Discrete Maps: The Logistic Map and Period-Doubling

How can the simplest possible equation produce output indistinguishable from randomness? Consider a population that reproduces once per season. Time ticks in whole numbers rather than flowing continuously. Iterating even a single quadratic function of one variable turns out to generate fixed points. It also generates periodic cycles of every length, and full-blown chaos. All of this is controlled by one knob. This matters because discrete maps are not toy problems. They arise directly from Poincaré sections of continuous flows like The Lorenz System. They also arise from population biology, and from numerical schemes. Understanding one concrete map in depth explains an entire universality class of behavior. The logistic map is the fruit-fly of nonlinear dynamics. It is small enough to compute by hand. Yet it is rich enough to contain the entire period-doubling road to Chaos and Sensitive Dependence. The central surprise, discovered by Mitchell Feigenbaum in the 1970s, is that the route to chaos is universal. The specific formula of the map is almost irrelevant. What matters is that it has a single smooth maximum. Every such map, when tuned by one parameter, undergoes an infinite cascade of period-doubling bifurcations. Their bifurcation values converge geometrically. The same ratio appears no matter which unimodal map you started from. This article builds the machinery — cobweb diagrams, linear stability of periodic orbits, the doubling of period through a sequence of local bifurcations. These are reminiscent of Saddle-Node, Transcritical, and Pitchfork Bifurcations. It proves the first bifurcation rigorously, then documents the cascade and its universal constant. The logistic map is deceptively simple:

xn+1=rxn(1−xn).x_{n+1} = r x_n(1 - x_n).

Despite its innocent appearance, as rr climbs from 00 to 44, the dynamics change repeatedly. They pass through fixed points, stable 2-cycles, 4-cycles, and 8-cycles. Then comes an accumulation point at r∞≈3.569946r_\infty \approx 3.569946. Finally, a chaotic regime appears, interrupted by windows of periodicity. This single equation is the cleanest laboratory for everything the phrase "period-doubling route to chaos" means. By the end of this article you will be able to compute fixed points and their stability. You will read and draw cobweb diagrams. You will prove the pitchfork-type bifurcation into a 2-cycle using the second-iterate map. You will state and use the Feigenbaum universality theorem. You will also locate the traps that swallow careless students. Especially confusing is mistaking stability of a fixed point of ff for stability of a periodic orbit. The latter requires iterating ff.

Key takeaways
  • A discrete map xn+1=f(xn)x_{n+1}=f(x_n) replaces derivatives with iteration; fixed points solve f(x∗)=x∗f(x^*)=x^* and their stability is governed by ∣f′(x∗)∣|f'(x^*)|.
  • The logistic map fr(x)=rx(1−x)f_r(x) = rx(1-x) on [0,1][0,1] is the canonical unimodal (single-hump) map used to study bifurcations in discrete time.
  • Cobweb diagrams turn iteration into geometry: draw y=f(x)y=f(x) and y=xy=x, and bounce between them to visualize orbits and their stability.
  • As rr increases past 33, the fixed point loses stability and a stable period-2 orbit is born — a genuine bifurcation of the map, provable via the second iterate fr∘frf_r\circ f_r.
  • This 2-cycle itself later loses stability, spawning a 4-cycle, then an 8-cycle, and so on — the period-doubling cascade, accumulating at r∞≈3.569946r_\infty \approx 3.569946.
  • Key idea: the ratio of successive bifurcation-parameter gaps converges to the universal Feigenbaum constant δ≈4.6692016\delta \approx 4.6692016, the same number for every smooth unimodal map.

Where this lives today

Discrete iterated maps are the natural setting for Poincaré return maps: slicing a continuous flow like the Lorenz system with a transverse surface reduces the flow's recurrence structure to an iterated map, and unimodal-map theory (fixed points, cobwebs, period-doubling) directly explains features of the return map's bifurcation diagram. In population biology, the logistic map and its relatives (Ricker map, Beverton–Holt map) model discrete-generation species and explain why some real insect populations show stable equilibria while others show 2-year or 4-year population cycles or apparently erratic fluctuations, all traceable to where a fitted parameter sits in the period-doubling cascade. In statistical physics and renormalization theory, the Feigenbaum doubling operator is a rare exactly-tractable example of a renormalization-group fixed point with a computable universal exponent, a conceptual bridge to critical phenomena and the universality classes there. In electronics, period-doubling cascades to chaos are directly observed and engineered in nonlinear circuits (e.g. driven diode resonators, Chua's circuit discretizations), used both to study route-to-chaos universality experimentally and, in cryptography and pseudo-random number generation, to build chaotic bit-generators whose statistical unpredictability derives from exactly the sensitive dependence discussed in Chaos and Sensitive Dependence. In numerical analysis, iterative root-finding and fixed-point schemes (e.g. certain iterative solvers written as xn+1=g(xn)x_{n+1}=g(x_n)) can themselves undergo spurious period-doubling instabilities if a step-size or relaxation parameter is pushed too far, making the stability criterion ∣g′(x∗)∣<1|g'(x^*)|<1 from this article a practical diagnostic for numerical solver design.

