5D Library › Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies · 中文
Chapter 6 Skill Learning: Freezing and Freeing Degrees of Freedom, External Focus and “Instructions as Costs”
Freezing and freeing
Bernstein imagined a sequence for learning: the beginner first “freezes” the redundant degrees of freedom, for example letting several joints move stiffly together, to simplify the problem; with practice these degrees of freedom are gradually “freed”, until the learner can exploit the passive forces of the body and the environment. The idea was brought into experimental research by Vereijken and colleagues in 1992 (Vereijken et al. 1992).
Evidence
The ski simulator. Vereijken, van Emmerik, Whiting and Newell had novices practise a side-to-side movement on a ski simulator over several days, recording the angles of the joints of the legs and trunk. Early in practice the joint movements were small in amplitude and tightly coupled — “freezing”; with practice the amplitudes grew and the coupling weakened — “freeing” (Vereijken et al. 1992).
But this sequence is not a general law. Newell and Vaillancourt pointed out in 2001 that practice can increase or decrease the dimension of movement, the direction depending on what the task demands rather than on a fixed developmental order (Newell and Vaillancourt 2001). Konczak, vander Velden and Jaeger in 2009 compared, cross-sectionally, 14 violin learners with different amounts of practice (aged 4–47, ten of them children) and found that the more cumulative practice, the smaller the shoulder flexion–extension of the bowing arm, with no systematic change at the elbow; the paper’s title is “motor control by freezing, not freeing degrees of freedom” (Konczak et al. 2009). van Ginneken and colleagues found in a throwing task in 2018 that, in the group that practised with many errors, a self-rated propensity for conscious control was associated with “freezing” of the throwing arm (Ginneken et al. 2018). Yang and Scholz used UCM analysis to study the learning of a throwing task in 2005 and found that practice affected the synergy structure of different task variables differently (Yang and Scholz 2005).
“Freezing” itself has different operational definitions: smaller joint amplitudes, higher correlations between joints and reduced variability do not always occur together. The illustrative model below shows that even on one of these definitions, freezing can be understood as a reasonable response to an inaccurate internal model.
An illustrative model: freezing as a response to model error
Calculation
The model (the author’s illustration, not data). Three joints and one task variable , . In each movement the joints first err by (degrees); the learner then corrects in two ways: task correction , which corrects the observed task error along the direction the learner “believes” affects the task, ( is observation noise with SD 0.3); and posture stiffness , which pulls every joint back towards a template ( is proprioceptive noise with SD 0.5). The corrections themselves add signal-dependent noise (SD 0.5 times the correction). The learner’s estimate of the Jacobian carries a fresh random error of size on every trial, and practice shrinks . For each , and are chosen to minimise “task variance + 0.05 × mean squared correction”; covariances are computed analytically, with expectations over the error in taken from 4,000 draws. Finally, at , different posture stiffnesses are imposed, mimicking an instruction to “hold the joint angles”.
