Numerical Errors in Logistic Map Calculation
J. C. Sprott
Department of Physics, University of Wisconsin, Madison,
WI 53706, USA
September 15, 1997
(Modified April 20, 2000)
When numerically iterating the logistic equation
Xn+1 = AXn(1 - Xn)
round-off errors can cause spurious results. Consider the case
A
= 4 which maps the unit interval (0, 1) back onto itself exactly twice.
Because of the shape of the parabolic logistic function, many values of
Xn
near 0.5 map into values of Xn+1 near 1.0
which in turn map into values of Xn+2 close
to zero. If these values of Xn+2 are
indistinguishable from zero to within machine precision, they will thereafter
remain at zero because zero is an unstable fixed point solution of the
logistic map. As a result, initial conditions X0
will iterate to zero on average after a relatively small number of iterations
N
whose value we estimate below.
Suppose that the logistic equation with A
= 4 is iterated using (4-byte) single-precision arithmetic. Four
bytes is 32 bits, of which 23 are used for the mantissa, 8 for the exponent,
and 1 for the sign according to the IEEE standard for floating point arithmetic.
Thus the computer can represent 223 = 8,388,608 unique values
in the range (0, 1) for a given exponent. Values near zero can be
represented more accurately by changing the exponent. Thus the problem
occurs for values of Xn+1 that are indistinguishable
from 1. Values larger than 1 - 2-24 will round to 1.0.
With this value of Xn+1, we can calculate
the range of Xn values whose next iterates round to 1.0
by solving the quadratic equation
X2 - X + 1/4 - 2-26 = 0
which has solutions X = 0.5 ± 2-13. Thus
the region of the interval (0, 1) over which Xn
is an eventually fixed point is of width 2-12 = 2.44 x 10-4
centered
on 0.5. Since the probability density of finding an iterate of the
logistic map with A = 4 is
P = 1 / pi[X(1 - X)]1/2
which has a value of P = 2 / pi = 0.637 for X = 0.5, the
probability that an iterate will fall within the region above is 1.55 x
10-4. Thus on average, we would expect the reciprocal
of this number of iterates before the fixed point at zero is reached, or
N
= 6400.
To test this prediction, the logistic map was iterated
in single precision many times using a uniform random initial condition
in (0,1), and the number of iterations required for zero to be reached
was determined. The source and object
code used for this test are available. Typical results follow:
Solution reached 0 after 3334 iterations - Average = 1981 - Cases =
989
Solution reached 0 after 1082 iterations - Average = 1980 - Cases =
990
Solution reached 0 after 3153 iterations - Average = 1981 - Cases =
991
Solution reached 0 after 4017 iterations - Average = 1984 - Cases =
992
Solution reached 0 after 2941 iterations - Average = 1985 - Cases =
993
Solution reached 0 after 2581 iterations - Average = 1985 - Cases =
994
Solution reached 0 after 1183 iterations - Average = 1984 - Cases =
995
Solution reached 0 after 3890 iterations - Average = 1986 - Cases =
996
Solution reached 0 after 2604 iterations - Average = 1987 - Cases =
997
Solution caught in a periodic cycle
Solution reached 0 after 3136 iterations - Average = 1988 - Cases =
998
Solution reached 0 after 542 iterations - Average = 1987 - Cases =
999
Solution reached 0 after 2082 iterations - Average = 1987 - Cases =
1000
Solution reached 0 after 927 iterations - Average = 1986 - Cases =
1001
Solution reached 0 after 1984 iterations - Average = 1986 - Cases =
1002
Solution reached 0 after 538 iterations - Average = 1984 - Cases =
1003
Solution caught in a periodic cycle
Solution reached 0 after 297 iterations - Average = 1982 - Cases =
1004
Solution reached 0 after 3455 iterations - Average = 1984 - Cases =
1005
Solution reached 0 after 2529 iterations - Average = 1984 - Cases =
1006
Solution reached 0 after 2051 iterations - Average = 1985 - Cases =
1007
Solution reached 0 after 1509 iterations - Average = 1984 - Cases =
1008
Solution reached 0 after 1056 iterations - Average = 1983 - Cases =
1009
Solution reached 0 after 2723 iterations - Average = 1984 - Cases =
1010
Occasionally the orbit gets stuck in some very high periodic cycle and
never reaches zero. Excluding such cases, the average over 32,000
runs is N = 2031. This result is about 3 ( or pi?) times worse
than expected for reasons that are not clear. Furthermore, the expectation
is that the number
N is inversely proportional to the square root
of the precision. Thus for (8-byte) double-precision with a mantissa
of 52 bits, the expectation is for N = 1.5 x 108, which
is in reasonable agreement with numerical results. This particular
problem can be solved by using a value of A slightly less than 4.0,
such as 3.999999, although the orbit still occasionally gets caught in
a cycle of some presumably high period.
