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

Chapter 3 Fitts’s Law and the Speed–Accuracy Trade-off: The Limits of an Information-Theoretic Analogy

Fitts’s experiments of 1954

Speed and accuracy are hard to have together: the farther and smaller the target, the slower the movement. In 1954 Fitts turned this common observation into a quantitative relation, borrowing the information theory Shannon had proposed a few years earlier (Fitts 1954; Shannon 1948).

Evidence

The original experiments. Fitts ran three experiments (Fitts 1954). The first was reciprocal tapping: holding a stylus, the participant tapped back and forth between two target plates as fast as possible without missing; 16 right-handed male college students took part, with target widths WW of 2, 1, 0.5 and 0.25 inches and amplitudes AA of 2, 4, 8 and 16 inches, 16 combinations in all, each done with a 1-ounce and a 1-pound stylus. The second experiment moved washers from one pin to another (16 further male students), the third moved pins from one hole to another (10 men and 10 women). Fitts defined the index of difficulty Id=log⁡2(2A/Ws)I_d = \log_2(2A/W_s) (bits) and the index of performance Ip=Id/tI_p = I_d/t (bits per second). In the eight best conditions with the 1-ounce stylus, IpI_p lay between 10.3 and 11.5 bits/s, and the 16 values of IpI_p for the two styli correlated at 0.97. From this Fitts concluded that the information-handling capacity of the motor system is roughly constant over a considerable range of amplitudes and tolerances — about 10 to 12 bits per second across the three tasks — and falls off outside that range.

Rewriting a constant IpI_p gives what is usually called Fitts’s law: MT=a+blog⁡22AW.MT = a + b\,\log_2\frac{2A}{W}. The framework’s MT=a+blog⁡2(2D/W)MT = a + b\log_2(2D/W) is this original form (DD and AA both denote amplitude).

Three indices of difficulty

Later research proposed several logarithmic forms. In a long 1992 review MacKenzie promoted the so-called Shannon formulation log⁡2(A/W+1)\log_2(A/W + 1) and compared it with Fitts’s original and with Welford’s log⁡2(A/W+0.5)\log_2(A/W + 0.5) (MacKenzie 1992).

Evidence

MacKenzie’s reasons. First, the Shannon form never gives a negative difficulty, whereas the original and Welford forms do when the two targets overlap; second, Fitts’s law was borrowed from Shannon’s Theorem 17, and the Shannon form corresponds directly to that theorem; third, with the Shannon form and target widths adjusted for the actual errors, the fit to Fitts’s 1-ounce tapping data improved from a correlation of 0.9831 to 0.9937. MacKenzie also noted that the improvement amounts only to “slightly better prediction power”. The effective width multiplies the standard deviation of the end-point coordinates by 4.133 to give WeW_e: the entropy of a normal distribution is log⁡2(4.133σ)\log_2(4.133\sigma), which corresponds to about 96 % hits and 4 % misses (MacKenzie 1992). Summarising 27 years of research in 2004, Soukoreff and MacKenzie recommended computing throughput with the effective width and the Shannon form when evaluating pointing devices (Soukoreff and MacKenzie 2004).

The first part of the program below lists the three indices for several conditions of Fitts’s tapping experiment. At low difficulty they differ greatly: at A=2A = 2, W=2W = 2 the original and Shannon forms are both 1 bit and Welford’s only 0.58; at A=16A = 16, W=0.25W = 0.25 the original is 7 bits and the Shannon form 6.02. So the same data give different slopes and intercepts with different indices, and values of bb from different papers cannot be compared directly.

Where the logarithm comes from

Fitts borrowed the language of information theory, but “the motor system is a channel of fixed capacity” is only an analogy: Shannon’s Theorem 17 concerns a band-limited channel with additive Gaussian noise, and an arm is no such channel. The logarithmic form can be obtained from several different mechanisms, and some mechanisms give other forms.

