Binary Integration Gain
Binary Integration Gain
28, 2017
1 Acknowledgement
Thanks to Gabriel Beltrão for bringing this issue to my attention, and for providing an independent check
of the calculations.
2 Introduction
Binary integration (BI), also called M-of-N processing or integration, is a technique for combining
multiple threshold detection test outcomes to form a single, final detection test outcome that achieves
specified probabilities of detection and false alarm, PD and PFA, with a lower single-sample signal-to-
noise ratio (SNR) than would be required for a single threshold test based on a single measurement.
Alternatively, for a given SNR, binary integration can achieve lower PFA and/or higher PD than a
detection test using a single measurement. A description of the basic technique is given in section 6.4 of
[1] and will not be repeated here.
A shortcoming of the discussion in [1] is that it only computes the effect of the processing on the post-BI
probabilities given a single-trial probability, and then suggests that the best choice of M for a given N is
the one that maximizes the range of pre-BI detection probabilities for which the post-BI detection
probability is increased. (Any reasonable pre-BI false alarm probability will be decreased for any choice
of M except M = 1.) A more useful basis for evaluation would be to compute a binary integration gain
GBI, i.e. the factor by which the required single-measurement SNR needed to achieve a given PD and PFA
is decreased when using BI, vs. that required when detection is based on only a single measurement.
This is in exact analogy to the idea of both coherent and noncoherent integration. Accordingly, the goal
of this memo is to develop a simple numerical procedure for evaluating GBI for any M and N and
compare the result to the corresponding noncoherent and coherent integration gains.
In the discussion that follows, the target is always assumed to be nonfluctuating. The Swerling target
fluctuations case is commented on briefly in Section 5.3.
1|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
separate threshold tests, and would only declare a detection if the threshold were crossed on any 3 or
more of those tests. When using the BI procedure, let the probabilities of detection and false alarm for a
single trial be PD1 and PFA1, and note that the desired probabilities after BI are just PD and PFA . Let the
reference SNR required to achieve a specified PD and PFA using a single measurement be , and the SNR
required to achieve a specified PD1 and PFA1 on a single measurement be 1. We restrict ourselves to the
following additional conditions, mainly because they comprise the easiest form of the problem:
The first two conditions in this list mean that when using BI, the “cumulative” probability of detection PD
is related to the single-trial detection probability PD1 according to Eq. (6.116) of [1], repeated here in the
notation of this memo:
N
N N r
PD r PDr1 1 PD1 (1)
r M
An identical equation relates PFA and PFA1. A numerical procedure for computing the binary integration
gain is straightforward:
1
In MATLAB®, the solution can be accomplished very efficiently by appropriate use of the fzero function.
2|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
4 Example
Figure 1 illustrates the binary integration gain for a square-law detector, PFA = 10−6, N = 8, and all
choices of M (1 through 8). Qualitatively similar results are obtained for other PFA values from 10−2 to
10−8, and for N from 3 through 7. Following is an example of computing one point on this chart, namely
the case of 3-of-8 processing with a PFA of 10−6 and PD of 0.7:
1. Shnidman’s SNR equation gives the SNR required to achieve this performance with a single
measurement as = 15.97 on a linear scale, which is 12.03 dB.
2. The single-trial probability PFA1 required to achieve PFA = 10−6 with 3-of-8 binary integration is
found by a numerical search to be 0.0026.
3. Similarly, the single-trial probability PD1 required to achieve PD = 0.7 with 3-of-8 binary
integration is found by a numerical search to be 0.4075.
4. Shnidman’s SNR equation gives the SNR required to achieve PFA1 = 0.0026 and PD1 = 0.4075 with
a single measurement to be 1 = 4.78 = 6.8 dB.
5. The binary integration gain for this case is GBI = /1 = 3.34 = 5.24 dB. Within roundoff errors,
this of course is also equal to the difference of the two SNRs in steps 1 and 4 in dB, i.e.
12.03 – 6.8 = 5.23 dB.
The data point in Figure 1 corresponding to this example is shown by the red circle.
Figure 1. Binary integration gain GBI due to binary integration for N = 8 and all possible values of M. In
all cases, PFA = 10−6. The circled data point corresponds to the numerical example above. Also shown
are the noncoherent gain Gnc and the coherent integration gain Gc for N = 8.
3|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
5 Discussion
In [3], Shnidman considers binary integration gain. He defines the optimum value of M to be that which
maximizes PD for a given N, PFA, SNR , and target fluctuation model. This is equivalent to our definition
of the optimum M as the one that maximizes integration gain, since that corresponds to minimizing the
required for a given PD, PFA, and N. He observes a similar broad maximum in the value of Mopt over all
of the target models considered as well as a much broader range of N (up to 1000) than considered
here. He also proposes an empirical estimate of Mopt for the nonfluctuating target:
M opt round 100.02 N 0.8 round 0.955 N 0.8 (nonfluctuating target) (2)
The round(∙) function was added here to give an integer result, as required in practice. This estimate
is stated to be valid for N from 5 to 700. Table 1 shows that the value of Mopt predicted by Eqn. (2)
usually matches those observed from calculations like those leading to Figure 1.
Table 1. Observed and predicted values of Mopt for a nonfluctuating target with PFA = 10−6.
N=3 4 5 6 7 8
Observed Mopt 2 3 4 4 5 5
Eqn. (2) 2 3 3 4 5 5
4|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
Figure 2. Reduction in integration gain (“loss”) when using binary integration with the optimum value
of M, as compared to using noncoherent integration for a nonfluctuating target and PFA = 10−6.