# A model of "freezing" and "freeing" degrees of freedom (the author's illustrative
# model, not data). Three joints, one task variable y = J q. Each trial the joints
# err by eta ~ N(0, I) (degrees). The learner then corrects in two ways:
# task correction -g * u * (J eta + s), u = Jh' / (Jh Jh'), s ~ N(0, 0.3^2)
# posture stiffness -lam * (eta + p), p ~ N(0, 0.5^2 I) (pull to a template)
# Corrections add signal-dependent noise (SD 0.5 x correction). Jh, the learner's
# internal estimate of J, is wrong by a fresh random error of size eps on every trial;
# practice shrinks eps. g and lam minimise task variance + 0.05 x mean sq. correction.
import math, random
J = [0.6, 0.5, 0.3]; SS, SP, C, ALPHA = 0.3, 0.5, 0.5, 0.05
EYE = [[float(i == j) for j in range(3)] for i in range(3)]
def moments(eps, rng, draws=4000): # E[u] and E[u u'] over model errors
Eu = [0.0] * 3; Euu = [[0.0] * 3 for _ in range(3)]
for _ in range(draws):
Jh = [j + eps * rng.gauss(0, 1) / math.sqrt(3) for j in J]
nh = sum(v * v for v in Jh); u = [v / nh for v in Jh]
for i in range(3):
Eu[i] += u[i] / draws
for k in range(3):
Euu[i][k] += u[i] * u[k] / draws
return Eu, Euu
def covs(g, lam, Eu, Euu):
jj = sum(v * v for v in J)
cross = [[Eu[i] * J[k] + J[i] * Eu[k] for k in range(3)] for i in range(3)]
corr = [[(lam ** 2 * (1 + SP ** 2)) * EYE[i][k] + lam * g * cross[i][k]
+ g * g * (jj + SS ** 2) * Euu[i][k] for k in range(3)] for i in range(3)]
fin = [[((1 - lam) ** 2 + lam ** 2 * SP ** 2 + C * C * corr[i][i]) * EYE[i][k]
- (1 - lam) * g * cross[i][k] + g * g * (jj + SS ** 2) * Euu[i][k]
for k in range(3)] for i in range(3)]
return fin, corr
def summary(S):
nj = math.sqrt(sum(v * v for v in J)); e = [v / nj for v in J]
vo = sum(e[i] * S[i][j] * e[j] for i in range(3) for j in range(3))
vu = (sum(S[i][i] for i in range(3)) - vo) / 2 # per DOF, UCM is 2-D
task = sum(J[i] * S[i][j] * J[j] for i in range(3) for j in range(3))
rr = [abs(S[i][j]) / math.sqrt(S[i][i] * S[j][j])
for i, j in ((0, 1), (0, 2), (1, 2))]
return math.sqrt(task), math.log(vu / vo), sum(rr) / 3, \
math.sqrt(sum(S[i][i] for i in range(3)) / 3)
def optimise(Eu, Euu, lam_fixed=None):
best = None
lams = [lam_fixed] if lam_fixed is not None else [i / 100 for i in range(101)]
for g in (i / 100 for i in range(151)):
for lam in lams:
fin, corr = covs(g, lam, Eu, Euu)
cost = sum(J[i] * fin[i][j] * J[j] for i in range(3) for j in range(3)) \
+ ALPHA * sum(corr[i][i] for i in range(3)) / 3
if best is None or cost < best[0]:
best = (cost, g, lam, fin)
return best
if __name__ == "__main__":
rng = random.Random(1992)
print("model error g* lam* task SD ln(Vucm/Vort) mean|r| joint SD")
for eps in (0.6, 0.45, 0.3, 0.2, 0.1, 0.0):
_, g, lam, fin = optimise(*moments(eps, rng)); t, lr, r, js = summary(fin)
print(f" {eps:4.2f} {g:4.2f} {lam:4.2f} {t:6.3f} {lr:+6.2f}"
f" {r:5.2f} {js:5.2f}")
print("accurate model (eps = 0) with posture stiffness imposed (an 'internal' cue):")
m0 = moments(0.0, rng)
for lam in (0.3, 0.6, 0.9):
_, g, _, fin = optimise(*m0, lam); t, lr, r, js = summary(fin)
print(f" lam = {lam:.1f}: g* {g:4.2f}, task SD {t:5.3f}, ln(Vucm/Vort)"
f" {lr:+5.2f}, joint SD {js:4.2f}")
# model error g* lam* task SD ln(Vucm/Vort) mean|r| joint SD
# 0.60 0.43 0.37 0.427 +0.78 0.21 0.68
# 0.45 0.50 0.30 0.399 +1.01 0.26 0.70
# 0.30 0.58 0.23 0.372 +1.25 0.30 0.72
# 0.20 0.60 0.20 0.360 +1.36 0.31 0.74
# 0.10 0.63 0.18 0.352 +1.43 0.33 0.74
# 0.00 0.64 0.17 0.349 +1.47 0.33 0.75
# accurate model (eps = 0) with posture stiffness imposed (an 'internal' cue):
# lam = 0.3: g* 0.52, task SD 0.356, ln(Vucm/Vort) +1.18, joint SD 0.67
# lam = 0.6: g* 0.26, task SD 0.435, ln(Vucm/Vort) +0.37, joint SD 0.59
# lam = 0.9: g* 0.00, task SD 0.571, ln(Vucm/Vort) +0.00, joint SD 0.68When model error is large (), the best strategy uses less task correction () and more posture stiffness (): if it is uncertain which direction affects the task, pull every joint back to the template. Joint variability is then smaller (SD 0.68 per joint), is only +0.78 and the task SD is 0.427. As model error shrinks, rises to 0.64, falls to 0.17, joint variability grows to 0.75, the log ratio rises to +1.47 and the task SD falls to 0.349 (Figure 5, right). In this model, then, “freezing → freeing” and “a strengthening UCM structure” are two faces of one process: the better the learner knows the structure of the task, the more freely variability that does not affect the task can be left alone.