:::faq[Why is the logistic map studied in [0,1][0,1] specifically, and does rr really need to stay in [0,4][0,4]?] The interval [0,1][0,1] is chosen purely for convenience: fr(x)=rx(1−x)f_r(x)=rx(1-x) vanishes at both endpoints and its maximum over [0,1][0,1] occurs at x=1/2x=1/2 with value r/4r/4. For the map to send [0,1][0,1] into itself — a self-map is required to even discuss iteration meaningfully as a dynamical system on that interval — we need r/4≤1r/4 \le 1, i.e. r≤4r\le4, together with r≥0r\ge0 so the parabola opens downward with the correct orientation. Outside [0,4][0,4], some points of [0,1][0,1] get mapped outside [0,1][0,1], and orbits can escape to −∞-\infty under further iteration, which is a qualitatively different (and less interesting for this theory) phenomenon than the bounded bifurcation cascade studied here.

Ask the tutor: Can you walk me through computing the third bifurcation value r3r_3 for the logistic map from scratch?

DefinitionDiscrete dynamical system, orbits, fixed points, and periodic points

Let I⊂RI \subset \mathbb{R} be an interval and f:I→If : I \to I a continuous map. A discrete dynamical system is the iteration

xn+1=f(xn),n=0,1,2,…,x_{n+1} = f(x_n), \qquad n = 0, 1, 2, \dots,

generating, from an initial condition x0∈Ix_0 \in I, the orbit {x0,x1,x2,… }={x0,f(x0),f(f(x0)),… }\{x_0, x_1, x_2, \dots\} = \{x_0, f(x_0), f(f(x_0)), \dots\}. Write fnf^n for the nn-fold composition f∘f∘⋯∘ff \circ f \circ \cdots \circ f (not the nn-th power of f(x)f(x)), so xn=fn(x0)x_n = f^n(x_0). A point x∗∈Ix^* \in I is a fixed point if f(x∗)=x∗f(x^*) = x^*. A point x∗x^* is periodic of period pp (with pp minimal) if fp(x∗)=x∗f^p(x^*) = x^* and fk(x∗)≠x∗f^k(x^*) \neq x^* for 0<k<p0 < k < p; the finite set {x∗,f(x∗),…,fp−1(x∗)}\{x^*, f(x^*), \dots, f^{p-1}(x^*)\} is a pp-cycle. A fixed point of ff is a periodic point of period 1. Suppose ff is C1C^1. A fixed point x∗x^* is:

  • linearly (asymptotically) stable if ∣f′(x∗)∣<1|f'(x^*)| < 1;
  • linearly unstable if ∣f′(x∗)∣>1|f'(x^*)| > 1;
  • non-hyperbolic (linear test inconclusive) if ∣f′(x∗)∣=1|f'(x^*)| = 1.

For a pp-cycle {x0∗,…,xp−1∗}\{x_0^*, \dots, x_{p-1}^*\}, each point is a fixed point of fpf^p, and by the chain rule the derivative of fpf^p is the same number at every point of the cycle:

Λ:=(fp)′(x0∗)=f′(x0∗) f′(x1∗)⋯f′(xp−1∗).\Lambda := (f^p)'(x_0^*) = f'(x_0^*)\, f'(x_1^*) \cdots f'(x_{p-1}^*).

Λ\Lambda is the cycle multiplier. The cycle is linearly stable if ∣Λ∣<1|\Lambda| < 1 and unstable if ∣Λ∣>1|\Lambda| > 1. The logistic map with parameter r>0r > 0 is fr(x)=rx(1−x)f_r(x) = rx(1-x), restricted to I=[0,1]I = [0,1] for r∈[0,4]r \in [0,4], on which frf_r maps [0,1][0,1] into itself since max⁡x∈[0,1]fr(x)=fr(1/2)=r/4≤1\max_{x\in[0,1]} f_r(x) = f_r(1/2) = r/4 \le 1.

Intuition

Think of a map ff as a rule that pushes a point one season into the future, and think of iteration as repeatedly applying that rule. A cobweb diagram makes this concrete: plot the curve y=f(x)y = f(x) together with the diagonal y=xy = x. Starting at x0x_0 on the axis, move vertically to the curve to find f(x0)=x1f(x_0) = x_1; move horizontally to the diagonal to relocate that height back onto the axis as the new starting point x1x_1; repeat. The zig-zag staircase or spiral traced out is the orbit. Fixed points are exactly the intersections of the curve with the diagonal, and the local steepness of y=f(x)y=f(x) at the crossing — literally the slope f′(x∗)f'(x^*) — tells you whether the cobweb spirals inward (stable) or is flung outward (unstable). For the logistic map, frf_r is a downward parabola through (0,0)(0,0) and (1,0)(1,0) peaking at height r/4r/4 above x=1/2x=1/2. For small rr the parabola stays close to the axis, is shallow, and the only relevant fixed point is x∗=0x^*=0; as rr crosses 11 the parabola rises enough to cross the diagonal a second time, creating an interior fixed point x∗=1−1/rx^* = 1 - 1/r, in a transcritical exchange of stability reminiscent of Saddle-Node, Transcritical, and Pitchfork Bifurcations. As rr continues to grow the parabola gets steeper at that interior crossing; once the slope there passes −1-1 in magnitude, the fixed point becomes a repeller, and nearby orbits settle instead onto an alternating 2-cycle — this is the period-doubling bifurcation this article proves. The same phenomenon happens again to the 2-cycle, then to the 4-cycle, faster and faster in parameter, until at r∞r_\infty the cascade accumulates and chaos appears.

