5D Library › Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies · 中文

Chapter 2 The Uncontrolled Manifold and the Minimal Intervention Principle: Control Only What Matters

Splitting the variance in two

The three-link model of the previous chapter assumed that variability falls near the solution curve. To test this in real data, Scholz and Schöner proposed the uncontrolled manifold (UCM) analysis in 1999 (Scholz and Schöner 1999). The idea is direct. First propose a hypothetical task variable (for example the horizontal position of the centre of mass); at each moment find the joint combinations that leave it unchanged, which form a manifold in joint space (after linearisation, the null space of the Jacobian); then split the joint deviations of the individual trials into a part along the manifold and a part perpendicular to it, and compare their sizes. If the variance along the manifold is clearly larger than the variance perpendicular to it, the task variable is “stabilised”: changes in the joints compensate one another so that it stays constant.

Evidence

The original study. Scholz and Schöner studied standing up from a chair: nine young adults (five women, four men) under three conditions — normal, wearing ski boots that restricted the ankle, and on a narrow base of support — with posture described by eight segment angles in the sagittal plane (Scholz and Schöner 1999). They tested separately whether the horizontal and vertical positions of the centre of mass, the head and the hand were stabilised variables. The centre of mass was strongly stabilised in both directions throughout; the horizontal head position was also stabilised, but much more weakly (and not before leaving the seat); the horizontal hand position was stabilised throughout, the vertical only at about 45 % of the movement. Variances were computed per degree of freedom so that hypotheses of different dimension could be compared.

Calculation

Variance per degree of freedom and the synergy index. Let joint space have nn dimensions and the task variable dd, let δ𝒒i\delta\mathbf{q}_i be the joint deviation of trial ii (minus the mean over trials), and let PP project onto the row space of JJ (ORT, dd dimensions), so that I−PI – P projects onto the null space (UCM, n−dn-d dimensions). With NN trials, VUCM=∑i∥(I−P)δ𝒒i∥2(N−1)(n−d),VORT=∑i∥Pδ𝒒i∥2(N−1)d.V_{\mathrm{UCM}} = \frac{\sum_i \|(I-P)\,\delta\mathbf{q}_i\|^2}{(N-1)(n-d)}, \qquad V_{\mathrm{ORT}} = \frac{\sum_i \|P\,\delta\mathbf{q}_i\|^2}{(N-1)\,d}. Later studies often use the synergy index ΔV=(VUCM−VORT)/VTOT\Delta V = (V_{\mathrm{UCM}} – V_{\mathrm{ORT}})/V_{\mathrm{TOT}}, where VTOTV_{\mathrm{TOT}} is the total variance per degree of freedom (Latash et al. 2001, 2002). ΔV\Delta V is bounded: −n/d-n/d when all the variance lies in ORT and n/(n−d)n/(n-d) when it all lies in the UCM. For statistical testing one may apply a log transform that sends the bounds to infinity (similar to Fisher’s zz transform); the general form written out in this book is ΔVz=12ln⁡ΔV+n/dn/(n−d)−ΔV.\Delta V_z = \tfrac12 \ln\frac{\Delta V + n/d}{\,n/(n-d) – \Delta V\,}. Note that at ΔV=0\Delta V = 0, ΔVz=12ln⁡((n−d)/d)\Delta V_z = \tfrac12\ln\big((n-d)/d\big), which is zero only when n=2dn = 2d. The ln⁡(VUCM/VORT)\ln(V_{\mathrm{UCM}}/V_{\mathrm{ORT}}) used in the framework is unbounded and zero for independent noise of equal variance; this book uses it as an auxiliary index.

Latash, Scholz and Schöner later defined a “synergy” as a neural organisation that shares a task among a set of elements (joints, fingers, muscles) and makes the changes of the elements co-vary so as to stabilise a performance variable (Latash et al. 2007). On this definition a synergy is not a fixed combination of joints but a structure of variability, and ΔV>0\Delta V > 0 is its operational index.

