According to the EL spec ch.1.7.1,
the A+B,A-B,A*B expressions when A is a "String containing ., e, or E" and B is a BigInteger have to be evaluated as follows:

"coerce both A and B to BigDecimal and apply operator"

The actual implementation though, instead of coercing A as String -> BigDecimal, does two coercions:  String -> Double -> BigDecimal, which produces a different result.

To reproduce, run the following JSP page:
<%
  pageContext.setAttribute("a", "1.1");
  pageContext.setAttribute("b", new java.math.BigInteger("1000000000000000000000"));
  // "c" is used to explicitly coerce arguments to BigDecimal
  pageContext.setAttribute("c", new java.math.BigDecimal("0"));
%>${a + b}<br>
${(a+c) + (b+c)}<br>
${a - b}<br>
${(a+c) - (b+c)}<br>
${a * b}<br>
${(a+c) * (b+c)}

Actual result:
1000000000000000000001.100000000000000088817841970012523233890533447265625
1000000000000000000001.1
-999999999999999999998.899999999999999911182158029987476766109466552734375
-999999999999999999998.9
1100000000000000088817.841970012523233890533447265625000000000000000000000
1100000000000000000000.0 

Expected result:
1000000000000000000001.1
1000000000000000000001.1
-999999999999999999998.9
-999999999999999999998.9
1100000000000000000000.0 
1100000000000000000000.0 

I think it is a minor issue, though.
This is reproducible in the current tc6.0.x as well.

Maybe o.a.el.lang.ELArithmetic#add(Object, Object) should delegate to BIGDECIMAL instead of DOUBLE with these types of arguments.