Meyer and colleagues proposed the stochastic optimised-submovement model in 1988: an aimed movement consists of a primary submovement and, when needed, a corrective submovement; the end-point SD of each submovement is proportional to its mean velocity; and the mover chooses the durations of the two submovements to minimise the total time while keeping the hit rate (Meyer et al. 1988). With two submovements the model predicts movement time approximately linear in A/W\sqrt{A/W}, not in log⁡(A/W)\log(A/W). The impulse-variability theory of Schmidt and colleagues in 1979, aimed at very fast single movements, predicts an effective width linear in A/MTA/MT (Schmidt et al. 1979).

Calculation

The second part of the program writes the idea of Meyer’s model as a numerical optimisation (illustrative parameters): the primary submovement takes T1T_1 and lands with SD kA/T1kA/T_1; if it misses, a corrective submovement starts from the miss, takes T2T_2 and has SD k|e|/T2k|e|/T_2; T1T_1 and T2T_2 are chosen to minimise the mean time subject to 96 % hits overall (k=0.04k = 0.04 s). The “optimal” mean times for the 16 conditions of Fitts’s tapping experiment are then regressed on each of the three indices.

# Part 1: index of difficulty in Fitts's (1954) tapping design (amplitudes 2-16 in,
# widths 0.25-2 in): Fitts's log2(2A/W), MacKenzie's Shannon form log2(A/W + 1) and
# Welford's log2(A/W + 0.5). Part 2: a stochastic optimised-submovement model in the
# spirit of Meyer et al. (1988): a primary submovement of duration T1 lands with SD
# k*A/T1; if it misses, a corrective submovement of duration T2 starts from the miss
# with SD k*|e|/T2. T1 and T2 minimise mean time subject to 96 % hits overall
# (k = 0.04 s, illustrative). Do log or square-root laws describe the optimised times?
import math
K, HIT = 0.04, 0.96
A_SET, W_SET = (2, 4, 8, 16), (0.25, 0.5, 1, 2)

def p_in(half, sd):
    return math.erf(half / (math.sqrt(2) * sd)) if sd > 0 else 1.0

def p_total(A, W, T1, T2, m=120):
    s1 = K * A / T1; h = W / 2; p1 = p_in(h, s1)
    if T2 <= 0:
        return p1
    top = h + 8 * s1; step = (top - h) / m; acc = 0.0
    for i in range(m + 1):                     # Simpson over the miss distance e > W/2
        e = h + i * step; wgt = 1 if i in (0, m) else (4 if i % 2 else 2)
        dens = math.exp(-0.5 * (e / s1) ** 2) / (s1 * math.sqrt(2 * math.pi))
        acc += wgt * dens * p_in(h, K * e / T2)
    return p1 + 2 * acc * step / 3

def best(A, W):
    def mean_time(T1):
        s1 = K * A / T1; p1 = p_in(W / 2, s1)
        if p1 >= HIT:
            return T1, 0.0
        if p_total(A, W, T1, 50.0) < HIT:
            return math.inf, 0.0
        lo, hi = 0.0, 50.0
        for _ in range(40):
            mid = (lo + hi) / 2
            lo, hi = (lo, mid) if p_total(A, W, T1, mid) >= HIT else (mid, hi)
        return T1 + (1 - p1) * hi, hi
    grid = [0.02 * i for i in range(2, 200)]
    t1 = min(grid, key=lambda t: mean_time(t)[0])
    a, b = max(0.02, t1 - 0.02), t1 + 0.02     # golden-section refinement
    g = (math.sqrt(5) - 1) / 2
    for _ in range(30):
        c, d = b - g * (b - a), a + g * (b - a)
        a, b = (a, d) if mean_time(c)[0] < mean_time(d)[0] else (c, b)
    t1 = (a + b) / 2; mt, t2 = mean_time(t1)
    return mt, t1, t2