Three traps in the index

The program below examines these indices with the three-link arm of Chapter 1 (fingertip position, n=3n = 3, d=2d = 2). There are two kinds of data: noise that is independent across joints with equal SDs (no synergy), and noise with an extra spread along the null space (synergy). Each is analysed both in relative joint angles (the angles between shoulder, elbow and wrist) and in absolute segment angles (each segment’s angle to the horizontal); finally the spread of the estimates is examined when there are only 20 trials each time.

Calculation

# UCM analysis of the three-link arm of p01 (n = 3 joints, task = fingertip x, y,
# d = 2). Joint deviations from the mean are split into the null space of J (UCM,
# n - d = 1 dimension) and its orthogonal complement (ORT, d = 2). Variances are
# compared PER DEGREE OF FREEDOM. Also: coordinate dependence and sampling error.
import math, random

def matmul(A, B):
    return [[sum(a * b for a, b in zip(r, c)) for c in zip(*B)] for r in A]

def inv(M):                                          # Gauss-Jordan, small matrices
    n = len(M)
    A = [row[:] + [float(i == j) for j in range(n)] for i, row in enumerate(M)]
    for c in range(n):
        p = max(range(c, n), key=lambda r: abs(A[r][c])); A[c], A[p] = A[p], A[c]
        A[c] = [v / A[c][c] for v in A[c]]
        for r in range(n):
            if r != c:
                A[r] = [a - A[r][c] * b for a, b in zip(A[r], A[c])]
    return [row[n:] for row in A]

def ucm(dev, J):
    """dev: list of joint-deviation vectors; J: d x n. Returns V_UCM, V_ORT per DOF."""
    d, n = len(J), len(J[0]); Jt = [list(c) for c in zip(*J)]
    P = matmul(matmul(Jt, inv(matmul(J, Jt))), J)    # projector onto ORT (row space)
    m = [sum(v[i] for v in dev) / len(dev) for i in range(n)]
    so = su = 0.0
    for v in dev:
        x = [a - b for a, b in zip(v, m)]
        po = [sum(P[i][j] * x[j] for j in range(n)) for i in range(n)]
        so += sum(t * t for t in po); su += sum((a - b) ** 2 for a, b in zip(x, po))
    k = len(dev) - 1
    return su / k / (n - d), so / k / d

def delta_v(vu, vo, n, d):
    vt = ((n - d) * vu + d * vo) / n
    dv = (vu - vo) / vt                              # bounded by -n/d and n/(n-d)
    return dv, 0.5 * math.log((dv + n / d) / (n / (n - d) - dv))

L = (0.30, 0.28, 0.18); Q0 = [math.radians(d) for d in (40, 70, 30)]

def jac(q):
    c = [sum(q[:k + 1]) for k in range(3)]
    return [[-sum(L[k] * math.sin(c[k]) for k in range(j, 3)) for j in range(3)],
            [sum(L[k] * math.cos(c[k]) for k in range(j, 3)) for j in range(3)]]

def sample(rng, kind, trials, s=math.radians(2.0)):
    J = jac(Q0); P = matmul(matmul([list(c) for c in zip(*J)],
                                   inv(matmul(J, [list(c) for c in zip(*J)]))), J)
    nul = [1 - P[0][0], -P[1][0], -P[2][0]]; z = math.sqrt(sum(v * v for v in nul))
    nul = [v / z for v in nul]
    out = []
    for _ in range(trials):
        e = [rng.gauss(0, s) for _ in range(3)]
        if kind == "synergy":                        # extra spread along the UCM
            g = rng.gauss(0, 3 * s); e = [a * 0.5 + g * b for a, b in zip(e, nul)]
        out.append(e)
    return out

def report(name, dev, J, n=3, d=2):
    vu, vo = ucm(dev, J); dv, dz = delta_v(vu, vo, n, d)
    print(f"  {name:37s} ln(Vucm/Vort) {math.log(vu / vo):+.2f}  dV {dv:+.2f}"
          f"  dVz {dz:+.2f}")

def absolute(dev, T):
    return [[sum(T[i][j] * x[j] for j in range(3)) for i in range(3)] for x in dev]