The mean correlation between joints in the model, however, does not rise early on as Vereijken and colleagues described (it goes from 0.21 to 0.33): “freezing” here is a matter of amplitude and stiffness, not of coupling. This is a reminder that different definitions of “freezing” may correspond to different mechanisms.
The last three lines of the model answer another question: if the model is already accurate but the learner is told to “hold the joint angles” (imposed ), what happens? At the task SD rises from 0.349 to 0.356, at to 0.435 and at to 0.571, while the log ratio falls from +1.47 to +1.18, +0.37 and 0.00. The imposed stiffness crowds out task correction ( falls from 0.64 to 0.52, 0.26 and 0), so the task gets worse.
Hypothesis · not established · Freezing as a rational response to internal-model uncertainty
The author’s interpretation: the “freezing” of beginners (small joint amplitudes, high stiffness, weak UCM structure) is at least partly a reasonable response to an inaccurate internal model, not merely a lack of ability. Testable predictions: in the same task, artificially increasing the learner’s uncertainty about the task structure (for example by randomly changing the mapping from joints to cursor during practice) should make learners who have already “freed” their degrees of freedom show smaller joint amplitudes and lower again, while explicit feedback about the task structure (for example a display of the solution manifold) should speed up “freeing”. What would refute it: uncertainty in the mapping has no effect on joint amplitudes and UCM structure, or an effect in the opposite direction.
External focus: recap and additions
Book 10 has discussed the research on attentional focus in detail; it is only recalled here, with the additions useful for transfer and trial design.
Evidence
Research on external focus. Wulf, Höß and Prinz in 1998 compared two instructions on a ski simulator and a balance platform: attending to one’s body movements (internal focus) or to the effect of the movement on the apparatus (external focus); the external-focus group learned better in retention tests (Wulf et al. 1998). Wulf, McNevin and Shea in 2001 proposed the “constrained action hypothesis”: an internal focus leads people to intervene consciously in control processes that ought to be automatic (Wulf et al. 2001). Wulf’s 2013 review summarised fifteen years of research (Wulf 2013). Masters found in 1992, in golf putting, that people who learned with explicit knowledge performed worse under pressure, whereas those who learned implicitly did not (Masters 1992).
Evidence
Meta-analysis and reporting bias. The 2021 meta-analysis in Psychological Bulletin by Chua, Jimenez-Diaz, Lewthwaite, Kim and Wulf reported effect sizes (Hedges’ ) of external over internal focus of 0.264 for performance tests (73 studies, 1,824 participants), 0.583 for retention, 0.584 for transfer (95 % confidence interval 0.325 to 0.842) and 0.833 for EMG measures (Chua et al. 2021). Correcting for reporting bias with a robust Bayesian meta-analysis, McKay and colleagues obtained 0.01 for performance, 0.15 for retention, 0.09 for transfer and 0.06 for EMG, with Bayes factors favouring a null effect in every analysis ( from 1.3 to 5.75) and moderate to strong evidence of publication bias in all analyses (McKay et al. 2024). The paper was published in Psychological Bulletin in 2024; the figures above are from version 3 of the preprint, updated in April 2024 (the month the journal version was accepted), and the full journal version was not seen.
