On Valve Circuits
Unknown
Submitted 1963-01-01 | SovietRxiv: ru-196301.74271 | Translated from Russian

Abstract Generated abstract

This paper studies the realization of Boolean matrices by valve circuits and estimates Shannon functions for circuit complexity under restrictions on circuit depth. Using decompositions of matrices into products and disjunctions of simpler matrices, it proves an asymptotic bound for depth two circuits realizing sparse Boolean matrices under specified density conditions. The main result extends previous estimates by showing that, for sequences of matrix dimensions whose logarithmic ratio has certain rational forms, arbitrary Boolean matrices can be realized with depth three circuits of asymptotic complexity equivalent to the unrestricted optimum, namely \(p_n q_n/\log_2(p_n q_n)\).

Full Text

CYBERNETICS AND CONTROL THEORY

E. I. NECHIPORUK

ON VALVE CIRCUITS

(Presented by Academician M. V. Keldysh, 4 VII 1962)

Let us consider the realization of Boolean matrices by valve circuits. We shall call the depth of a valve circuit the greatest of the lengths of chains connecting the input and output poles of the circuit. Denote by \(B_m(p,q)\) Shannon’s function for realizations of \((p,q)\)-matrices* by circuits of depth not exceeding \(m\), and by \(B(p,q)\) Shannon’s function for realizations by circuits of arbitrary depth. In \((^1)\) an asymptotic estimate was obtained for

\[ B(p_n,q_n) \]

under the condition

\[ v_n=\frac{\lg_2 q_n}{\lg_2 p_n}\to 0 \]

(and some others). At the same time it turns out that \(B(p_n,q_n)\sim B_2(p_n,q_n)\). Below we succeed in obtaining the estimate \(B(p_n,q_n)\sim B_3(p_n,q_n)\) under the condition that \(\lim v_n\) exists and belongs to some nowhere dense set of points of the interval \([0,1]\).

Estimates of the complexity of valve circuits have numerous applications; for example, the realization of linear transformations is almost uniquely described by valve circuits.

\(1^\circ\). Denote by \(\chi(\mathfrak A)\) the matrix realized by the circuit \(\mathfrak A\). Let \(\mathfrak A'\) be a \((p,r)\)-circuit and \(\mathfrak A''\) an \((r,q)\)-circuit. Denote by \(\mathfrak A'\times \mathfrak A''\) the \((p,q)\)-circuit obtained as a result of identifying the \(l\)-th output pole of the circuit \(\mathfrak A'\) with the \(l\)-th input pole of the circuit \(\mathfrak A''\) \((l=1,\ldots,r)\).

Lemma.

\[ \chi(\mathfrak A'\times \mathfrak A'')=\chi(\mathfrak A')\times \chi(\mathfrak A''). \]

(The symbol \(\times\) also denotes multiplication of matrices.)

Denote by \(\|A\|\) the number of ones in the matrix \(A\). Denote by \(\mathfrak B(p_n,q_n,\alpha_n)\) the class of all Boolean \((p_n,q_n)\)-matrices \(A\) such that \(\|A\|=\alpha_n p_n q_n\) (the conditions \(0\le \alpha_n\le 1\) and that \(\alpha_n p_n q_n\) is a natural number will be understood henceforth without explicit mention).

Theorem 1. Suppose the following conditions are satisfied:

a) \(q_n\le p_n\);

b) \(\alpha_n q_n\to\infty\);

c) \(\lg_2 p_n/\lg_2 \dfrac{1}{\alpha_n}\to \rho\), where \(\rho\) is an integer greater than zero;

d) \(p_n\alpha_n^\rho\to\infty\).

Then for \(\mathfrak B(p_n,q_n,\alpha_n)\)

\[ B_2(p_n,q_n)\sim \frac{\alpha_n p_n q_n}{\rho}. \]

Proof. The lower bound is obtained as in \((^1)\).