if __name__ == "__main__":
    rng = random.Random(1999); J = jac(Q0)
    T = [[1, 0, 0], [1, 1, 0], [1, 1, 1]]             # absolute = T x relative angles
    Jabs = matmul(J, inv(T))
    print("50,000 trials each (essentially the population values):")
    for kind in ("independent", "synergy"):
        dev = sample(rng, kind, 50000)
        report(f"{kind}, relative joint angles", dev, J)
        report(f"{kind}, absolute segment angles", absolute(dev, T), Jabs)
    print("bounds of dV for n = 3, d = 2: -1.5 (all ORT) to +3.0 (all UCM)")
    print("sampling: 20 trials per estimate, 4,000 repetitions")
    for kind in ("independent", "synergy"):
        est = []
        for _ in range(4000):
            vu, vo = ucm(sample(rng, kind, 20), J); est.append(math.log(vu / vo))
        est.sort()
        print(f"  {kind:11s} ln ratio, 2.5/50/97.5th pct: {est[100]:+.2f} /"
              f" {est[2000]:+.2f} / {est[3900]:+.2f}; share > +0.5: "
              f"{sum(e > 0.5 for e in est) / 4000:.3f}")
# 50,000 trials each (essentially the population values):
#   independent, relative joint angles    ln(Vucm/Vort) +0.00  dV +0.00  dVz -0.35
#   independent, absolute segment angles  ln(Vucm/Vort) -0.94  dV -0.76  dVz -0.82
#   synergy, relative joint angles        ln(Vucm/Vort) +3.60  dV +2.77  dVz +1.45
#   synergy, absolute segment angles      ln(Vucm/Vort) +1.80  dV +1.89  dVz +0.56
# bounds of dV for n = 3, d = 2: -1.5 (all ORT) to +3.0 (all UCM)
# sampling: 20 trials per estimate, 4,000 repetitions
#   independent ln ratio, 2.5/50/97.5th pct: -0.82 / -0.01 / +0.74; share > +0.5: 0.094
#   synergy     ln ratio, 2.5/50/97.5th pct: +2.77 / +3.60 / +4.37; share > +0.5: 1.000

The first trap is coordinate dependence. The same “no synergy” data give ln⁡(VUCM/VORT)=0.00\ln(V_{\mathrm{UCM}}/V_{\mathrm{ORT}}) = 0.00 in relative joint angles but −0.94 in absolute segment angles, and ΔV\Delta V changes from 0.00 to −0.76; the “synergy” data change from +3.60 to +1.80. The reason is that noise that is isotropic in one set of coordinates is no longer isotropic in another, while UCM analysis treats isotropy as the no-synergy baseline. Sternad and colleagues discussed the problem systematically in 2010 and showed that inferences about neural control drawn from the anisotropy of a covariance matrix are highly sensitive to the choice of coordinates, unless one knows in advance which coordinates the nervous system uses (Sternad et al. 2010).

The second trap is the zero point after transformation. In the three-link example n=3n = 3 and d=2d = 2, so at ΔV=0\Delta V = 0 we have ΔVz=−0.35\Delta V_z = -0.35, not zero. Comparing ΔVz\Delta V_z directly with zero would misread “no synergy” as “negative synergy”.

The third trap is sampling error. With only 20 trials, even data with no synergy at all give a 95 % range of the log ratio from −0.82 to +0.74, and 9.4 % of estimates exceed +0.5. A “synergy” for a single participant at a single moment needs a confidence interval and a comparison with a null model; the checklist of Chapter 8 makes this a required step.

Evidence

Why coordinates matter. Sternad, Park, Müller and Hogan pointed out in PLoS Computational Biology two problems: using the anisotropy of a covariance matrix as evidence of “preferential variability”, and using orthogonality to judge whether variability affects task performance; both are highly sensitive to the choice of coordinates. They proposed a two-level analysis that maps execution variability through a function to the task result, which depends much less on coordinates (Sternad et al. 2010). The “tolerance–noise–covariation” decomposition of Müller and Sternad belongs to this family of methods (Müller and Sternad 2004).

