[best.penalty, on=.(test.fold, penalty)]
+
+(fold.model.err.tall <- err.pred[, .(
+  test.errors=sum(errors),
+  test.labels=sum(labels)
+), by=.(test.fold, model.name)])
+fold.model.err.tall[, prop.errors := test.errors / test.labels]
+fold.model.err.wide <- dcast(
+  fold.model.err.tall,
+  test.fold ~ model.name,
+  value.var = "prop.errors")
+fold.model.err.wide[, diff := OPART-LOPART]
+fold.model.err.wide
+
+prob.err.wide <- dcast(
+  err.pred,
+  test.fold + sequenceID ~ model.name,
+  value.var = "errors")
+prob.err.wide[, diff := OPART-LOPART]
+prob.err.wide.counts <- prob.err.wide[, .(
+  count=.N
+), keyby=.(test.fold, diff)]
+
+gg <- ggplot()+
+  geom_tile(aes(
+    diff, factor(test.fold),
+    fill=log10(count)),
+    data=prob.err.wide.counts)+
+  coord_equal()+
+  scale_fill_gradient(
+    "log10(sequences)",
+    low="white",
+    high="orange")+
+  theme_bw()+ 
+  geom_text(aes(
+    diff, factor(test.fold),
+    label=count),
+    data=prob.err.wide.counts)+
+  ylab("Test fold")+
+  xlab("Difference of incorrectly predicted
+labels in test set (OPART-LOPART)")
+pdf("figure-cv.pdf", width=4, height=2)
+print(gg)
+dev.off()
diff --git a/figure-cv.pdf b/figure-cv.pdf
new file mode 100644
index 0000000..1e890ba
Binary files /dev/null and b/figure-cv.pdf differ