For this book the two sets of figures have two implications. First, transfer is exactly the primary outcome of the Chapter 8 trial, and it is where the two estimates diverge most: about 0.58 standard deviations uncorrected, under 0.1 corrected. Second, if the true effect is only about 0.1 standard deviations, a trial designed for 0.375 standard deviations will most likely conclude “no difference worth acting on”; the sample-size program of Chapter 8 computes this. Book 33 discussed how publication bias and small studies together inflate effects, and that is not repeated here.
Instructions as cost functions
Why would an external focus be better (if it is)? The optimal feedback control of Chapter 2 offers one way to put it: an instruction changes the learner’s cost function. An external focus (“keep the cursor on the target”) pays only for the task result; an internal focus (“keep the two joints in the prescribed ratio”) also pays for task-irrelevant directions.
Calculation
The program reuses the two-effector sum task of Chapter 2, but with signal-dependent motor noise (SD N). The “external” instruction pays only for the sum of positions; the “internal” instruction adds , i.e. asks the two effectors to keep equal shares — note that this constrains only the task-irrelevant direction and leaves the cost on the sum unchanged. Feedback gains are computed for each cost function and the outcome is simulated 3,000 times under signal-dependent noise, with the same random numbers in every row.
# Instructions as cost functions (the author's model, not data). The two-effector sum
# task of p03 with SIGNAL-DEPENDENT noise: each motor command u carries noise of SD
# 0.25 |u| + 0.5 N. "External" cue: pay only for the task sum p1 + p2 = 0.20 m.
# "Internal" cue: also pay lam x (p1 - p2)^2, i.e. keep the two effectors on a fixed
# template (equal shares). The weight on the task sum itself is unchanged, so the cue
# only constrains the task-irrelevant direction. Gains are the LQR gains for each cost
# (certainty-equivalent design); outcomes are simulated with the signal-dependent
# noise (3,000 trials each).
import math, random
DT, TAU, N, TARGET, C_SDN, C_ADD = 0.01, 0.04, 50, 0.20, 0.25, 0.5
def mm(A, B):
return [[sum(a * b for a, b in zip(r, c)) for c in zip(*B)] for r in A]
def tr(A):
return [list(c) for c in zip(*A)]
def add(*Ms):
return [[sum(v) for v in zip(*rows)] for rows in zip(*Ms)]
def inv2(M):
d = M[0][0] * M[1][1] - M[0][1] * M[1][0]
return [[M[1][1] / d, -M[0][1] / d], [-M[1][0] / d, M[0][0] / d]]
A = [[0.0] * 7 for _ in range(7)]; B = [[0.0] * 2 for _ in range(7)]
for k in (0, 3):
A[k][k] = 1; A[k][k + 1] = DT; A[k + 1][k + 1] = 1; A[k + 1][k + 2] = DT
A[k + 2][k + 2] = 1 - DT / TAU; B[k + 2][k // 3] = DT / TAU
A[6][6] = 1
def outer(c, w):
return [[w * a * b for b in c] for a in c]
def cost(lam):
Q = add(outer([1, 0, 0, 1, 0, 0, -TARGET], 1.0),
*[outer([float(i == j) for j in range(7)], w)
for i, w in ((1, 0.05), (4, 0.05), (2, 1e-4), (5, 1e-4))])
if lam:
Q = add(Q, outer([1, 0, 0, -1, 0, 0, 0], lam))
return Q
def gains(Q, r=1e-6):
S, L = Q, []
for _ in range(N):
K = mm(inv2(add(mm(mm(tr(B), S), B), [[r, 0], [0, r]])), mm(mm(tr(B), S), A))
S = mm(mm(tr(A), S), add(A, [[-v for v in row] for row in mm(B, K)]))
L.append(K)
return L[::-1]
def simulate(L, rng, trials=3000):
out = []
for _ in range(trials):
x = [0.0] * 6 + [1.0]; eff = 0.0
for t in range(N):
u = [-sum(k * v for k, v in zip(row, x)) for row in L[t]]
eff += sum(v * v for v in u) * DT