Optimal feedback control and the minimal intervention principle

UCM analysis describes the structure of variability but does not explain where it comes from. In 2002 Todorov and Jordan offered an explanation in Nature Neuroscience: if the nervous system works like an optimal feedback controller — at each moment computing commands from its estimate of the current state so as to minimise the combined cost of task error and effort — it will naturally correct only the deviations that interfere with the task and leave the rest alone (Todorov and Jordan 2002). They called this the minimal intervention principle: deviations from the average trajectory are corrected only when they interfere with task performance. Correcting task-irrelevant deviations costs effort and (under signal-dependent noise) introduces new noise, so leaving them alone is better. Todorov’s 2004 review discussed this idea together with other optimality principles (Todorov 2004), and Diedrichsen, Shadmehr and Ivry summarised the evidence for and against the framework in 2010 (Diedrichsen et al. 2010).

Evidence

Evidence at the level of muscles. Valero-Cuevas, Venkadesan and Todorov analysed the activation of several muscles in fingertip force tasks in 2009 and found that the variability of muscle activation is structured: larger in directions that do not affect the fingertip force and smaller in directions that do, as the minimal intervention principle predicts (Valero-Cuevas et al. 2009).

The program below demonstrates the principle with the “sum” task discussed by Todorov and Jordan. Two one-dimensional effectors (each a 1 kg mass whose force passes through a first-order filter with a 40 ms time constant) start at zero and must bring the sum of their positions to 0.20 m and stop at 0.5 s; every 10 ms command carries additive noise with SD 1 N. The optimal feedback controller (OFC) pays only for the sum of positions; the “servo” controller used for comparison pays for each effector reaching its own 0.10 m, which amounts to prescribing a target for every joint in advance.

Calculation

The linear–quadratic regulator. The state 𝒙\mathbf{x} contains the positions, velocities and forces of the two effectors and a constant 1 (to represent the target), with dynamics 𝒙t+1=A𝒙t+B𝒖t+noise\mathbf{x}_{t+1} = A\mathbf{x}_t + B\mathbf{u}_t + \text{noise}. The cost is the terminal cost 𝒙N⊤Q𝒙N\mathbf{x}_N^\top Q\,\mathbf{x}_N plus the effort r∥𝒖t∥2r\|\mathbf{u}_t\|^2 at every step; starting from SN=QS_N = Q and recursing backwards, Lt=(rI+B⊤St+1B)−1B⊤St+1A,St=A⊤St+1(A−BLt),L_t = (r I + B^\top S_{t+1} B)^{-1} B^\top S_{t+1} A,\qquad S_t = A^\top S_{t+1}(A – B L_t), gives the feedback law 𝒖t=−Lt𝒙t\mathbf{u}_t = -L_t\mathbf{x}_t. The program computes the task variance, the variance of the difference between the effectors and the corrective effort by exact propagation of the covariance (not by simulation), finds by bisection the servo weight with the same corrective effort as OFC, checks with 2,000 simulated trials, and finally pushes effector 1 by 10 mm at 0.15 s to see how the two controllers respond.

Calculation

# Optimal feedback control (finite-horizon LQR with additive motor noise) for the
# redundant "sum" task of Todorov & Jordan: two 1-D effectors (mass 1 kg, first-order
# force filter, tau = 40 ms) must bring p1 + p2 to 0.20 m and stop at T = 0.5 s.
# OFC pays only for the task sum; a "servo" controller pays for each effector reaching
# its own planned position (0.10 m). dt = 10 ms. All parameters are illustrative.
import math, random
DT, TAU, N, TARGET = 0.01, 0.04, 50, 0.20
NOISE = 1.0                                  # SD of additive motor noise per step, N