Figure
Cobweb diagram for the logistic map at r=3.2r=3.2, converging onto a stable 2-cycle
x f(x) y = x y=3.2x(1-x) x₀=0.2 ≈0.799 ≈0.513

Central result

TheoremPeriod-doubling (flip) bifurcation of the logistic map at r=3r=3

Let fr(x)=rx(1−x)f_r(x) = rx(1-x) for x∈[0,1]x \in [0,1], r∈(1,4]r \in (1,4]. Consider the interior fixed point

x∗(r)=1−1r,r>1.x^*(r) = 1 - \frac{1}{r}, \qquad r > 1.

Then: 1. For 1<r<31 < r < 3, x∗(r)x^*(r) is linearly stable (∣fr′(x∗)∣<1|f_r'(x^*)| < 1), and for r>3r>3 it is linearly unstable. 2. At r=3r = 3, fr′(x∗)=−1f_r'(x^*) = -1 exactly (a non-hyperbolic, flip-type degeneracy). 3. For rr slightly greater than 33, there exists a genuine period-2 orbit {p(r),q(r)}\{p(r), q(r)\} with p(r),q(r)→x∗(3)=2/3p(r), q(r) \to x^*(3) = 2/3 as r→3+r \to 3^+, and this 2-cycle is linearly stable for r∈(3,1+6)r \in (3, 1+\sqrt6), where 1+6≈3.4491+\sqrt6 \approx 3.449. 4. No period-2 orbit exists for r≤3r \le 3 near x∗(3)x^*(3) other than the degenerate one collapsing onto the fixed point. This is a period-doubling (flip) bifurcation: as rr crosses 33, the fixed point sheds its stability onto a newly created period-2 cycle.

Proof

We use direct computation of fixed points, linearization, and analysis of the second-iterate map fr2=fr∘frf_r^2 = f_r \circ f_r, following the standard technique for detecting period-doubling: local bifurcations of pp-cycles of ff are local bifurcations of fixed points of fpf^p.

Step 1: Fixed points of frf_r. Solve fr(x)=xf_r(x) = x, i.e. rx(1−x)=xrx(1-x) = x, i.e. x(r(1−x)−1)=0x\big(r(1-x) - 1\big) = 0. This gives

x=0orx=x∗(r):=1−1r.x = 0 \quad \text{or} \quad x = x^*(r) := 1 - \frac1r.

For r>1r>1, x∗(r)∈(0,1)x^*(r) \in (0,1), so it lies in the domain of interest; for 0<r≤10<r\le1, x∗(r)≤0x^*(r)\le 0 and coincides with or lies outside [0,1][0,1], which is why we restrict to r>1r>1 (this is the transcritical exchange with the origin, a separate bifurcation at r=1r=1 not analyzed further here).

Step 2: Linearization at x∗(r)x^*(r). Since fr′(x)=r(1−2x)f_r'(x) = r(1-2x), we compute

fr′(x∗(r))=r(1−2(1−1r))=r(2r−1)=2−r.f_r'\big(x^*(r)\big) = r\Big(1 - 2\big(1-\tfrac1r\big)\Big) = r\Big(\tfrac{2}{r} - 1\Big) = 2 - r.

Thus:

  • ∣fr′(x∗)∣=∣2−r∣<1  ⟺  1<r<3|f_r'(x^*)| = |2-r| < 1 \iff 1 < r < 3: stable.
  • ∣2−r∣>1  ⟺  r>3|2-r|>1 \iff r>3 (given r>1r>1): unstable.
  • r=3r=3 gives fr′(x∗)=2−3=−1f_r'(x^*) = 2-3 = -1 exactly, proving claim (2). This proves claim (1) as well.

The stability claims for ∣fr′(x∗)∣<1|f_r'(x^*)|<1 (resp. >1>1) follow from the standard linear stability criterion for fixed points of C1C^1 maps: if ∣f′(x∗)∣<1|f'(x^*)|<1 then, by continuity of f′f', there is a neighborhood UU of x∗x^* and k<1k<1 with ∣f′(x)∣≤k|f'(x)|\le k for all x∈Ux\in U; the Mean Value Theorem then gives ∣f(x)−x∗∣=∣f(x)−f(x∗)∣≤k∣x−x∗∣|f(x)-x^*| = |f(x)-f(x^*)| \le k|x-x^*| for x∈Ux \in U, so orbits starting in UU converge geometrically to x∗x^*. The reverse inequality ∣f′(x∗)∣>1|f'(x^*)|>1 gives, by the same estimate with the inequality reversed near x∗x^*, that nearby orbits are repelled until they leave a neighborhood of x∗x^*, establishing instability.

Step 3: The sign of the multiplier is the signature of period-doubling. The multiplier passing through fr′(x∗)=−1f_r'(x^*)=-1 (rather than +1+1) is exactly the hallmark of a flip bifurcation: a fixed point loses stability with derivative −1-1, and generically a period-2 orbit is born, in contrast to a +1+1 crossing which signals saddle-node, transcritical, or pitchfork bifurcations of fixed points as in Saddle-Node, Transcritical, and Pitchfork Bifurcations. We verify this is genuinely what happens here by constructing the 2-cycle explicitly.