Upper bound. Let \(D_n\in \mathfrak B(p_n,q_n,\gamma_n)\), \(\beta_n\le \gamma_n\le \alpha_n\), where \(\beta_n\) is an arbitrary parameter satisfying the conditions

\[ \frac{\beta_n}{\alpha_n}\to 0,\qquad \beta_n q_n\to\infty,\qquad p_n\beta_n^\rho\to\infty. \]

* In the present paper the definition of a valve \((p,q)\)-circuit introduced in \((^1)\) is used.

** By a \((p,q)\)-matrix we mean a matrix having \(p\) rows and \(q\) columns.

If \(n\) is sufficiently large, then in the matrix \(D_n\) there exists a \((t_n,\rho)\)-submatrix consisting entirely of ones, with \(t_n \to \infty\). Indeed, the number of all \((1,\rho)\)-submatrices of the matrix \(D_n\) consisting entirely of ones is not less than \(p_n C^\rho_{[\gamma_n q_n]}\) (because the smallest number of such submatrices occurs in the case of a uniform distribution of the ones among the rows of the matrix \(D_n\)). Let us call the type of a \((1,\rho)\)-submatrix the system of numbers of the columns of the matrix \(D_n\) in which this submatrix is located. The \(t_n\) different \((1,\rho)\)-submatrices of one type form a \((t_n,\rho)\)-submatrix. Since the number of types is equal to \(C^\rho_{q_n}\), there exists at least one type to which belong at least

\[ t_n=\left]\frac{p_n C^\rho_{[\gamma_n q_n]}}{C^\rho_{q_n}}\right[ \]

\((1,\rho)\)-submatrices consisting entirely of ones. \(t_n \sim p_n\gamma_n^\rho\), since \(\gamma_n q_n \to \infty\).

We describe the process of constructing the matrices \(D_n^{(k)}\), \(k=1,\ldots,N_n(\beta_n)+1\). Denote by \(\gamma_n^{(k)}\) the number \(\|D_n^{(k)}\|/p_n q_n\). Put \(D_n^{(1)}=A_n\). If \(\beta_n \leqslant \gamma_n^{(k)}\), then, by the preceding, we select in the matrix \(D_n^{(k)}\) a \((t_n^{(k)},\rho)\)-submatrix consisting entirely of ones; denote it by \(a_n^{(k)}\). Denote by \(A_n^{(k)}\) the matrix obtained from \(D_n^{(k)}\) by replacing all elements by zeros except the elements of the submatrix \(a_n^{(k)}\), and by \(D_n^{(k+1)}\) the matrix obtained from \(D_n^{(k)}\) by replacing by zeros all elements of the submatrix \(a_n^{(k)}\); we have \(D_n^{(k)}=A_n^{(k)}\vee D_n^{(k+1)}\). If \(\gamma_n^{(k)}<\beta_n\), then we put \(k=N_n(\beta_n)+1\). Thus, the process consists in successively selecting submatrices filled with ones; the matrices formed become increasingly “sparse,” and the process terminates as soon as a matrix is formed containing fewer than \(\beta_n p_n q_n\) ones.

Denote by \(H_n\) the matrix \(D_n^{(N_n(\beta_n)+1)}\). We have
\(A_n=A_n^{(1)}\vee\ldots\vee A_n^{(N_n(\beta_n))}\vee H_n\),
\(\min_{k=1,\ldots,N_n(\beta_n)} t_n^{(k)} \gtrsim p_n\beta_n^\rho\),
\(\|H_n\|<\beta_n p_n q_n\).
Represent each matrix \(A_n^{(k)}\) in the form \(A_n^{(k)}=F_n^{(k)}\times G_n^{(k)}\), where \(F_n^{(k)}\) is a \((p_n,1)\)-matrix, \(G_n^{(k)}\) is a \((1,q_n)\)-matrix, \(\|F_n^{(k)}\|=t_n^{(k)}\), \(\|G_n^{(k)}\|=\rho\), and let

\[ F_n=\bigl(F_n^{(1)},\ldots,F_n^{(N_n(\beta_n))}\bigr),\qquad G_n= \begin{pmatrix} G_n^{(1)}\\ \vdots\\ G_n^{(N_n(\beta_n))} \end{pmatrix}. \]

