5D Library › Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies · 中文
Chapter 1 Bernstein’s Problem: Redundancy, Abundance and “Repetition without Repetition”
Where the problem came from
From the 1920s onwards Nikolai Bernstein (1896–1966) used photographic methods to record the movements of workers, athletes and children and to study how movement is organised. Most of his papers appeared in Russian, and the English-speaking world came to know his work mainly through the translated collection of 1967, The Co-ordination and Regulation of Movements (Bernstein 1967). Its most quoted passage is his definition of coordination: coordination is the process of mastering the redundant degrees of freedom of the moving organ, that is, of converting it into a controllable system (paraphrased from the usual English quotation; the page in the original volume was not checked for this book).
“Redundant” is relative to a task. The human arm from shoulder to wrist has about seven joint degrees of freedom: three at the shoulder, one at the elbow, one for forearm rotation and two at the wrist. If the task is only to place the fingertip at a point in space, three coordinates must be fixed and the other four degrees of freedom may take any value; if the orientation of the hand is also required (six coordinates in all), one is still left over. One level down, each joint is pulled by several muscles, and the same joint torque can be produced by different combinations of muscles; one level further down, each muscle consists of many motor units. In his 1990 review, Turvey put the problem this way: if the nervous system had to specify the value of every degree of freedom individually, the control problem would be unmanageably large, so coordination must somehow reduce the number of quantities that are controlled independently (Turvey 1990).
Calculation
Counting degrees of freedom. Let the joint space have dimension and the task variable dimension (). Near a given posture , the task variable linearises to , where is the Jacobian. If has full rank, the joint changes with form an -dimensional null space: changing the joint angles along these directions leaves the task variable unchanged (to first order). For the arm and the fingertip position, ; for the position and orientation of the hand, .
Redundancy or abundance
Bernstein framed the problem as “overcoming” or “mastering” redundant degrees of freedom, and later literature often calls it the motor redundancy problem. The framing carries a judgement: the extra degrees of freedom are a nuisance that the nervous system must somehow eliminate, for example by “locking” several joints together or by using a cost function to pick out a single solution.
In 1998 Gelfand and Latash proposed a different framing, which they called the principle of abundance (Gelfand and Latash 1998). In a 2012 mini-review entitled “The bliss (not the problem) of motor abundance (not redundancy)”, Latash reviewed the studies that support it and argued that the problem had been formulated incorrectly: the nervous system does not eliminate the extra degrees of freedom but uses them, letting many combinations of joints or muscles accomplish the task, thereby stabilising the important variables, coping with perturbations and performing several tasks at once (Latash 2012).
The two framings do not differ in their description of anatomy (the extra degrees of freedom are certainly there) but in their assumption about what the nervous system does. Under the redundancy framing we should expect the joint combination to be roughly the same from one movement to the next (the problem has been eliminated); under the abundance framing we should expect the combination to differ every time, but with the differences coordinated so that the task variable stays stable. The uncontrolled-manifold analysis of Chapter 2 was designed to tell these cases apart. Note that “abundance” is a theoretical framing, not a measurable “degree of abundance”; what can be measured is how variability is distributed over different directions.
A curve of solutions
The simplest example makes the problem concrete: a three-link arm moving in a plane (shoulder, elbow and wrist), whose task is to place the fingertip at a point in the plane. Here and , leaving one spare degree of freedom, so the joint combinations that reach the same point form a curve.
Calculation
The program first computes the null-space direction of the Jacobian at a nominal posture (shoulder 40°, elbow 70°, wrist 30°). It then sweeps the orientation of the last link from −180° to 180°, solves for the first two joint angles at each orientation, and keeps the solutions that lie within the (assumed) joint ranges. The second part simulates 200 movements: each time a new solution is chosen (SD of the last-link orientation 8°), plus independent noise of 0.5° in every joint; as a contrast, joint variability of the same size is made independent across joints, and the spread of the fingertip is compared.