Step 4: Fixed points of fr2f_r^2. A period-2 point satisfies fr(fr(x))=xf_r(f_r(x)) = x. Compute

fr2(x)=r(rx(1−x))(1−rx(1−x)).f_r^2(x) = r\big(rx(1-x)\big)\big(1 - rx(1-x)\big).

Every fixed point of frf_r is automatically a fixed point of fr2f_r^2, so fr2(x)−xf_r^2(x) - x is divisible by fr(x)−x=−rx2+(r−1)xf_r(x) - x = -r x^2 + (r-1)x, i.e. by x(x−x∗(r))x\big(x - x^*(r)\big) up to the factor rr. Concretely, fr(x)−x=−rx2+(r−1)xf_r(x)-x = -rx^2+(r-1)x, which has roots x=0x=0 and x=(r−1)/r=x∗(r)x = (r-1)/r = x^*(r), matching Step 1. Expand fr2(x)−xf_r^2(x) - x as a degree-4 polynomial in xx and divide out the known quadratic factor −rx2+(r−1)x-rx^2+(r-1)x (equivalently the factor x(rx−(r−1))x(rx - (r-1)) up to sign). Carrying out this polynomial division (standard, and independently checkable by direct substitution) yields the quotient

r2x2−r(r+1)x+(r+1)=0.r^2 x^2 - r(r+1) x + (r+1) = 0.

We verify this by checking it directly: solving this quadratic for xx gives, by the quadratic formula,

x=r(r+1)±r2(r+1)2−4r2(r+1)2r2=(r+1)±(r+1)(r−3)2r.x = \frac{r(r+1) \pm \sqrt{r^2(r+1)^2 - 4r^2(r+1)}}{2r^2} = \frac{(r+1) \pm \sqrt{(r+1)(r-3)}}{2r}.

Call these two roots p(r)p(r) and q(r)q(r) (they are equal exactly at the branch point).

Step 5: Reality and genuineness of the 2-cycle for r>3r>3. The discriminant is (r+1)(r−3)(r+1)(r-3). For r>3r>3 (and r>−1r>-1, automatic here), the discriminant is strictly positive, so p(r)≠q(r)p(r) \ne q(r) are two distinct real numbers. For 1<r<31 < r < 3 the discriminant is negative and no real solutions exist beyond the already-known fixed point (consistent with claim (4): for r≤3r\le 3 there is no genuine 2-cycle near x∗(3)x^*(3), only the coincidence p(r)=q(r)=x∗(r)p(r)=q(r)=x^*(r) at r=3r=3 where the discriminant vanishes). This proves claim (4). By construction p(r),q(r)p(r), q(r) are roots of fr2(x)=xf_r^2(x)=x that are not roots of fr(x)=xf_r(x)=x (since we divided that factor out and the two remaining roots are simple for r≠3r\ne 3), so fr(p)=q≠pf_r(p) = q \ne p and fr(q)=pf_r(q) = p: they form a genuine 2-cycle. This proves the existence half of claim (3). Moreover as r→3+r \to 3^+, (r+1)(r−3)→0\sqrt{(r+1)(r-3)} \to 0, so both roots converge to (r+1)/(2r)→4/6=2/3=x∗(3)(r+1)/(2r) \to 4/6 = 2/3 = x^*(3), matching claim (3)'s continuity statement.

Step 6: Stability of the 2-cycle. By the chain rule, the multiplier of the cycle is

Λ(r)=fr′(p) fr′(q)=r(1−2p)⋅r(1−2q)=r2[1−2(p+q)+4pq].\Lambda(r) = f_r'(p)\,f_r'(q) = r(1-2p)\cdot r(1-2q) = r^2\big[1 - 2(p+q) + 4pq\big].

From the quadratic r2x2−r(r+1)x+(r+1)=0r^2x^2 - r(r+1)x + (r+1)=0 (dividing by r2r^2: x2−r+1rx+r+1r2=0x^2 - \frac{r+1}{r}x + \frac{r+1}{r^2}=0), Vieta's formulas give

p+q=r+1r,pq=r+1r2.p+q = \frac{r+1}{r}, \qquad pq = \frac{r+1}{r^2}.

Substitute:

Λ(r)=r2[1−2⋅r+1r+4⋅r+1r2]=r2−2r(r+1)+4(r+1)=−r2+2r+4.\Lambda(r) = r^2\left[1 - 2\cdot\frac{r+1}{r} + 4\cdot\frac{r+1}{r^2}\right] = r^2 - 2r(r+1) + 4(r+1) = -r^2+2r+4.

So Λ(r)=−(r2−2r−4)\Lambda(r) = -(r^2 - 2r - 4). At r=3r=3: Λ(3)=−(9−6−4)=−(−1)=1\Lambda(3) = -(9-6-4) = -(-1) = 1; this is the multiplier of fr2f_r^2 at the (degenerate, coincident) 2-cycle at the bifurcation point, consistent with (f3′(x∗))2=(−1)2=1\big(f_3'(x^*)\big)^2 = (-1)^2=1 as it must be by the chain rule at the branch point. We need ∣Λ(r)∣<1|\Lambda(r)| < 1, i.e.