def fit(x, y):
    n = len(x); mx, my = sum(x) / n, sum(y) / n
    b = sum((u - mx) * (v - my) for u, v in zip(x, y)) / sum((u - mx) ** 2 for u in x)
    a = my - b * mx; ss = sum((v - a - b * u) ** 2 for u, v in zip(x, y))
    return a, b, 1 - ss / sum((v - my) ** 2 for v in y)

if __name__ == "__main__":
    print("   A     W   Fitts ID  Shannon ID  Welford ID   model MT    T1     T2")
    rows = []
    for A in A_SET:
        for W in W_SET:
            mt, t1, t2 = best(A, W); rows.append((A, W, mt))
            if (A, W) in ((2, 2), (2, 0.25), (4, 1), (8, 0.5), (16, 2), (16, 0.25)):
                print(f"{A:4d} {W:5.2f}   {math.log2(2 * A / W):6.2f}     "
                      f"{math.log2(A / W + 1):6.2f}      {math.log2(A / W + 0.5):6.2f}"
                      f"     {mt:6.3f} s  {t1:5.3f}  {t2:5.3f}")
    mt = [r[2] for r in rows]
    for name, f in (("Fitts log2(2A/W)", lambda A, W: math.log2(2 * A / W)),
                    ("Shannon log2(A/W+1)", lambda A, W: math.log2(A / W + 1)),
                    ("square root sqrt(A/W)", lambda A, W: math.sqrt(A / W))):
        x = [f(A, W) for A, W, _ in rows]; a, b, r2 = fit(x, mt)
        res = max(abs(m - a - b * u) for u, m in zip(x, mt))
        print(f"MT on {name:21s}: a = {a:+.3f} s, b = {b:.3f}, R^2 = {r2:.5f},"
              f" max |res| {1000 * res:4.1f} ms")
    print("16 conditions; A/W ranges from 1 to 64; MT depends on A/W only in this model")
#    A     W   Fitts ID  Shannon ID  Welford ID   model MT    T1     T2
#    2  0.25     4.00       3.17        3.09      0.419 s  0.245  0.248
#    2  2.00     1.00       1.00        0.58      0.110 s  0.079  0.095
#    4  1.00     3.00       2.32        2.17      0.276 s  0.171  0.178
#    8  0.50     5.00       4.09        4.04      0.623 s  0.348  0.350
#   16  0.25     7.00       6.02        6.01      1.319 s  0.697  0.698
#   16  2.00     4.00       3.17        3.09      0.419 s  0.245  0.248
# MT on Fitts log2(2A/W)     : a = -0.261 s, b = 0.190, R^2 = 0.90814, max |res| 250.5 ms
# MT on Shannon log2(A/W+1)  : a = -0.239 s, b = 0.226, R^2 = 0.94570, max |res| 196.6 ms
# MT on square root sqrt(A/W): a = -0.069 s, b = 0.173, R^2 = 0.99996, max |res|  5.7 ms
# 16 conditions; A/W ranges from 1 to 64; MT depends on A/W only in this model

In this model movement time depends only on A/WA/W and is almost perfectly linear in A/W\sqrt{A/W} (R2=0.99996R^2 = 0.99996, largest residual 5.7 ms, intercept −0.069 s); the residuals curve systematically, so the square-root relation is a very good approximation, not an exact law (only when A/WA/W is large and the primary submovement almost always misses does the optimal time tend to proportionality with A/W\sqrt{A/W}). Yet fitting Fitts’s original form to the same 16 points still gives R2=0.908R^2 = 0.908, and the Shannon form 0.946 — with the noise of real data, such fits are easily read as “consistent with Fitts’s law”. The negative intercepts (−0.261 and −0.239 s) are a clue that the true curve is steeper than a logarithm. Figure 3 (right) plots the same points against A/W\sqrt{A/W} and against log⁡2(2A/W)\log_2(2A/W). That a form fits well does not show that the mechanism behind the form produced the data.