Update (November 7, 2026): The "3 (or pi?)" puzzle resolved
The estimate above used the wrong threshold.
With IEEE round-to-nearest arithmetic, the single-precision numbers just
below 1 are spaced 2-24 apart, so a result rounds to 1.0 when
it exceeds the halfway point 1 - 2-25, not 1 - 2-24.
The quadratic equation then becomes
X2 - X + 1/4 - 2-27 = 0
with solutions X = 0.5 ± 2-13.5, a region of
width 2-12.5 = 1.73 x 10-4, a probability of 1.10
x 10-4, and an expected N = (pi/2)212.5 ≈
9100. A direct count of the 4345 single-precision values near 0.5
whose correctly rounded image is exactly 1.0 gives N = 9098.
Thus the measured N = 2031 is about 4.5 times (not 3 or pi times)
shorter than expected, and the near-agreement with pi was a coincidence.
New simulations with 200,000 random initial conditions
each (C on a modern PC with IEEE arithmetic) give:
-
Product calculated exactly and rounded once to single precision (as a
BASIC that evaluates expressions in extended precision and stores the result
in single precision would presumably do): 93.0% reach zero with an average
of N = 2034 ± 3, in agreement with the 2031 above, and the
other 7.0% are caught in cycles of period 930, 431, 205, 106, 5, or 4.
No orbit took longer than 4891 iterations to reach zero.
-
A tiny random perturbation of one part in 107 added to each
iterate (except those near 1): all orbits reach zero, with an average of
N = 9195 ± 20 and the exponential distribution expected for
random events, within about 1% of the estimate.
-
Every multiplication and subtraction rounded to single precision: 81% of
the orbits are caught in cycles (67% of all orbits in a single cycle of
period 4344), and the remaining 19% reach zero after an average of only
1169 iterations.
The explanation is that the single-precision logistic
map is a many-to-one mapping of a finite set of numbers onto itself.
X and 1 - X have the same image, and thousands of neighboring
values near 0.5 (or near 0 and 1) map onto the same number. Once
two orbits land on the same number, they coincide forever. Starting
from 106 random initial conditions, only about 270,000 different
values remain after 79 iterations and 57,000 after 421 iterations, compared
with 866,000 and 836,000 when the small random perturbation is added.
The orbits thus merge into a tree that drains into zero (or into one of
a few cycles), and the time to reach zero is set by the depth of that tree
rather than by random sampling of the probability density P.
The estimate of 9100 is correct for a random sequence of iterates, but
the actual orbits are not random.
A similar thing happens in double precision.
The corrected estimate for round-to-nearest is N = (pi/2)227
≈ 2.1 x 108 (2.2 x 108 from a direct count), and
1200 orbits with a small random perturbation gave N = (2.20 ±
0.06) x 108. Without the perturbation, however, 84% of
8000 double-precision orbits were caught in cycles (68% of all orbits in
a single cycle of period 5,638,349), and the other 16% reached zero after
an average of only 1.0 x 107 iterations. Thus the agreement
with numerical results claimed above for double precision is not borne
out by these calculations, although the 1997 test may have used different
arithmetic, which I cannot check. Details, figures, and a short Python
program are in the note Finite-Precision Funnels
in Chaotic Maps.
Updated November 7, 2026 by Rob - Clint Sprott's AI assistant.
Back to Sprott's Technical Notes