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 inv(M):
    n = len(M)
    A = [row[:] + [float(i == j) for j in range(n)] for i, row in enumerate(M)]
    for c in range(n):
        p = max(range(c, n), key=lambda r: abs(A[r][c])); A[c], A[p] = A[p], A[c]
        A[c] = [v / A[c][c] for v in A[c]]
        for r in range(n):
            if r != c:
                A[r] = [a - A[r][c] * b for a, b in zip(A[r], A[c])]
    return [row[n:] for row in A]

# state: p1 v1 f1 p2 v2 f2 1   (the constant 1 carries the target)
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 add(*Ms):
    return [[sum(v) for v in zip(*rows)] for rows in zip(*Ms)]

def final_cost(kind, w):
    stop = add(*[outer([float(i == j) for j in range(7)], wt)
                 for i, wt in ((1, 0.05), (4, 0.05), (2, 1e-4), (5, 1e-4))])
    if kind == "ofc":                         # (p1 + p2 - target)^2
        return add(stop, outer([1, 0, 0, 1, 0, 0, -TARGET], w))
    return add(stop, outer([1, 0, 0, 0, 0, 0, -TARGET / 2], w),
               outer([0, 0, 0, 1, 0, 0, -TARGET / 2], w))

def gains(Q, r=1e-6):
    S, L = Q, []
    for _ in range(N):
        K = mm(inv(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 run(L, rng, kick=None):
    x = [0.0] * 6 + [1.0]; effort = 0.0
    for t in range(N):
        if kick and t == kick[0]:
            x[0] += kick[1]
        u = [-sum(k * v for k, v in zip(row, x)) for row in L[t]]
        effort += sum(v * v for v in u) * DT
        if rng:
            u = [v + rng.gauss(0, NOISE) for v in u]
        x = [sum(a * v for a, v in zip(row, x)) + sum(b * w for b, w in zip(brow, u))
             for row, brow in zip(A, B)]
    return x[0], x[3], effort

def exact(L):
    """Propagate the state covariance exactly: task SD, UCM SD, ln ratio, effort."""
    S = [[0.0] * 7 for _ in range(7)]; effort = 0.0
    for t in range(N):
        effort += sum(sum(a * b for a, b in zip(r, c)) for r, c in
                      zip(mm(L[t], S), L[t])) * DT              # E[u'u] about the mean
        M = add(A, [[-v for v in r] for r in mm(B, L[t])])
        S = add(mm(mm(M, S), tr(M)), [[NOISE ** 2 * sum(b1 * b2 for b1, b2 in
                zip(B[i], B[j])) for j in range(7)] for i in range(7)])
    vo = (S[0][0] + S[3][3] + 2 * S[0][3]) / 2; vu = (S[0][0] + S[3][3] - 2 * S[0][3]) / 2
    p1, p2, _ = run(L, None)                       # mean end point (noise-free run)
    return (math.sqrt(2 * vo) * 1000, math.sqrt(2 * vu) * 1000, math.log(vu / vo),
            effort, 1000 * (p1 + p2 - TARGET))

def monte_carlo(L, seed, trials=2000):
    rng = random.Random(seed); s = [run(L, rng) for _ in range(trials)]
    m = sum(a + b for a, b, _ in s) / trials
    return 1000 * math.sqrt(sum((a + b - m) ** 2 for a, b, _ in s) / (trials - 1))

def row(name, r, extra=""):
    print(f"{name:20s} {r[0]:5.2f} mm {r[4]:+5.2f} mm {r[1]:6.2f} mm  {r[2]:+5.2f}"
          f"    {r[3]:.4f}{extra}")