# Bernstein's problem in miniature: a planar three-link arm (shoulder, elbow, wrist;
# lengths 0.30, 0.28, 0.18 m) must put its fingertip on a point. 3 joints, 2 task
# coordinates: one redundant degree of freedom. Part 1 traces the whole set of joint
# configurations that reach the same point; part 2 compares "repetition without
# repetition" (a new solution every trial) with rigid repetition of one solution.
import math, random
L = (0.30, 0.28, 0.18)
LIMITS = ((-30, 150), (0, 150), (-60, 90)) # joint ranges, degrees (assumed)
def tip(q): # q in radians, relative angles
a = x = y = 0.0
for li, qi in zip(L, q):
a += qi; x += li * math.cos(a); y += li * math.sin(a)
return x, y
def jacobian(q):
a, cum = 0.0, []
for qi in q:
a += qi; cum.append(a)
return [[-sum(L[k] * math.sin(cum[k]) for k in range(j, 3)) for j in range(3)],
[sum(L[k] * math.cos(cum[k]) for k in range(j, 3)) for j in range(3)]]
def ik(target, phi): # end-link orientation phi fixed -> two-link problem
wx = target[0] - L[2] * math.cos(phi); wy = target[1] - L[2] * math.sin(phi)
c2 = (wx * wx + wy * wy - L[0] ** 2 - L[1] ** 2) / (2 * L[0] * L[1])
if abs(c2) > 1:
return None
q2 = math.acos(c2) # elbow-flexed branch
q1 = math.atan2(wy, wx) - math.atan2(L[1] * math.sin(q2), L[0] + L[1] * math.cos(q2))
return q1, q2, phi - q1 - q2
def ok(q):
return all(lo <= math.degrees(v) <= hi for v, (lo, hi) in zip(q, LIMITS))
def sd(v):
m = sum(v) / len(v)
return math.sqrt(sum((x - m) ** 2 for x in v) / (len(v) - 1))
if __name__ == "__main__":
q0 = tuple(math.radians(d) for d in (40, 70, 30)); T = tip(q0)
J = jacobian(q0)
n = [J[0][1] * J[1][2] - J[0][2] * J[1][1], J[0][2] * J[1][0] - J[0][0] * J[1][2],
J[0][0] * J[1][1] - J[0][1] * J[1][0]] # cross product: null direction
s = math.sqrt(sum(v * v for v in n)); n = [v / s for v in n]
print(f"target ({T[0]:.3f}, {T[1]:.3f}) m; unit null-space direction of J at the"
f" nominal posture:\n ({n[0]:+.3f}, {n[1]:+.3f}, {n[2]:+.3f})"
" (shoulder, elbow, wrist)")
sols = [q for k in range(-1800, 1801)
if (q := ik(T, math.radians(k / 10))) and ok(q)]
for j, name in enumerate(("shoulder", "elbow", "wrist")):
v = [math.degrees(q[j]) for q in sols]
print(f" {name:8s} angle over the solution set:"
f" {min(v):6.1f} to {max(v):6.1f} deg")
worst = max(math.dist(tip(q), T) for q in sols)
print(f"{len(sols)} sampled solutions (elbow-flexed branch) within joint limits;"
f" max miss {worst:.1e} m")
rng = random.Random(1967); phi0 = sum(q0)
trials = [] # repetition without repetition
for _ in range(200):
q = ik(T, phi0 + math.radians(rng.gauss(0, 8)))
trials.append([v + math.radians(rng.gauss(0, 0.5)) for v in q])
jsd = [math.degrees(sd([q[j] for q in trials])) for j in range(3)]
tips = [tip(q) for q in trials]
esd = 1000 * math.hypot(sd([p[0] for p in tips]), sd([p[1] for p in tips]))
print("200 trials, a new solution each time (end-link SD 8 deg + 0.5 deg noise):")
print(f" joint SDs {jsd[0]:.1f}, {jsd[1]:.1f}, {jsd[2]:.1f} deg;"
f" fingertip SD {esd:.1f} mm")
tot = math.sqrt(sum(v * v for v in jsd) / 3)
rigid = [tip([v + math.radians(rng.gauss(0, tot)) for v in q0]) for _ in range(200)]
rsd = 1000 * math.hypot(sd([p[0] for p in rigid]), sd([p[1] for p in rigid]))
print(f"same total joint variance, independent joints ({tot:.1f} deg each):"
f" fingertip SD {rsd:.1f} mm")
print("degrees of freedom: arm 7 (shoulder 3, elbow 1, forearm 1, wrist 2);")
print(" fingertip point (3) leaves 4 spare; full hand pose (6) leaves 1 spare")
# target (-0.004, 0.572) m; unit null-space direction of J at the nominal posture:
# (+0.162, -0.504, +0.849) (shoulder, elbow, wrist)
# shoulder angle over the solution set: 37.1 to 63.4 deg
# elbow angle over the solution set: 18.4 to 95.1 deg
# wrist angle over the solution set: -59.9 to 89.8 deg
# 885 sampled solutions (elbow-flexed branch) within joint limits; max miss 2.8e-16 m
# 200 trials, a new solution each time (end-link SD 8 deg + 0.5 deg noise):
# joint SDs 2.8, 8.3, 14.1 deg; fingertip SD 6.5 mm
# same total joint variance, independent joints (9.6 deg each): fingertip SD 121.4 mm
# degrees of freedom: arm 7 (shoulder 3, elbow 1, forearm 1, wrist 2);
# fingertip point (3) leaves 4 spare; full hand pose (6) leaves 1 spareAt the nominal posture the null-space direction is about (shoulder +0.16, elbow −0.50, wrist +0.85): flexing the wrist, extending the elbow and slightly flexing the shoulder together barely moves the fingertip. On the elbow-flexed branch, all 885 sampled solutions respect the joint ranges, with fingertip errors at the level of numerical precision ( m). Across these solutions the shoulder angle ranges from 37.1° to 63.4°, the elbow from 18.4° to 95.1°, and the wrist across almost the whole assumed range (−59.9° to 89.8°). A single spare degree of freedom thus lets one fingertip position correspond to very different postures (Figure 1, left).
