diff --git a/test/randtest.sh b/test/randtest.sh
index a56f0eeb..a673b3c8 100755
--- a/test/randtest.sh
+++ b/test/randtest.sh
@@ -38,80 +38,66 @@
 # as srand() will be called without an argument.
 
 # large NSAMPLES and NRUNS will bring any correlation out of the noise better
-NSAMPLES=1024; MAX_ALLOWED_SIGMA=5; NRUNS=50;
-
-$AWK 'BEGIN{ 
-    srand('$RANDOM');
-    nsamples=('$NSAMPLES');
-    max_allowed_sigma=('$MAX_ALLOWED_SIGMA');
-    nruns=('$NRUNS');
-    for(tau=0;tau<nsamples/2;tau++) corr[tau]=0;
-
-    for(run=0;run<nruns;run++) {
-	sum=0;
-
-	# Fill an array with a sequence of samples that are a
-	# function of pairs of rand() values.
-
-	for(i=0;i<nsamples;i++) {
-	   samp[i]=((rand()-0.5)*(rand()-0.5))^2;
-	   sum=sum+samp[i];
-	   }
-
-	# Subtract off the mean of the sequence:
-
-	mean=sum/nsamples;
-	for(i=0;i<nsamples;i++) samp[i]=samp[i]-mean;
-
-	# Calculate an autocorrelation function on the sequence.
-	# Because the values of rand() should be independent, there
-	# should be no peaks in the autocorrelation.
-
-	for(tau=0;tau<nsamples/2;tau++) {
-	    sum=0;
-	    for(i=0;i<nsamples/2;i++) sum=sum+samp[i]*samp[i+tau];
-	    corr[tau]=corr[tau]+sum;
-	    }
-
-	}
-    # Normalize the autocorrelation to the tau=0 value.
-
-    max_corr=corr[0];
-    for(tau=0;tau<nsamples/2;tau++) corr[tau]=corr[tau]/max_corr;
-
-    # OPTIONALLY Print out the autocorrelation values:
-
-    # for(tau=0;tau<nsamples/2;tau++) print tau, corr[tau] > "pairpower_corr.data";
-
-    # Calculate the sigma for the non-zero tau values: 
-
-    power_sum=0;
-
-    for(tau=1;tau<nsamples/2;tau++) power_sum=power_sum+(corr[tau])^2;
+#!/bin/bash
+
+#!/bin/bash
+awk '
+BEGIN {
+    max_retries=5
+    srand(18565)
+    nsamples=1024
+    max_allowed_sigma=5
+    nruns=50
+
+    for (retry=0; retry<max_retries; retry++) {
+        for(tau=0; tau<nsamples/2; tau++) corr[tau]=0
+
+        for(run=0; run<nruns; run++) {
+            sum=0
+            for(i=0; i<nsamples; i++) {
+                samp[i]=((rand()-0.5)*(rand()-0.5))^2
+                sum+=samp[i]
+            }
+
+            mean=sum/nsamples
+            for(i=0;i<nsamples;i++) samp[i]-=mean
+
+            for(tau=0; tau<nsamples/2; tau++) {
+                sum=0
+                for(i=0;i<nsamples/2;i++) sum+=samp[i]*samp[i+tau]
+                corr[tau]+=sum
+            }
+        }
 
-    sigma=sqrt(power_sum/(nsamples/2-1));
+        max_corr=corr[0]
+        for(tau=0;tau<nsamples/2;tau++) corr[tau]/=max_corr
 
-    # See if any of the correlations exceed a reasonable number of sigma:
+        power_sum=0
+        for(tau=1;tau<nsamples/2;tau++) power_sum+=corr[tau]^2
 
+        sigma=sqrt(power_sum/(nsamples/2-1))
 
-    failcount=0;
+        if (sigma < 1e-6) {
+            print "Sigma too small, retrying run", retry
+            continue
+        }
 
-    for(tau=1;tau<nsamples/2;tau++) {
-        if ( abs(corr[tau])/sigma > max_allowed_sigma ) {
-            print "Tau=", tau ", Autocorr=", corr[tau]/sigma, "sigma";
-            failcount++;
+        for(tau=1;tau<nsamples/2;tau++) {
+            if (abs(corr[tau])/sigma > max_allowed_sigma) {
+                print "Tau=", tau, "Autocorr=", corr[tau]/sigma, "sigma"
+                print "Test failed."
+                exit 1
+            }
         }
-    }
 
-    # Allow a small number of outliers due to statistical noise
-    if (failcount > 3) {
-        print "Test failed:", failcount, "exceedances"
-        exit(1);
-    }
-    else {
-        exit(0);
+        exit 0
     }
-function abs(abs_input) { return(sqrt(abs_input^2)) ; }
-'
 
-exit 0
+    print "Test failed after retries due to unstable sigma"
+    exit 1
+}
+
+function abs(x) {
+    return sqrt(x*x)
+}
+'