Then

\[ A_n=F_n\times G_n\vee H_n, \tag{1} \]

\[ \|F_n\|\leqslant \frac{\alpha_n p_n q_n}{\rho},\qquad \|G_n\|\leqslant \frac{\alpha_n q_n}{\beta_n^\rho},\qquad \|H_n\|<\beta_n p_n q_n. \tag{2} \]

Realizing each matrix \(F_n\), \(G_n\), \(H_n\) by a circuit of depth 1, we obtain, by the lemma, a circuit for \(A_n\). By (2),
\(B_2(p_n,q_n)\lesssim \alpha_n p_n q_n/\rho\). The theorem is proved.

Remark. If conditions a), b), d) are satisfied and in condition c) \(\rho\) is not an integer, then

\[ B_2(p_n,q_n)\lesssim \frac{\alpha_n p_n q_n}{[\rho]}. \]

It can be shown that in this case the trivial power lower bound is ineffective, i.e., it can be improved.

2°. Theorem 2. Suppose the following conditions are satisfied:

a) \(q_n \leqslant p_n\);

b) \(q_n \to \infty\);

c)
\[ \lim_{n\to\infty}\frac{\lg_2 q_n}{\lg_2 p_n} = \frac{\mu}{\mu(\rho-1)+\rho}, \]
where \(\mu,\rho\) are positive integers.*

Then
\[ B(p_n,q_n)\sim B_3(p_n,q_n)\sim \frac{p_n q_n}{\lg_2(p_n q_n)}. \]

Proof. The lower estimate is the cardinality estimate.

The upper estimate. Let \(A_n\) be a \((p_n,q_n)\)-matrix. Introduce the parameter \(\vartheta_n\) and divide the matrix \(A_n\) into
\[ T_n=\left]\frac{q_n}{\vartheta_n}\right[ \]
nonoverlapping \((p_n,\vartheta_{n,k})\)-submatrices
\[ A_n^{(k)},\qquad A_n=\bigl(A_n^{(1)},\ldots,A_n^{(T_n)}\bigr), \]
so that \(\vartheta_{n,k}=\vartheta_n\) for \(k=1,\ldots,T_n-1\) and \(\vartheta_{n,T_n}\leqslant\vartheta_n\). Form a \((2^{\vartheta_{n,k}},\vartheta_{n,k})\)-matrix \(\Sigma_{n,k}\), whose rows are all possible distinct vectors, taken in arbitrary order. Obviously, there exist \((p_n,2^{\vartheta_{n,k}})\)-matrices \(B^{(k)}(A_n)\), having exactly one one in each row, such that
\[ A_n^{(k)}=B^{(k)}(A_n)\times \Sigma_{n,k},\qquad k=1,\ldots,T_n. \]
Introduce the matrices
\[ B(A_n)=\bigl(B^{(1)}(A_n),\ldots,B^{(T_n)}(A_n)\bigr), \]
\[ \Sigma_n= \begin{pmatrix} \Sigma_{n,1} & 0\\ \ddots & \\ 0 & \Sigma_{n,T_n} \end{pmatrix}. \]
Then
\[ A_n=B(A_n)\times \Sigma_n. \]