Signal-dependent noise and two trajectory theories

Why do speed and accuracy conflict? One simple explanation is that the noise in motor commands grows with the size of the command: the faster the movement, the larger the forces needed and the larger the noise.

Evidence

Signal-dependent noise. Jones, Hamilton and Wolpert studied force fluctuations during isometric force production in 2002, found that the SD of force grows roughly linearly with the mean force, and analysed how the recruitment and firing properties of motor units contribute to this noise (Jones et al. 2002). The review of Franklin and Wolpert summarises such noise as follows: over a considerable range, the SD of motor noise is proportional to the mean signal, so the coefficient of variation is roughly constant (Franklin and Wolpert 2011).

On this basis Harris and Wolpert proposed in 1998 that the trajectory of a movement is chosen so as to minimise, under signal-dependent noise, the variance of the eye or hand position over a period after the movement (Harris and Wolpert 1998). With this single assumption they accounted for the velocity profiles of saccadic eye movements and of aimed arm movements, and argued that it predicts the speed–accuracy trade-off described by Fitts’s law.

An older theory is the minimum-jerk model of Flash and Hogan in 1985: given the start, the end and the duration, choose the trajectory that minimises the integral of the squared jerk (the third derivative of position) (Flash and Hogan 1985). Its solution is a fifth-order polynomial whose normalised velocity is 30τ2−60τ3+30τ430\tau^2 – 60\tau^3 + 30\tau^4, peaking at the midpoint at 1.875 times the mean velocity. Morasso observed in 1981 that in planar arm movements the hand path is nearly straight and the velocity bell-shaped (Morasso 1981), a fact both theories must explain; Uno, Kawato and Suzuki proposed in 1989 a dynamic model whose cost is the change in joint torque (Uno et al. 1989).

The framework writes “minimising end-point variance yields a bell-shaped velocity profile” together with “minimum jerk / minimum variance”. They should be kept apart: minimum jerk is a kinematic smoothness cost that has nothing to do with noise and does not depend on the mechanics of the limb; minimum variance is derived from noise, and its predictions depend on the dynamics of the limb. Both can give bell-shaped profiles, but for different reasons.

Calculation

The program represents the limb as a unit mass with stiffness KK (damping ratio 0.7) and a 40 ms force filter, and finds the minimum-variance command under signal-dependent noise (SD k|u|k|u|): subject to the mean trajectory being at rest at distance DD at time TT, minimise the summed positional variance over the following 300 ms (Lagrange multipliers; a 3 × 3 linear system). It then compares the velocity profiles for different KK with the minimum-jerk profile; the second part finds, for each index of difficulty, the shortest time with 4.133SD≤W4.133\,\mathrm{SD} \le W.

# Minimum-jerk (Flash & Hogan 1985) versus minimum-variance (Harris & Wolpert 1998).
# Plant: unit mass with stiffness K, damping ratio 0.7 and a first-order force filter
# (tau = 40 ms); dt = 5 ms. Motor noise is signal-dependent, SD = k |u|. The minimum-
# variance command minimises the summed positional variance over a 300 ms hold after
# the movement, subject to the MEAN trajectory reaching distance D at time T at rest.
# Part 2: shortest T with 4.133 SD <= W for each index of difficulty (K = 10).
import math
DT, TAU, HOLD = 0.005, 0.04, 60

def plant(K):
    b = 2 * 0.7 * math.sqrt(K)
    return [[1, DT, 0], [-DT * K, 1 - DT * b, DT], [0, 0, 1 - DT / TAU]], [0, 0, DT / TAU]

def simulate(A, B, u):
    x, out = [0.0, 0.0, 0.0], []
    for ui in u:
        x = [sum(a * v for a, v in zip(r, x)) + bi * ui for r, bi in zip(A, B)]
        out.append(x)
    return out