if __name__ == "__main__":
    ofc = gains(final_cost("ofc", 1.0)); so = exact(ofc)
    print("controller          task SD  mean err  diff SD  ln ratio  corr. effort")
    row("OFC (task sum only)", so)
    lo, hi = 0.05, 50.0                       # servo weight with equal correction effort
    for _ in range(40):
        mid = math.sqrt(lo * hi)
        lo, hi = (mid, hi) if exact(gains(final_cost("servo", mid)))[3] < so[3] \
            else (lo, mid)
    ss = exact(gains(final_cost("servo", lo)))
    row("servo, equal effort", ss, f" (w {lo:.3f})")
    for w in (0.5, 4.0):
        row(f"servo, weight {w:3.1f}", exact(gains(final_cost("servo", w))))
    mc = (monte_carlo(ofc, 2002), monte_carlo(gains(final_cost("servo", lo)), 2002))
    print(f"check by simulation (2,000 trials): task SD OFC {mc[0]:.2f} mm,"
          f" servo {mc[1]:.2f} mm")
    print("perturbation: effector 1 displaced +10 mm at t = 0.15 s (noise-free run)")
    for name, Lg in (("OFC", ofc), ("servo", gains(final_cost("servo", lo)))):
        p1, p2, _ = run(Lg, None, (15, 0.01)); q1, q2, _ = run(Lg, None)
        print(f"  {name:5s} final change: p1 {1000 * (p1 - q1):+5.2f} mm, p2"
              f" {1000 * (p2 - q2):+5.2f} mm, sum {1000 * (p1 + p2 - q1 - q2):+5.2f} mm")
# controller          task SD  mean err  diff SD  ln ratio  corr. effort
# OFC (task sum only)   1.14 mm -1.46 mm  11.87 mm  +4.68    0.1534
# servo, equal effort   2.84 mm -10.17 mm   2.84 mm  +0.00    0.1534 (w 0.275)
# servo, weight 0.5     2.16 mm -5.71 mm   2.16 mm  +0.00    0.1659
# servo, weight 4.0     0.83 mm -0.73 mm   0.83 mm  +0.00    0.2141
# check by simulation (2,000 trials): task SD OFC 1.14 mm, servo 2.82 mm
# perturbation: effector 1 displaced +10 mm at t = 0.15 s (noise-free run)
#   OFC   final change: p1 +5.13 mm, p2 -4.87 mm, sum +0.25 mm
#   servo final change: p1 +1.59 mm, p2 +0.00 mm, sum +1.59 mm

The results agree with the minimal intervention principle. OFC has a task SD of 1.14 mm but an SD of 11.87 mm for the difference between the effectors, ln⁡(VUCM/VORT)=+4.68\ln(V_{\mathrm{UCM}}/V_{\mathrm{ORT}}) = +4.68: it does not correct “one too far, the other too short” at all. At the same corrective effort the servo controller has a task SD of 2.84 mm, 2.5 times that of OFC, with equal variance in both directions (log ratio 0.00); its mean end point also falls 10.17 mm short (1.46 mm for OFC), because the effort cost keeps a weakly weighted controller from getting all the way. To reach a smaller task error the servo controller must spend more effort (at weight 4 the task SD is 0.83 mm and the mean shortfall 0.73 mm, with corrective effort 0.2141, about 40 % more than OFC). The 2,000 simulated trials give 1.14 and 2.82 mm, in agreement with the exact calculation.

The perturbation shows the other side (Figure 2, right). After effector 1 is pushed by 10 mm, OFC shares the correction between the effectors: effector 1 ends 5.13 mm further and effector 2 travels 4.87 mm less, so the sum is off by only 0.25 mm. The servo controller sends effector 1 back on its own (it still ends 1.59 mm further) and ignores effector 2 completely, leaving an error of 1.59 mm in the sum. The phenomenon of “one element perturbed, the others compensating” is called motor equivalence in the motor control literature.

Left: final positions of the two effectors over 400 trials under optimal feedback control (OFC, blue) and servo control with equal corrective effort (gold); the dashed line is the set of combinations satisfying the task, p_1 + p_2 = 200 mm. The OFC points spread along the line, the servo points form a round cloud. Right: change in the position of each effector relative to an unperturbed run after effector 1 is pushed by 10 mm at 0.15 s.