Denote by \(q'_n\) the number
\[ \sum_{k=1}^{T_n}2^{\vartheta_{n,k}}. \]
Introduce the parameter \(\lambda_n\) and divide the matrix \(B(A_n)\) into
\[ U_n=\left]\frac{p_n}{\lambda_n}\right[ \]
nonoverlapping \((\lambda_{n,i},q'_n)\)-submatrices
\[ B_i(A_n), \]
\[ B(A_n)= \begin{pmatrix} B_1(A_n)\\ \vdots\\ B_{U_n}(A_n) \end{pmatrix} \]
so that \(\lambda_{n,i}=\lambda_n\) for \(i=1,\ldots,U_n-1\) and \(\lambda_{n,U_n}\leqslant\lambda_n\). Form a \((\lambda_{n,i},C_{\lambda_{n,i}}^{\mu+1})\)-matrix \(\mathfrak S^{(n,i)}\), whose columns are all possible distinct vectors containing exactly \(\mu+1\) ones, taken in arbitrary order (if \(\lambda_{n,i}<\mu+1\), then we put \(C_{\lambda_{n,i}}^{\mu+1}=0\)). Obviously, there exist \((C_{\lambda_{n,i}}^{\mu+1},q'_n)\)-matrices \(C_i(A_n)\) \((i=1,\ldots,U_n)\) such that
\[ B_i(A_n)=\mathfrak S^{(n,i)}\times C_i(A_n)\vee B_i^*(A_n), \]
where
\[ \|B_i^*(A_n)\|\leqslant(\mu+1)q'_n,\qquad \frac{\lambda_n T_n}{\mu+1}-q'_n\leqslant \|C_i(A_n)\|\leqslant \frac{\lambda_n T_n}{\mu+1} \]
(the product of an \((a,0)\)-matrix by a \((0,b)\)-matrix is defined as an \((a,b)\)-matrix consisting entirely of zeros). Introduce the matrices
\[ \mathfrak S^{(n)}= \begin{pmatrix} \mathfrak S^{(n,1)} & 0\\ \ddots & \\ 0 & \mathfrak S^{(n,U_n)} \end{pmatrix}, \qquad C(A_n)= \begin{pmatrix} C_1(A_n)\\ \vdots\\ C_{U_n}(A_n) \end{pmatrix}, \qquad B^*(A_n)= \begin{pmatrix} B_1^*(A_n)\\ \vdots\\ B_{U_n}^*(A_n) \end{pmatrix}. \]

* Condition c) includes the important case for applications when \(p_n \succ q_n\).

Then

\[ B(A_n)=\mathfrak S^{(n)}\times C(A_n)\vee B^*(A_n). \tag{4} \]

Denote by \(p'_n\) the number \(\displaystyle \sum_{i=1}^{U_n} C_{\lambda_n,i}^{\mu+1}\); \(C(A_n)\) is a \((p'_n,q'_n)\)-matrix; denote by \(\alpha_n\) the number \(\|C(A_n)\|/p'_nq'_n\), and by \(v_n\) the number \(\lg_2 q_n/\lg_2 p_n\).

Introduce, for the parameters \(\vartheta_n,\lambda_n\), the conditions: 1) \(\lambda_n/2^{\vartheta_n}\to\infty\); 2) \(q_n/\vartheta_n\lambda_n^\mu\to\infty\). Then
\[ T_n\sim q_n/\vartheta_n,\quad q'_n\sim q_n2^{\vartheta_n}/\vartheta_n,\quad U_n\sim p_n/\lambda_n,\quad p'_n\sim p_n\lambda_n^\mu/(\mu+1)!,\quad \|C(A_n)\|\sim p_nq_n/(\mu+1)\vartheta_n, \]
\[ \alpha_n\sim \mu!/2^{\vartheta_n}\lambda_n^\mu,\quad \alpha_nq'_n\to\infty. \]

Introduce, further, the conditions: 3) \(\lg_2 p'_n\big/\lg_2\dfrac{1}{\alpha_n}\to\rho\); 4) \(p'_n\alpha_n^\rho\to\infty\).