def solve(M, b):
    n = len(M); A = [r[:] + [x] for r, x in zip(M, b)]
    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]
        for r in range(n):
            if r != c:
                A[r] = [a - A[r][c] / A[c][c] * x for a, x in zip(A[r], A[c])]
    return [A[i][n] / A[i][i] for i in range(n)]

def min_variance(K, nt, D=1.0):
    A, B = plant(K); g = simulate(A, B, [1.0] + [0.0] * (nt + HOLD))
    c = [sum(g[t - s][0] ** 2 for t in range(nt, nt + HOLD)) for s in range(nt)]
    G = [[g[nt - s][k] for s in range(nt)] for k in range(3)]
    M = [[sum(G[i][s] * G[j][s] / c[s] for s in range(nt)) for j in range(3)]
         for i in range(3)]
    lam = solve(M, [D, 0.0, K * D])                   # at rest at D: force = K D
    u = [sum(lam[k] * G[k][s] for k in range(3)) / c[s] for s in range(nt)]
    var = sum(ci * ui * ui for ci, ui in zip(c, u))  # hold-phase holding command K D
    var += (K * D) ** 2 * sum(sum(g[t - s][0] ** 2 for s in range(nt, t))
                              for t in range(nt, nt + HOLD))
    return u, math.sqrt(var / HOLD)                   # u and positional SD / k

def profile(K, nt):
    A, B = plant(K); u, _ = min_variance(K, nt)
    return [x[1] * nt * DT for x in simulate(A, B, u)]   # velocity / mean velocity

if __name__ == "__main__":
    nt = 100; q = (0.1, 0.25, 0.5, 0.75, 0.9)
    mj = [30 * x ** 2 - 60 * x ** 3 + 30 * x ** 4
          for x in ((i + 1) / nt for i in range(nt))]
    print("T = 0.5 s; velocity/mean velocity at 10/25/50/75/90 % of T; peak (time);"
          " RMS gap")
    rows = [("minimum jerk", mj)] + [(f"min-variance, K = {K:2d}", profile(K, nt))
                                     for K in (0, 5, 10, 20)]
    for name, v in rows:
        pk = max(range(nt), key=lambda i: v[i])
        rms = math.sqrt(sum((a - b) ** 2 for a, b in zip(v, mj)) / nt)
        print(f"  {name:22s} " + " ".join(f"{v[round(x * nt) - 1]:5.2f}" for x in q)
              + f"   {v[pk]:.2f} ({(pk + 1) / nt:.2f} T)  {rms:.2f}")
    print("shortest MT with 4.133 SD <= W; k set so that ID 2 needs T = 0.10 s:")
    for K in (0, 10):
        table = [(n * DT, min_variance(K, n)[1]) for n in range(20, 401, 2)]
        k = 2 / (4.133 * 4 * table[0][1])           # SDN coefficient for this plant
        ids, mts = [], []
        for idx in (2, 3, 4, 5, 6, 7):
            need = 2 / (4.133 * k * 2 ** idx)        # allowed SD/(k D) when W = 2D/2^ID
            ids.append(idx)                          # tolerance: float rounding at ID 2
            mts.append(next(t for t, r in table if r <= need * (1 + 1e-9)))
        n = len(ids); mx, my = sum(ids) / n, sum(mts) / n
        b = sum((x - mx) * (y - my) for x, y in zip(ids, mts)) / \
            sum((x - mx) ** 2 for x in ids)
        r2 = 1 - sum((y - my - b * (x - mx)) ** 2 for x, y in zip(ids, mts)) / \
            sum((y - my) ** 2 for y in mts)
        lx = [i * math.log(2) for i in ids]; ly = [math.log(t) for t in mts]
        ex = sum((x - sum(lx) / n) * (y - sum(ly) / n) for x, y in zip(lx, ly)) / \
            sum((x - sum(lx) / n) ** 2 for x in lx)
        print(f"  K = {K:2d}: " + " ".join(f"{t:.2f}" for t in mts) +
              f" s (ID 2-7); linear R^2 {r2:.3f}; MT ~ (2D/W)^{ex:.2f}")
# T = 0.5 s; velocity/mean velocity at 10/25/50/75/90 % of T; peak (time); RMS gap
#   minimum jerk            0.24  1.05  1.88  1.05  0.24   1.88 (0.50 T)  0.00
#   min-variance, K =  0    0.16  0.68  1.51  1.65  0.79   1.75 (0.67 T)  0.41
#   min-variance, K =  5    0.28  0.98  1.61  1.31  0.53   1.63 (0.55 T)  0.20
#   min-variance, K = 10    0.36  1.17  1.64  1.12  0.40   1.64 (0.48 T)  0.14
#   min-variance, K = 20    0.52  1.50  1.63  0.81  0.22   1.76 (0.39 T)  0.27
# shortest MT with 4.133 SD <= W; k set so that ID 2 needs T = 0.10 s:
#   K =  0: 0.10 0.15 0.21 0.33 0.56 1.09 s (ID 2-7); linear R^2 0.816; MT ~ (2D/W)^0.67
#   K = 10: 0.10 0.14 0.20 0.29 0.44 0.77 s (ID 2-7); linear R^2 0.860; MT ~ (2D/W)^0.58