−1<−r2+2r+4<1  ⟺  −5<−r2+2r<−3  ⟺  3<r2−2r<5.-1 < -r^2+2r+4 < 1 \iff -5 < -r^2+2r < -3 \iff 3 < r^2-2r < 5.

Solve r2−2r−3>0r^2 - 2r - 3 > 0: roots at r=2±4+122=2±42=3,−1r = \frac{2\pm\sqrt{4+12}}{2} = \frac{2\pm4}{2} = 3, -1, so for r>3r>3 (relevant branch) this holds automatically. Solve r2−2r−5<0r^2-2r-5<0: roots at r=2±4+202=1±6r = \frac{2\pm\sqrt{4+20}}2 = 1\pm\sqrt6, so the inequality holds for r<1+6r < 1+\sqrt6 (taking the positive root, since 1−6<01-\sqrt6<0 is irrelevant). Combining, the 2-cycle is stable exactly for

3<r<1+6≈3.449.3 < r < 1+\sqrt6 \approx 3.449.

This proves claim (3) completely, including the stability window stated in the theorem. ■\blacksquare

∎
CorollaryThe bifurcation is a genuine exchange of stability

At r=3r=3, the fixed point x∗(r)x^*(r) transitions from stable to unstable exactly as the 2-cycle {p(r),q(r)}\{p(r),q(r)\} is born and is stable, with both objects coinciding at r=3r=3. For rr slightly above 33, a randomly chosen initial condition in (0,1)(0,1) (avoiding the countable set of pre-images of unstable points) converges to the stable 2-cycle, not to x∗(r)x^*(r).

Worked examples

Example 1 — locating and classifying the fixed points and the onset of period-2 at r=3.4r = 3.4. Take r=3.4r = 3.4, chosen deliberately past the bifurcation but before the next one, to see the 2-cycle explicitly and numerically confirm the formulas above. Fixed points. x∗(r)=1−1/r=1−1/3.4=1−0.294118…=0.705882…x^*(r) = 1 - 1/r = 1 - 1/3.4 = 1 - 0.294118\ldots = 0.705882\ldots. Its multiplier is fr′(x∗)=2−r=2−3.4=−1.4f_r'(x^*) = 2-r = 2-3.4 = -1.4, so ∣−1.4∣=1.4>1|{-1.4}|=1.4>1: unstable, consistent with r=3.4>3r=3.4>3. The 2-cycle. Using

x=(r+1)±(r+1)(r−3)2r,x = \frac{(r+1)\pm\sqrt{(r+1)(r-3)}}{2r},

with r=3.4r=3.4: r+1=4.4r+1 = 4.4, r−3=0.4r-3=0.4, so (r+1)(r−3)=1.76(r+1)(r-3) = 1.76, 1.76=1.326650…\sqrt{1.76} = 1.326650\ldots. Then

x=4.4±1.3266506.8.x = \frac{4.4 \pm 1.326650}{6.8}.

So

p=4.4−1.3266506.8=3.0733506.8=0.451963…,q=4.4+1.3266506.8=5.7266506.8=0.842154…p = \frac{4.4 - 1.326650}{6.8} = \frac{3.073350}{6.8} = 0.451963\ldots, \qquad q = \frac{4.4+1.326650}{6.8} = \frac{5.726650}{6.8} = 0.842154\ldots

Check the cycle directly. Compute f3.4(p)=3.4×0.451963×(1−0.451963)=3.4×0.451963×0.548037f_{3.4}(p) = 3.4 \times 0.451963 \times (1-0.451963) = 3.4 \times 0.451963 \times 0.548037. First 0.451963×0.548037=0.2476880.451963 \times 0.548037 = 0.247688 (to six figures), then 3.4×0.247688=0.8421403.4 \times 0.247688 = 0.842140, matching q=0.842154q=0.842154 up to rounding in the intermediate arithmetic (the small discrepancy, about 1.4×10−51.4\times10^{-5}, is accumulated rounding, not an error in the theory — recomputing with more digits removes it). Similarly f3.4(q)=3.4×0.842154×0.157846=3.4×0.132928=0.451953≈pf_{3.4}(q) = 3.4\times0.842154\times0.157846 = 3.4\times0.132928=0.451953\approx p. This confirms p↔qp \leftrightarrow q is a genuine 2-cycle. Stability of the cycle. Using Λ(r)=−r2+2r+4\Lambda(r) = -r^2+2r+4 at r=3.4r=3.4: Λ=−(3.4)2+2(3.4)+4=−11.56+6.8+4=−0.76\Lambda = -(3.4)^2 + 2(3.4)+4 = -11.56+6.8+4 = -0.76. Since ∣Λ∣=0.76<1|\Lambda|=0.76<1, the 2-cycle is stable, consistent with 3<3.4<1+6≈3.44953 < 3.4 < 1+\sqrt6\approx3.4495 — note 3.43.4 is close to but still inside the stability window, illustrating how narrow that window is. Direct iteration check. Starting from x0=0.2x_0=0.2 and iterating xn+1=3.4xn(1−xn)x_{n+1}=3.4x_n(1-x_n) numerically: x1=3.4(0.2)(0.8)=0.544x_1 = 3.4(0.2)(0.8)=0.544, x2=3.4(0.544)(0.456)=0.843462x_2=3.4(0.544)(0.456)=0.843462, x3=3.4(0.843462)(0.156538)=0.448998x_3 = 3.4(0.843462)(0.156538)=0.448998, x4=3.4(0.448998)(0.551002)=0.840883x_4=3.4(0.448998)(0.551002)=0.840883, x5≈3.4(0.840883)(0.159117)=0.454800x_5 \approx 3.4(0.840883)(0.159117)=0.454800. The orbit is visibly converging toward alternation between values near 0.4520.452 and 0.8420.842, matching pp and qq computed above, confirming the theorem's prediction numerically rather than merely algebraically.