Introduce the parameter \(\beta_n\) and the conditions: 5) \(\beta_n2^{\vartheta_n}\lambda_n^\mu\to0\); 6) \(\beta_nq_n2^{\vartheta_n}/\vartheta_n\to\infty\); 7) \(p_n\lambda_n^\mu\beta_n^\rho\to\infty\).

Apply to \(C(A_n)\) the construction of Theorem 1; by virtue of (1), (3), (4),
\[ C(A_n)=F_n\times G_n\vee H_n \]
and
\[ A_n=\mathfrak S^{(n)}\times F_n\times(G_n\times\Sigma_n)\vee (\mathfrak S^{(n)}\times H_n\vee B^*(A_n))\times\Sigma_n. \]

We obtain the circuit for \(A_n\) from circuits of depth 1 for \(F_n,\mathfrak S^{(n)},G_n\times\Sigma_n,H_n,B^*(A_n),\Sigma_n\) (lemma). We have
\[ \|\mathfrak S^{(n)}\|\le p_n\lambda_n^\mu,\quad \|\Sigma_n\|\le q_n2^{\vartheta_n},\quad \|B^*(A_n)\|\le \frac{p_nq_n2^{\vartheta_n}}{\vartheta_n\lambda_n}, \]
and, by virtue of (2),
\[ \|F_n\|\le p_nq_n(\mu+1)\rho^{\vartheta_n},\quad \|G_n\times\Sigma_n\|\le \vartheta_n\|G_n\|\le \frac{q_n}{\lambda_n^\mu p_n^{\beta_n^\rho}}, \]
\[ \|H_n\|\le \frac{\beta_np_nq_n2^{\vartheta_n}\lambda_n^\mu}{\vartheta_n}. \]

In order that the relations
\[ B_3(p_n,q_n)\sim\|F_n\|\le p_nq_n/\lg_2(p_nq_n) \]
hold, it is sufficient that the following additional conditions hold:
\[ \begin{gathered} 8)\;(\mu+1)\rho^{\vartheta_n}\ge(1+v_n)\lg_2p_n;\qquad 9)\;\lambda_n^\mu\lg_2p_n/q_n\to0;\qquad 10)\;2^{\vartheta_n}\lg_2p_n/p_n\to0;\\ 11)\;2^{\vartheta_n}\lg_2p_n/\vartheta_n\lambda_n\to0;\quad 12)\;\lg_2p_n/p_n\lambda_n^\mu\beta_n^\rho\to0;\quad 13)\;\beta_n2^{\vartheta_n}\lambda_n^\mu\lg_2p_n/\vartheta_n\to0. \end{gathered} \]

By the conditions of the theorem,
\[ v_n=\frac{\mu}{\mu(\rho-1)+\rho}+\varphi_n, \]
where \(\varphi_n\to0\). Put, for sufficiently large \(n\),
\[ \lambda_n=\left[p_n^{\frac{1}{\mu(\rho-1)+\rho}+\delta_n}\right],\quad \vartheta_n=\left[\left(\frac{1}{\mu(\rho-1)+\rho}+\xi_n\right)\lg_2p_n\right]; \]
\[ \frac{1}{\beta_n}=p_n^{\frac{\mu+1}{\mu(\rho-1)+\rho}+\eta_n}; \]
\[ \delta_n=\frac{\varphi_n-3\,\lg_2\lg_2p_n/\lg_2p_n}{\mu}; \quad \xi_n=-3\,\frac{\lg_2\lg_2p_n}{\lg_2p_n} +\min\left[\delta_n,\mu\left(\frac{1}{\rho}-1\right)\delta_n\right]; \]
\[ \eta_n=\mu\delta_n+\xi_n+\frac{\lg_2\lg_2p_n}{\lg_2p_n}. \]

Then conditions 1)—13) are fulfilled. The theorem is proved.

Leningrad State University
named after A. A. Zhdanov

Received
3 VII 1962

REFERENCES

  1. O. B. Lupanov, DAN, 111, 6, 1171 (1956).

Submission history

On Valve Circuits