There are two results. First, the minimum-variance velocity profile changes with the mechanics of the limb: for pure inertia (K=0K = 0) the peak is at 0.67TT and the curve leans late; at K=10K = 10 the peak is at 0.48TT and the profile is closest to minimum jerk (RMS difference 0.14 mean velocities); at K=20K = 20 the peak moves forward to 0.39TT. Peak velocity is 1.63–1.76 times the mean, below the 1.875 of minimum jerk. So the skew of the velocity profile can in principle distinguish the two theories, but only once the mechanical parameters of the limb are known (Figure 3, left).

Second, in this simplified open-loop model, as the index of difficulty goes from 2 to 7 bits the shortest time rises from 0.10 s to 1.09 s (K=0K = 0) or 0.77 s (K=10K = 10), approximately a power function (2D/W)0.67(2D/W)^{0.67} or (2D/W)0.58(2D/W)^{0.58}, and a linear fit on the index of difficulty has R2R^2 of only 0.82–0.86. Harris and Wolpert reported that their model agrees with Fitts’s law; the book’s simplified model has no feedback corrections and gives a relation closer to a power function. This does not show that they were wrong; it shows that “minimum variance implies Fitts’s law” depends on the details of the model, and that a logarithmic speed–accuracy relation may well require feedback corrections during the movement (such as the corrective submovement of Meyer’s model).

Left: normalised velocity profiles at T = 0.5 s for minimum jerk (grey) and minimum variance at three stiffnesses (program p05). Right: mean times of a Meyer-type optimised-submovement model for the 16 conditions of Fitts’s tapping experiment, plotted against \sqrt{A/W} (blue dots, bottom axis) and against \log_2(2A/W) (open squares, top axis); the former is a straight line, the latter bends slightly upwards.

Note

How to read the “information-theoretic boundary”. Fitts’s index of performance measures the index of difficulty completed per unit time, in “bits per second”, but these “bits” are defined by A/WA/W; they are not an amount of information measured in neural signals. Saying that the motor system has an “upper limit of 10 bits per second” mistakes an analogy for a physical law. What can be tested is whether, for a given task and definition of the index, MTMT is linear in the index of difficulty, what the intercept and slope are, and how they change with practice, age and device (Soukoreff and MacKenzie 2004).

