On round-off errors in linear programming