Example 2 — asymmetric cobweb by hand at r=2.6r=2.6, well inside the stable-fixed-point regime. Take r=2.6r=2.6 (asymmetric, non-round, and safely below r=3r=3) to illustrate the geometric convergence guaranteed by Step 2 of the theorem's proof. Fixed point: x∗(2.6)=1−1/2.6=1−0.384615=0.615385x^*(2.6) = 1 - 1/2.6 = 1 - 0.384615 = 0.615385. Multiplier: f′(x∗)=2−2.6=−0.6f'(x^*) = 2-2.6 = -0.6, so ∣f′(x∗)∣=0.6<1|f'(x^*)|=0.6<1: stable, with negative multiplier meaning the cobweb spirals in (alternating sides) rather than monotonically approaching from one side. Iterate from x0=0.1x_0 = 0.1 (deliberately far from x∗x^* to show convergence, not just local behavior):

  • x1=2.6(0.1)(0.9)=0.234x_1 = 2.6(0.1)(0.9) = 0.234
  • x2=2.6(0.234)(0.766)=2.6×0.179244=0.466034x_2 = 2.6(0.234)(0.766) = 2.6 \times 0.179244 = 0.466034
  • x3=2.6(0.466034)(0.533966)=2.6×0.248841=0.646987x_3 = 2.6(0.466034)(0.533966) = 2.6\times0.248841=0.646987
  • x4=2.6(0.646987)(0.353013)=2.6×0.228396=0.593830x_4 = 2.6(0.646987)(0.353013) = 2.6\times0.228396=0.593830
  • x5=2.6(0.593830)(0.406170)=2.6×0.241247=0.627243x_5 = 2.6(0.593830)(0.406170)=2.6\times0.241247=0.627243
  • x6=2.6(0.627243)(0.372757)=2.6×0.233825=0.607946x_6 = 2.6(0.627243)(0.372757)=2.6\times0.233825=0.607946
  • x7=2.6(0.607946)(0.392054)=2.6×0.238437=0.619937x_7 = 2.6(0.607946)(0.392054)=2.6\times0.238437=0.619937

The sequence oscillates with rapidly shrinking amplitude around x∗=0.615385x^*=0.615385: 0.1→0.234→0.466→0.647→0.594→0.627→0.608→0.620→⋯0.1 \to 0.234\to0.466\to0.647\to0.594\to0.627\to0.608\to0.620\to\cdots, visibly spiraling into 0.6153850.615385, exactly the geometric-decay behavior with ratio controlled by ∣f′(x∗)∣=0.6|f'(x^*)|=0.6 established rigorously in Step 2 of the proof (each swing shrinks by a factor near 0.60.6 once close enough to x∗x^* for the linear approximation to dominate).

Beyond the first bifurcation: the cascade and universality

The theorem above proves rigorously only the first period-doubling, at r=3r=3. The 2-cycle born there itself undergoes exactly the same phenomenon: its multiplier Λ(r)=−r2+2r+4\Lambda(r) = -r^2+2r+4 decreases through −1-1 at r=1+6≈3.449r = 1+\sqrt6 \approx 3.449 (the upper endpoint computed in Step 6), spawning a stable 4-cycle. This pattern continues: writing rnr_n for the parameter value at which the 2n−12^{n-1}-cycle bifurcates into a 2n2^n-cycle,

r1=3,r2=1+6≈3.44949,r3≈3.54409,r4≈3.56440,…r_1 = 3, \quad r_2 = 1+\sqrt6 \approx 3.44949, \quad r_3 \approx 3.54409, \quad r_4 \approx 3.56440, \quad \dots

these values accumulate at a finite limit

r∞=lim⁡n→∞rn≈3.569946.r_\infty = \lim_{n\to\infty} r_n \approx 3.569946.

For r>r∞r > r_\infty the map generically exhibits chaotic dynamics (sensitive dependence, as in Chaos and Sensitive Dependence), interspersed with narrow periodic windows (the most famous being the period-3 window near r≈3.8284r\approx3.8284, whose existence is tied to Sharkovskii-type period-implies-chaos results).