u = [v + rng.gauss(0, C_SDN * abs(v) + C_ADD) for v in u]
x = [sum(a * v for a, v in zip(row, x)) + sum(b * w for b, w in zip(br, u))
for row, br in zip(A, B)]
out.append((x[0], x[3], eff))
return out
def var(v):
m = sum(v) / len(v); return sum((a - m) ** 2 for a in v) / (len(v) - 1)
if __name__ == "__main__":
print("cue task SD bias ln(Vucm/Vort) mean effort")
for name, lam in (("external (task only)", 0.0), ("internal, lam = 0.1", 0.1),
("internal, lam = 1", 1.0), ("internal, lam = 10", 10.0)):
s = simulate(gains(cost(lam)), random.Random(2013)) # common random numbers
so = [a + b for a, b, _ in s]; su = [a - b for a, b, _ in s]
bias = 1000 * (sum(so) / len(so) - TARGET); eff = sum(e for *_, e in s) / len(s)
print(f"{name:26s} {1000 * math.sqrt(var(so)):6.2f} mm {bias:+5.2f} mm"
f" {math.log(var(su) / var(so)):+6.2f} {eff:6.3f}")
print("Monte Carlo SE of each task SD about 1.3 %; same noise draws in every row")
# cue task SD bias ln(Vucm/Vort) mean effort
# external (task only) 1.11 mm -1.45 mm +4.69 2.959
# internal, lam = 0.1 1.11 mm -1.45 mm +1.77 2.974
# internal, lam = 1 1.11 mm -1.45 mm +0.01 2.991
# internal, lam = 10 1.11 mm -1.46 mm -1.94 3.019
# Monte Carlo SE of each task SD about 1.3 %; same noise draws in every rowThe result is surprisingly “flat”: as goes from 0 to 10, falls from +4.69 to −1.94, a change of 6.6 units, while the task SD stays at 1.11 mm throughout (Monte Carlo standard error about 1.3 %), the mean error stays around −1.45 mm and effort rises by only 2 %. In this model an instruction that constrains only the task-irrelevant direction barely affects the task result: the extra corrections are small, and so is the signal-dependent noise they add. The contrast with the static model of the previous section is instructive: there the imposed posture stiffness also pulled on the task-relevant direction and crowded out task correction, so the task got worse.
The result has two direct implications for the trial design of Chapter 8. First, can change greatly while the task result does not change at all, so it should not be the primary outcome; the primary outcome should be the accuracy of the task itself. Second, if external focus has an advantage in task accuracy, the advantage cannot be explained by “leaving UCM variability alone” alone; other mechanisms are needed, for example whether the internal instruction also interferes with task-relevant corrections, whether it adds attentional load, or whether UCM variability shows its value only when transferring to new conditions (the hypothesis of Chapter 1).
Summary
“Freezing and freeing” is the learning process described by Vereijken and colleagues on a ski simulator in 1992, but it is not a general law: dimension can rise or fall with practice, and violin learners with more practice show less shoulder motion, not more. In an illustrative model, freezing (less task correction, more posture stiffness) is a reasonable response to internal-model error, and as the model error shrinks rises from +0.78 to +1.47. The meta-analytic external-focus effect (0.584 for transfer) falls below about 0.1 after correction for reporting bias. Treating instructions as cost functions, an “internal” instruction that constrains only the task-irrelevant direction can move the log ratio from +4.69 to −1.94 with the task SD unchanged; only an imposed stiffness that also acts on the task direction makes the task worse. The next chapter reads these studies beside ancient texts on skill.
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).