Repetition without repetition
Another of Bernstein’s famous phrases is “repetition without repetition”. In the usual quotation (secondary sources give p. 134 of the 1967 English volume) he was talking about practice: properly undertaken practice does not repeat the same means of solving a motor problem again and again, but solves the problem again and again, with techniques that are changed and improved from one repetition to the next (Bernstein 1967). It is first a claim about the learning process, and only secondarily a description of movement variability.
The second part of the program turns the latter meaning into a kinematic model (the author’s illustration, not Bernstein’s model). If every movement uses a new solution but all solutions lie near the solution curve, the joint SDs can be large (shoulder 2.8°, elbow 8.3°, wrist 14.1°) while the fingertip SD is only 6.5 mm. If joint variability of the same size (about 9.6° per joint) is independent across joints, the fingertip SD is 121.4 mm, nearly 19 times larger. Joint variability by itself says nothing about the quality of control; what matters is the direction in which it falls.
Note
What this model does not answer. The program assumes that variability falls near the solution curve and then computes the consequences; it does not say how the nervous system achieves this. There are at least three possible answers: choose one solution in advance and lock the remaining degrees of freedom (“freezing”, Chapter 6); organise the joints into a few fixed combinations (“synergies”, Chapter 5); or keep correcting the deviations that affect the task while leaving the rest alone (“minimal intervention”, Chapter 2). The three answers make different predictions about data, which is what the following chapters examine.
Ways of analysing variability
To tell these answers apart in data, variability in joint space has to be split into a part that affects the task and a part that does not. The uncontrolled-manifold analysis of Chapter 2 is the most widely used method: at each moment, the joint deviations of the individual trials are projected onto the null space (no effect on the task) and onto its orthogonal complement (effect on the task), and the variances per degree of freedom are compared (Scholz and Schöner 1999). Another method is the “tolerance–noise–covariation” decomposition of Müller and Sternad, which asks directly how much of the variability of the task result comes from choosing a region with a large tolerance, how much from the size of the noise and how much from covariation among the variables (Müller and Sternad 2004). The two methods do not answer quite the same question: the first concerns the structure of variance in joint space, the second the contribution of that structure to the task result. Chapter 2 will also show that the conclusions of the first depend on the choice of coordinates.
Hypothesis · not established · Variability as the raw material of practice
The author’s interpretation: under the abundance framing, variability in the null space is not merely “tolerated noise” but may also be raw material for learning — the learner keeps trying different solutions along directions that do not affect the current task, and finds among them solutions that work better under new conditions; this is one reading of “repetition without repetition”. If the interpretation holds, artificially compressing null-space variability during practice (for example by requiring the joint angles to stay fixed) should impair transfer to new task conditions, not merely change the synergy index. What would refute it: in a pre-registered randomised controlled trial (for example the protocol of Chapter 8), a practice condition that compresses null-space variability shows no worse transfer accuracy than the control condition, or better.
Summary
Bernstein defined the coordination of movement as the process of mastering redundant degrees of freedom. In a planar three-link arm, one spare degree of freedom is enough for a single fingertip position to correspond to a whole curve of solutions, with the elbow angle varying between 18° and 95°. The “redundancy” framing treats extra degrees of freedom as a nuisance to be eliminated, the “abundance” framing as a resource for stabilising the task; the latter is a reformulation of the problem, not a measurable quantity. “Repetition without repetition” is first a claim about practice; when it is written as a kinematic model, the key is the direction of variability: joint variability along the solution curve (SD up to 14.1°) spreads the fingertip by only 6.5 mm, whereas independent variability of the same size spreads it by 121.4 mm. The next chapter shows how this directionality can be measured in data, and what kind of controller produces it.
References
Fan, Y. (2026). Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies: From the Uncontrolled Manifold Hypothesis and Internal Forward Models to Fitts's Law, Read Beside Classical Chinese Thought on Skill and the Way [具身认知、伯恩斯坦自由度问题与运动协同的控制理论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828057 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).