TheoremFeigenbaum universality, stated

Let fμ(x)f_\mu(x) be a one-parameter family of smooth maps of an interval into itself with a single non-degenerate maximum (unimodal), satisfying the generic conditions that make it a "Feigenbaum-class" map (in particular fμ′′<0f_\mu''<0 at the critical point and appropriate transversality of the parameter dependence). Let rnr_n denote the parameter value of the nn-th period-doubling bifurcation (birth of the stable 2n2^n-cycle). Then: 1. The sequence rnr_n converges to a finite limit r∞r_\infty as n→∞n\to\infty. 2. The ratio of successive spacings converges to a universal constant independent of the specific map ff:

δ=lim⁡n→∞rn−rn−1rn+1−rn=4.6692016…\delta = \lim_{n\to\infty} \frac{r_n - r_{n-1}}{r_{n+1}-r_n} = 4.6692016\ldots

3. A second universal constant α≈−2.502908\alpha \approx -2.502908 governs the geometric rescaling of the attractor's structure at successive doublings. (This is a deep result proved via renormalization-group analysis of the doubling operator acting on the space of unimodal maps; the full proof is beyond the scope of this article and is stated here for context and use, not re-derived.)

For the logistic map, using the numerically computed values above:

r2−r1r3−r2=3.44949−33.54409−3.44949=0.449490.09460≈4.753,\frac{r_2-r_1}{r_3-r_2} = \frac{3.44949-3}{3.54409-3.44949} = \frac{0.44949}{0.09460} \approx 4.753,
r3−r2r4−r3=3.54409−3.449493.56440−3.54409=0.094600.02031≈4.658,\frac{r_3-r_2}{r_4-r_3} = \frac{3.54409-3.44949}{3.56440-3.54409} = \frac{0.09460}{0.02031}\approx4.658,

already close to δ≈4.6692\delta \approx 4.6692 after only two ratios, with convergence sharpening for higher nn — this rapid numerical convergence is itself part of why Feigenbaum's discovery, originally made by hand calculator on exactly this sequence, was so striking. The same ratio δ\delta appears for entirely different-looking unimodal maps, such as fμ(x)=μsin⁡(πx)f_\mu(x) = \mu \sin(\pi x) on [0,1][0,1], whose own bifurcation sequence rnsin⁡r_n^{\sin} satisfies the identical limit δ≈4.6692016\delta \approx 4.6692016 despite the sine map bearing no algebraic relation to the logistic parabola. This is the sense in which δ\delta is universal: it depends only on the local quadratic shape of the map near its maximum, not on the global formula.

Figure
Bifurcation diagram sketch: stable orbit values of the logistic map versus rr, showing successive doublings
r x* 1 3 r∞≈3.57 4 stable fixed point 2-cycle 4-cycle chaos + windows
CounterexampleDropping unimodality: a map with two humps can bypass the universal cascade

The theorem on Feigenbaum universality requires ff to be unimodal — a single smooth interior maximum on the interval, with negative Schwarzian derivative in the fully general statement (a technical condition automatically satisfied by the logistic and sine families). Drop this hypothesis by taking a bimodal map, e.g.

gμ(x)=μ x2(1−x)2⋅c,x∈[0,1],g_\mu(x) = \mu\, x^2(1-x)^2 \cdot c, \quad x\in[0,1],

rescaled so its range fits in [0,1][0,1]; more simply, consider

g(x)=4μ x(1−x)(1−2x)2/Mg(x) = 4\mu\, x(1-x)(1 - 2x)^2 \Big/ M

where MM normalizes the maximum to 11 — the point of the construction is only that gg has two local maxima on (0,1)(0,1) instead of one, separated by a local minimum. Such a map need not follow a period-doubling cascade governed by the single constant δ\delta: its bifurcation structure can instead follow the sequence associated with maps with two critical points, whose renormalization fixed point (and universal constants) is different from Feigenbaum's δ≈4.6692\delta \approx 4.6692. The mechanism of the proof above — reducing to the second-iterate map and tracking a single multiplier through −1-1 — still applies locally to any single bifurcation, but the global, universal, map-independent ratio δ\delta is a theorem about the renormalization operator restricted to unimodal maps; abandoning unimodality removes the map from the basin of that renormalization fixed point, and the accumulation ratio changes accordingly. This shows the unimodality hypothesis is not decorative: it is precisely what selects the universality class, and without it the numerical value 4.6692…4.6692\ldots need not appear at all.

Scope, edge cases & common traps

SituationWhat to checkCommon error
Fixed point with ∣f′(x∗)∣=1|f'(x^*)|=1Linear test is inconclusive; need higher-order (Schwarzian derivative) analysisDeclaring stability or instability from the linear test alone at rr exactly 33
Period-pp orbit stabilityMust use the multiplier Λ=(fp)′\Lambda=(f^p)', the product of derivatives around the cycle, not f′f' at a single pointTesting ∣f′(p)∣<1|f'(p)|<1 at one point of a 2-cycle and ignoring the other point
Domain of validityLogistic map maps [0,1]→[0,1][0,1]\to[0,1] only for r∈[0,4]r\in[0,4]Iterating with r>4r>4 and treating exits from [0,1][0,1] as meaningful orbit values
Bifurcation sequence rnr_nGenuinely infinite, accumulating at finite r∞≈3.569946r_\infty\approx3.569946Assuming the cascade continues past r=4r=4 or diverges to infinity
Universality constant δ\deltaApplies to the ratio of parameter gaps for unimodal maps, a renormalization-theoretic factExpecting δ\delta to equal some simple algebraic number derived from the logistic formula itself

The mathematician behind it

Origins · 1 minMitchell J. Feigenbaum and how this was discovered
MJ
Mitchell J. Feigenbaum
1944–2019 · Los Alamos National Laboratory, 1970s

Feigenbaum was a mathematical physicist who, using a programmable HP-65 calculator, computed successive period-doubling bifurcation values for the logistic map by hand and noticed the ratio of gaps was converging to approximately 4.66924.6692. He then discovered the identical constant governing entirely different unimodal maps, such as sin⁡(πx)\sin(\pi x), and developed the renormalization-group framework — borrowed conceptually from statistical mechanics phase-transition theory — that explains why the constant is universal: it is the multiplier of a fixed point of a doubling operator acting on the infinite-dimensional space of unimodal maps.