2
There are some cases with Swerling 2 and 4 targets where the noncoherent integration gain actually exceeds the
coherent integration gain. However, even in these cases the SNR required to achieve a given detection
5|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
measurements of a nonfluctuating target in noise, and assuming the goal is to detect the target’s
presence with the minimum required SNR, coherent integration should be performed first to the
maximum extent possible. If coherent integration is not feasible or if it is limited to fewer than N
samples at a time (perhaps due to radar-target motion-induced phase errors, transceiver phase
instabilities, etc.), then there will still be N N samples available to be combined. The results above
show that these samples should be noncoherently integrated if possible. Finally, if noncoherent
integration of the remaining samples is not feasible, binary integration can be applied with only a minor
additional loss.
It is worth noting that binary integration is often combined with multiple pulse repetition frequency
(PRF) data acquisition to achieve more than just detection. In particular, collecting N coherent
processing intervals (CPIs) of data in a pulse Doppler radar, each at a different PRF; applying appropriate
threshold detection processing to each CPI; and then applying binary integration across the CPIs in each
range-Doppler cell not only achieves an integration gain for detection but can also be used to reduce or
avoid range-Doppler blind zones and resolve range and Doppler ambiguities. See [1] for an introduction
to this style of processing.
6 References
[1] M. A. Richards, Fundamentals of Radar Signal Processing, second edition. McGraw-Hill, 2014.
[2] M. A. Richards, “Notes on Noncoherent Integration Gain”, technical memorandum, July 17,
2014. Available at www.radarsp.com.
[3] D. A. Shnidman, “Binary Integration for Swerling Target Fluctuations”, IEEE Trans. Aerospace and
Electronic Systems, vol. 34, no. 3, pp. 1043-1053, July 1998.
7 MATLAB® Code
Versions of the following code were used to generate the figures in this memo.
%
% binary_int_gain_nonfluc
%
% Computes the integration gain from using binary integration. A
% nonfluctuating target is assumed. Shnidman's equation is used to
% estimate various SNRs needed, so the result is approximate but probably
% pretty good.
%
% Reference: Section 6.4 of Richards, "Fundamentals of Radar Signal
% Processing".
%
% Mark A. Richards, September 2016.
clear all
close all
performance is less with coherent integration than with noncoherent integration. A discussion of this phenomenon
can be found in [2].
6|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
% results for all choices of M from M=1 (1-of-N) to M=N (N-of-N). We'll
% also loop over multiple values of N.
NN = 3:8';
NNlen = length(NN);
PD = [0.3:0.05:0.9,0.92,0.94,0.95];
% PD = [0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.95]';
PDlen = length(PD);
% Compute the SNR required to achieve this Pd and Pfa with a single
% test, i.e. not using M-of-N processing.
SNR_1 = shnidman(1,Pd,Pfa,0);
SNR_1dB = 10*log10(SNR_1);
for Nn = 1:NNlen
N = NN(Nn);
% Compute the SNR for achieving this Pd and Pfa using N samples
% and noncoherent integration; use that to compute noncoherent
% integration gain for this case
SNR_N = shnidman(N,Pd,Pfa,0);
SNR_NdB = 10*log10(SNR_N);
gain_noncoh_dB(PFAn,PDn,Nn) = SNR_1dB - SNR_NdB;
gain_coh_dB(PFAn,PDn,Nn) = 10*log10(N);
for M = 1:N
M;
7|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
% M-of-N method.
SNR_MN = shnidman(1,Pd1,Pfa1,0);
SNR_MNdB = 10*log10(SNR_MN);
% All of the results are in the gain(:,:,:,:) matrix. Now some plots.
% This parameter chooses which Pfa to use. PFAn = 1 for 1e-2, PFAn = 2 for
% 1e-3, ,,, , PFAn = 7 for 1e-8
8|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
x = PD;
y = [gain_BI_dB(PFAn,:,2,1); % Pfa = 1e-6, all Pd values, N = 4, M = 1
gain_BI_dB(PFAn,:,2,2); % Pfa = 1e-6, all Pd values, N = 4, M = 2
gain_BI_dB(PFAn,:,2,3); % Pfa = 1e-6, all Pd values, N = 4, M = 3
gain_BI_dB(PFAn,:,2,4); % Pfa = 1e-6, all Pd values, N = 4, M = 4
gain_noncoh_dB(PFAn,:,2); % noncoherent gain, Pfa = 1e-6, all Pd values, N = 4
gain_coh_dB(PFAn,:,2)]; % coherent gain, N = 4
plot(x,y')
xlim([0.3 1]);
ylim([0,10]);
grid
xlabel('Pd')
ylabel('Gain (dB)')
title(['M-of-4, Pfa = ',num2str(PFA(PFAn))])
legend('1 of 4','2 of 4','3 of 4','4 of 4','Noncoherent','Coherent')
9|P age
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
% Now let's use our data to see what the integratino loss is compared to
% noncoherent. Generate a plot of the difference in integration gain
% between the noncoherent case and the best binary case, as a function of
% Pd, for a fixed Pfa.
loss_NCmBI_dB = zeros(PDlen,NNlen);
for Nn = 1:NNlen
Mopt = round((10^-0.02)*NN(Nn)^0.8)
for PDn = 1:PDlen
% loss_NCmBI_dB(PDn,Nn) = gain_noncoh_dB(PFAn,PDn,Nn) -
max(gain_BI_dB(PFAn,PDn,Nn,:));
loss_NCmBI_dB(PDn,Nn) = gain_noncoh_dB(PFAn,PDn,Nn) -
gain_BI_dB(PFAn,PDn,Nn,Mopt);
end
end
figure
plot(PD,loss_NCmBI_dB)
grid
xlabel('Pd')
10 | P a g e
M. A. Richards, “Binary Integration Gain” Nov. 28, 2017
11 | P a g e