Note

What this demonstration shows and what it does not. It shows that in a stated linear–quadratic model, an optimal controller that pays only for the task produces the structure VUCM≫VORTV_{\mathrm{UCM}} \gg V_{\mathrm{ORT}} and reaches a smaller task error with less effort. It is not a “proof” of optimal feedback control theory: the model assumes that the state is fully observed, that the noise is additive and that the cost function is known, whereas the real nervous system has sensory delays and noise (Chapter 4) and its cost function can only be inferred from behaviour. Conversely, observing VUCM≫VORTV_{\mathrm{UCM}} \gg V_{\mathrm{ORT}} does not prove that the nervous system performs optimal feedback control: passive mechanical properties, fixed synergies learned in advance, or merely the choice of coordinates (previous section) can produce similar structure (Diedrichsen et al. 2010; Sternad et al. 2010).

Summary

UCM analysis divides variability in joint space into a part that does not affect the task (UCM) and a part that does (ORT), compared per degree of freedom; the original study was a sit-to-stand task in nine people, in which the centre of mass was strongly stabilised. The synergy index ΔV\Delta V is bounded, and a log transform of it need not be zero at ΔV=0\Delta V = 0; the conclusions depend on coordinates (the same independent noise gives −0.94 in segment coordinates), and estimates from 20 trials carry sizeable error. The optimal feedback control of Todorov and Jordan offers one explanation of this structure: the minimal intervention principle. In the two-effector sum task at equal corrective effort, the task SD is 1.14 mm under OFC and 2.84 mm under servo control, and when one effector is pushed the other compensates in the opposite direction. This is a demonstration in a particular model, not a proof. The next chapter turns to another classical problem: the trade-off between the speed and the accuracy of movement.

References

Diedrichsen, Jörn, Reza Shadmehr, and Richard B. Ivry. 2010. “The Coordination of Movement: Optimal Feedback Control and Beyond.” Trends in Cognitive Sciences 14 (1): 31–39. https://doi.org/10.1016/j.tics.2009.11.004.
Latash, Mark L., John P. Scholz, Frédéric Danion, and Gregor Schöner. 2001. “Structure of Motor Variability in Marginally Redundant Multifinger Force Production Tasks.” Experimental Brain Research 141 (2): 153–65. https://doi.org/10.1007/s002210100861.
Latash, Mark L., John P. Scholz, and Gregor Schöner. 2002. “Motor Control Strategies Revealed in the Structure of Motor Variability.” Exercise and Sport Sciences Reviews 30 (1): 26–31. https://doi.org/10.1097/00003677-200201000-00006.
Latash, Mark L., John P. Scholz, and Gregor Schöner. 2007. “Toward a New Theory of Motor Synergies.” Motor Control 11 (3): 276–308. https://doi.org/10.1123/mcj.11.3.276.
Müller, Hermann, and Dagmar Sternad. 2004. “Decomposition of Variability in the Execution of Goal-Oriented Tasks: Three Components of Skill Improvement.” Journal of Experimental Psychology: Human Perception and Performance 30 (1): 212–33. https://doi.org/10.1037/0096-1523.30.1.212.
Scholz, John P., and Gregor Schöner. 1999. “The Uncontrolled Manifold Concept: Identifying Control Variables for a Functional Task.” Experimental Brain Research 126 (3): 289–306. https://doi.org/10.1007/s002210050738.
Sternad, Dagmar, Se-Woong Park, Hermann Müller, and Neville Hogan. 2010. “Coordinate Dependence of Variability Analysis.” PLoS Computational Biology 6 (4): e1000751. https://doi.org/10.1371/journal.pcbi.1000751.
Todorov, Emanuel. 2004. “Optimality Principles in Sensorimotor Control.” Nature Neuroscience 7 (9): 907–15. https://doi.org/10.1038/nn1309.
Todorov, Emanuel, and Michael I. Jordan. 2002. “Optimal Feedback Control as a Theory of Motor Coordination.” Nature Neuroscience 5 (11): 1226–35. https://doi.org/10.1038/nn963.
Valero-Cuevas, Francisco J., Madhusudhan Venkadesan, and Emanuel Todorov. 2009. “Structured Variability of Muscle Activations Supports the Minimal Intervention Principle of Motor Control.” Journal of Neurophysiology 102 (1): 59–68. https://doi.org/10.1152/jn.90324.2008.

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).