History
- 1825: Pierre François Verhulst introduces the continuous logistic differential equation for bounded population growth.
- 1976: Robert May's influential paper "Simple mathematical models with very complicated dynamics" popularizes the discrete logistic map and its route to chaos among ecologists and physicists.
- 1975–1978: Mitchell Feigenbaum discovers the universal ratio δ≈4.6692016\delta \approx 4.6692016 governing period-doubling cascades in unimodal maps, and develops the renormalization-group explanation.
- 1978: Feigenbaum publishes "Quantitative Universality for a Class of Nonlinear Transformations," establishing the theory rigorously for a broad class of maps.
- 1980s–present: Feigenbaum's constants δ\delta and α\alpha are confirmed experimentally in fluid convection, electronic circuits, lasers, and chemical oscillators, cementing period-doubling as a universal route to chaos alongside routes seen in continuous systems such as The Lorenz System.
Is the period-doubling bifurcation at r=3r=3 the same type of bifurcation as a pitchfork bifurcation in continuous systems?

It is closely analogous but not identical. In continuous-time pitchfork bifurcations, a fixed point's stability changes as a real eigenvalue crosses zero, spawning two new nearby fixed points. In discrete time, the analogous "eigenvalue" is the multiplier f′(x∗)f'(x^*), and the special discrete phenomenon is that it can cross −1-1, not just +1+1: this negative crossing is called a flip or period-doubling bifurcation and has no continuous-time analogue, because continuous flows do not have a notion of "period-2 in one time step." Crossing +1+1 in discrete time reproduces the familiar saddle-node/transcritical/pitchfork menu; crossing −1-1 is a genuinely new, purely discrete phenomenon.

Does every unimodal map undergo infinitely many period-doublings before r=4r=4, or could the cascade stop early?

For the logistic map specifically, the cascade is proved to be infinite and to accumulate at a finite r∞≈3.569946<4r_\infty\approx3.569946 < 4, after which chaotic parameter values dominate (interspersed with periodic windows of odd period, by results connected to Sharkovskii's theorem). This finiteness of r∞r_\infty within (0,4](0,4] is a specific, provable fact about this particular family; the general Feigenbaum universality theorem asserts that whenever such an infinite cascade occurs in a unimodal family satisfying the appropriate regularity hypotheses, the ratio of gaps converges to the same δ\delta, not that every unimodal family must have an infinite cascade before its own natural parameter range ends.

What exactly breaks if we only check ∣f′(x∗)∣=1|f'(x^*)|=1 and stop, instead of computing the second-iterate map?

The linear (first-derivative) test is inconclusive exactly at ∣f′(x∗)∣=1|f'(x^*)|=1: it tells you neither stability nor instability, and by itself gives zero information about what new orbit, if any, is created. Only by examining the next iterate — here fr2f_r^2 — and its own fixed points and their multipliers can you detect that a period-2 orbit is actually born and determine its own stability range. This is precisely the reduction used in the proof of the theorem: local bifurcations of pp-periodic orbits of ff are local bifurcations of fixed points of fpf^p, so higher iterates are not optional bookkeeping but the actual object whose bifurcation theory answers the question.

How is the Feigenbaum constant δ\delta actually computed or defined rigorously, beyond fitting numbers?

Rigorously, δ\delta is defined as the unique expanding eigenvalue of the linearization of the doubling operator T\mathcal{T} acting on suitably normalized unimodal maps, where (Tg)(x)=−1αg(g(αx))(\mathcal{T}g)(x) = -\frac{1}{\alpha} g\big(g(\alpha x)\big) for an appropriate rescaling constant α\alpha (the second Feigenbaum constant). Feigenbaum conjectured, and later work (notably by Lanford using computer-assisted rigorous analysis) proved, that T\mathcal{T} has an attracting fixed point g∗g^* in an appropriate function space, and the derivative of T\mathcal{T} at g∗g^* has a single unstable eigenvalue equal to δ\delta. This renormalization-group framework, not curve-fitting to the numerically computed rnr_n, is the rigorous content behind the empirically observed constant.

Can period-doubling and chaos in the logistic map be observed in a real physical experiment, or is it purely a mathematical curiosity?

It is directly observable. Classic experiments include Rayleigh–Bénard convection cells, driven nonlinear electronic oscillators (such as diode resonators), and certain laser systems, all of which show period-doubling cascades in an externally tuned control parameter (analogous to rr), with measured ratios of successive bifurcation-parameter spacings converging to values statistically consistent with δ≈4.6692\delta\approx4.6692. This experimental universality — wildly different physical systems obeying the same numerical constant — is considered one of the most striking confirmations that abstract one-dimensional map theory captures genuine, measurable structure in physical nonlinear systems, not merely an artifact of the logistic formula.

Before this