Summary

Fitts’s 1954 tapping experiment (16 participants, amplitudes 2–16 inches, widths 0.25–2 inches) gave MT=a+blog⁡2(2A/W)MT = a + b\log_2(2A/W), with an index of performance of about 10.3–11.5 bits/s in the best conditions. The Shannon form log⁡2(A/W+1)\log_2(A/W + 1) is another index, promoted by MacKenzie, which differs greatly from the original at low difficulty. The logarithmic form is not the only possibility: in a Meyer-type optimised-submovement model, movement time is almost perfectly linear in A/W\sqrt{A/W} (R2=0.99996R^2 = 0.99996), yet the logarithmic form still fits with R2=0.908R^2 = 0.908. Minimum jerk and minimum variance are different theories: the former is a kinematic smoothness cost, the latter is derived from signal-dependent noise and its velocity profile changes with limb stiffness; in the book’s simplified open-loop model, minimum variance gives a speed–accuracy relation closer to a power function. The next chapter discusses another limitation — the delay of sensory feedback — and how the nervous system gets around it with internal models.

References

Fitts, Paul M. 1954. “The Information Capacity of the Human Motor System in Controlling the Amplitude of Movement.” Journal of Experimental Psychology 47 (6): 381–91. https://doi.org/10.1037/h0055392.
Flash, Tamar, and Neville Hogan. 1985. “The Coordination of Arm Movements: An Experimentally Confirmed Mathematical Model.” Journal of Neuroscience 5 (7): 1688–703. https://doi.org/10.1523/JNEUROSCI.05-07-01688.1985.
Franklin, David W., and Daniel M. Wolpert. 2011. “Computational Mechanisms of Sensorimotor Control.” Neuron 72 (3): 425–42. https://doi.org/10.1016/j.neuron.2011.10.006.
Harris, Christopher M., and Daniel M. Wolpert. 1998. “Signal-Dependent Noise Determines Motor Planning.” Nature 394 (6695): 780–84. https://doi.org/10.1038/29528.
Jones, Kelvin E., Antonia F. de C. Hamilton, and Daniel M. Wolpert. 2002. “Sources of Signal-Dependent Noise During Isometric Force Production.” Journal of Neurophysiology 88 (3): 1533–44. https://doi.org/10.1152/jn.2002.88.3.1533.
MacKenzie, I. Scott. 1992. “Fitts’ Law as a Research and Design Tool in Human-Computer Interaction.” Human-Computer Interaction 7 (1): 91–139. https://doi.org/10.1207/s15327051hci0701_3.
Meyer, David E., Richard A. Abrams, Sylvan Kornblum, Charles E. Wright, and J. E. Keith Smith. 1988. “Optimality in Human Motor Performance: Ideal Control of Rapid Aimed Movements.” Psychological Review 95 (3): 340–70. https://doi.org/10.1037/0033-295X.95.3.340.
Morasso, Pietro. 1981. “Spatial Control of Arm Movements.” Experimental Brain Research 42 (2): 223–27. https://doi.org/10.1007/BF00236911.
Schmidt, Richard A., Howard Zelaznik, Brian Hawkins, James S. Frank, and Jr. Quinn John T. 1979. “Motor-Output Variability: A Theory for the Accuracy of Rapid Motor Acts.” Psychological Review 86 (5): 415–51. https://doi.org/10.1037/0033-295X.86.5.415.
Shannon, Claude E. 1948. “A Mathematical Theory of Communication.” Bell System Technical Journal 27 (3): 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x.
Soukoreff, R. William, and I. Scott MacKenzie. 2004. “Towards a Standard for Pointing Device Evaluation, Perspectives on 27 Years of Fitts’ Law Research in HCI.” International Journal of Human-Computer Studies 61 (6): 751–89. https://doi.org/10.1016/j.ijhcs.2004.09.001.
Uno, Yoji, Mitsuo Kawato, and Rika Suzuki. 1989. “Formation and Control of Optimal Trajectory in Human Multijoint Arm Movement: Minimum Torque-Change Model.” Biological Cybernetics 61 (2): 89–101. https://doi.org/10.1007/BF00204593.

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).